Upload 232 files

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+modeling_whisper.cpython-312[[:space:]](1).pyc filter=lfs diff=lfs merge=lfs -text
+modeling_whisper.cpython-312.pyc filter=lfs diff=lfs merge=lfs -text
+realtime-whisper-webgpu/public/banner.png filter=lfs diff=lfs merge=lfs -text
+realtime-whisper-webgpu/public/logo.png filter=lfs diff=lfs merge=lfs -text
+realtime-whisper-webgpu/public/realtime-whisper-webgpu.mp4 filter=lfs diff=lfs merge=lfs -text

configuration_whisper (1).py

@@ -0,0 +1,344 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Whisper model configuration"""
+
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig, OnnxSeq2SeqConfigWithPast
+from ...utils import logging
+
+
+if TYPE_CHECKING:
+    from ...feature_extraction_utils import FeatureExtractionMixin
+    from ...tokenization_utils_base import PreTrainedTokenizerBase
+    from ...utils import TensorType
+
+logger = logging.get_logger(__name__)
+
+
+# fmt: off
+NON_SPEECH_TOKENS = [
+    1, 2, 7, 8, 9, 10, 14, 25,
+    26, 27, 28, 29, 31, 58, 59, 60, 61, 62,
+    63, 90, 91, 92, 93, 357, 366, 438, 532, 685,
+    705, 796, 930, 1058, 1220, 1267, 1279, 1303, 1343, 1377,
+    1391, 1635, 1782, 1875, 2162, 2361, 2488, 3467, 4008, 4211,
+    4600, 4808, 5299, 5855, 6329, 7203, 9609, 9959, 10563, 10786,
+    11420, 11709, 11907, 13163, 13697, 13700, 14808, 15306, 16410, 16791,
+    17992, 19203, 19510, 20724, 22305, 22935, 27007, 30109, 30420, 33409,
+    34949, 40283, 40493, 40549, 47282, 49146, 50257, 50359, 50360, 50361
+]
+NON_SPEECH_TOKENS_MULTI = [
+    1, 2, 7, 8, 9, 10, 14, 25,
+    26, 27, 28, 29, 31, 58, 59, 60, 61, 62,
+    63, 90, 91, 92, 93, 359, 503, 522, 542, 873,
+    893, 902, 918, 922, 931, 1350, 1853, 1982, 2460, 2627,
+    3246, 3253, 3268, 3536, 3846, 3961, 4183, 4667, 6585, 6647,
+    7273, 9061, 9383, 10428, 10929, 11938, 12033, 12331, 12562, 13793,
+    14157, 14635, 15265, 15618, 16553, 16604, 18362, 18956, 20075, 21675,
+    22520, 26130, 26161, 26435, 28279, 29464, 31650, 32302, 32470, 36865,
+    42863, 47425, 49870, 50254, 50258, 50360, 50361, 50362
+]
+# fmt: on
+
+
+class WhisperConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`WhisperModel`]. It is used to instantiate a
+    Whisper model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Whisper
+    [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 51865):
+            Vocabulary size of the Whisper model. Defines the number of different tokens that can be represented by the
+            `decoder_input_ids` passed when calling [`WhisperModel`]
+        num_mel_bins (`int`, *optional*, defaults to 80):
+            Number of mel features used per input features. Should correspond to the value used in the
+            `WhisperProcessor` class.
+        encoder_layers (`int`, *optional*, defaults to 4):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 4):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 6):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 6):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 1536):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in encoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 1536):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_start_token_id (`int`, *optional*, defaults to 50257):
+            Corresponds to the "<|startoftranscript|>" token, which is automatically used when no `decoder_input_ids`
+            are provided to the `generate` function. It is used to guide the model`s generation process depending on
+            the task.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        is_encoder_decoder (`bool`, *optional*, defaults to `True`):
+            Whether the model is used as an encoder/decoder or not.
+        activation_function (`str`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        d_model (`int`, *optional*, defaults to 384):
+            Dimensionality of the layers.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        scale_embedding (`bool`, *optional*, defaults to False):
+            Scale embeddings by diving by sqrt(d_model).
+        max_source_positions (`int`, *optional*, defaults to 1500):
+            The maximum sequence length of log-mel filter-bank features that this model might ever be used with.
+        max_target_positions (`int`, *optional*, defaults to 448):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        pad_token_id (`int`, *optional*, defaults to 50256):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 50256):
+            Begin of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 50256):
+            End of stream token id.
+        suppress_tokens (`List[int]`, *optional*):
+            A list containing the non-speech tokens that will be used by the logit processor in the `generate`
+            function. NON_SPEECH_TOKENS and NON_SPEECH_TOKENS_MULTI each correspond to the `english-only` and the
+            `multilingual` model.
+        begin_suppress_tokens (`List[int]`, *optional*, defaults to `[220,50256]`):
+            A list containing tokens that will be supressed at the beginning of the sampling process. Initialized as
+            the token for `" "` (`blank_token_id`) and the `eos_token_id`
+        use_weighted_layer_sum (`bool`, *optional*, defaults to `False`):
+            Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an
+            instance of [`WhisperForAudioClassification`].
+        classifier_proj_size (`int`, *optional*, defaults to 256):
+            Dimensionality of the projection before token mean-pooling for classification. Only relevant when using an
+            instance of [`WhisperForAudioClassification`].
+        apply_spec_augment (`bool`, *optional*, defaults to `False`):
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
+            [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
+            Recognition](https://arxiv.org/abs/1904.08779).
+        mask_time_prob (`float`, *optional*, defaults to 0.05):
+            Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking
+            procecure generates `mask_time_prob*len(time_axis)/mask_time_length` independent masks over the axis. If
+            reasoning from the propability of each feature vector to be chosen as the start of the vector span to be
+            masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the
+            actual percentage of masked vectors. This is only relevant if `apply_spec_augment == True`.
+        mask_time_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the time axis.
+        mask_time_min_masks (`int`, *optional*, defaults to 2),:
+            The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,
+            irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <
+            mask_time_min_masks''
+        mask_feature_prob (`float`, *optional*, defaults to 0.0):
+            Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The
+            masking procecure generates `mask_feature_prob*len(feature_axis)/mask_time_length` independent masks over
+            the axis. If reasoning from the propability of each feature vector to be chosen as the start of the vector
+            span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap
+            may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is
+            True`.
+        mask_feature_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the feature axis.
+        mask_feature_min_masks (`int`, *optional*, defaults to 0),:
+            The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time
+            step, irrespectively of `mask_feature_prob`. Only relevant if
+            `mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks`.
+        median_filter_width (`int`, *optional*, defaults to 7):
+            Width of the median filter used to smoothen to cross-attention outputs when computing token timestamps.
+            Should be an odd number.
+
+    Example:
+
+    ```python
+    >>> from transformers import WhisperConfig, WhisperModel
+
+    >>> # Initializing a Whisper tiny style configuration
+    >>> configuration = WhisperConfig()
+
+    >>> # Initializing a model (with random weights) from the tiny style configuration
+    >>> model = WhisperModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "whisper"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_key_value_heads": "encoder_attention_heads",
+        "num_attention_heads": "encoder_attention_heads",
+        "hidden_size": "d_model",
+    }
+
+    def __init__(
+        self,
+        vocab_size=51865,
+        num_mel_bins=80,
+        encoder_layers=4,
+        encoder_attention_heads=6,
+        decoder_layers=4,
+        decoder_attention_heads=6,
+        decoder_ffn_dim=1536,
+        encoder_ffn_dim=1536,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        decoder_start_token_id=50257,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=384,
+        dropout=0.0,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        scale_embedding=False,
+        max_source_positions=1500,
+        max_target_positions=448,
+        pad_token_id=50256,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        suppress_tokens=None,
+        begin_suppress_tokens=[220, 50256],
+        use_weighted_layer_sum=False,
+        classifier_proj_size=256,
+        apply_spec_augment=False,
+        mask_time_prob=0.05,
+        mask_time_length=10,
+        mask_time_min_masks=2,
+        mask_feature_prob=0.0,
+        mask_feature_length=10,
+        mask_feature_min_masks=0,
+        median_filter_width=7,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.num_mel_bins = num_mel_bins
+        self.d_model = d_model
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+        self.max_source_positions = max_source_positions
+        self.max_target_positions = max_target_positions
+
+        # Audio Classification-specific parameters. Feel free to ignore for other classes.
+        self.classifier_proj_size = classifier_proj_size
+        self.use_weighted_layer_sum = use_weighted_layer_sum
+
+        # fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
+        self.apply_spec_augment = apply_spec_augment
+        self.mask_time_prob = mask_time_prob
+        self.mask_time_length = mask_time_length
+        self.mask_time_min_masks = mask_time_min_masks
+        self.mask_feature_prob = mask_feature_prob
+        self.mask_feature_length = mask_feature_length
+        self.mask_feature_min_masks = mask_feature_min_masks
+
+        self.median_filter_width = median_filter_width
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            suppress_tokens=suppress_tokens,
+            begin_suppress_tokens=begin_suppress_tokens,
+            **kwargs,
+        )
+
+
+class WhisperOnnxConfig(OnnxSeq2SeqConfigWithPast):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_inputs = OrderedDict(
+            [
+                ("input_features", {0: "batch", 1: "feature_size", 2: "encoder_sequence"}),
+            ]
+        )
+        if self.use_past:
+            common_inputs["decoder_input_ids"] = {0: "batch"}
+        else:
+            common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
+
+        if self.use_past:
+            self.fill_with_past_key_values_(common_inputs, direction="inputs")
+
+        return common_inputs
+
+    def generate_dummy_inputs(
+        self,
+        preprocessor: Union["PreTrainedTokenizerBase", "FeatureExtractionMixin"],
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional["TensorType"] = None,
+        sampling_rate: int = 22050,
+        time_duration: float = 5.0,
+        frequency: int = 220,
+    ) -> Mapping[str, Any]:
+        dummy_inputs = OrderedDict()
+        encoder_inputs = OnnxConfig.generate_dummy_inputs(
+            self,
+            preprocessor=preprocessor.feature_extractor,
+            batch_size=batch_size,
+            framework=framework,
+            sampling_rate=sampling_rate,
+            time_duration=time_duration,
+            frequency=frequency,
+        )
+        encoder_sequence_length = encoder_inputs["input_features"].shape[2]
+        seq_length = encoder_sequence_length // 2 if self.use_past else seq_length
+
+        decoder_inputs = super().generate_dummy_inputs(
+            preprocessor.tokenizer, batch_size, seq_length, is_pair, framework
+        )
+
+        dummy_inputs["input_features"] = encoder_inputs.pop("input_features")
+        dummy_inputs["decoder_input_ids"] = decoder_inputs.pop("decoder_input_ids")
+
+        if "past_key_values" in decoder_inputs:
+            dummy_inputs["past_key_values"] = decoder_inputs.pop("past_key_values")
+
+        return dummy_inputs
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-3

configuration_whisper.py

@@ -0,0 +1,344 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Whisper model configuration"""
+
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig, OnnxSeq2SeqConfigWithPast
+from ...utils import logging
+
+
+if TYPE_CHECKING:
+    from ...feature_extraction_utils import FeatureExtractionMixin
+    from ...tokenization_utils_base import PreTrainedTokenizerBase
+    from ...utils import TensorType
+
+logger = logging.get_logger(__name__)
+
+
+# fmt: off
+NON_SPEECH_TOKENS = [
+    1, 2, 7, 8, 9, 10, 14, 25,
+    26, 27, 28, 29, 31, 58, 59, 60, 61, 62,
+    63, 90, 91, 92, 93, 357, 366, 438, 532, 685,
+    705, 796, 930, 1058, 1220, 1267, 1279, 1303, 1343, 1377,
+    1391, 1635, 1782, 1875, 2162, 2361, 2488, 3467, 4008, 4211,
+    4600, 4808, 5299, 5855, 6329, 7203, 9609, 9959, 10563, 10786,
+    11420, 11709, 11907, 13163, 13697, 13700, 14808, 15306, 16410, 16791,
+    17992, 19203, 19510, 20724, 22305, 22935, 27007, 30109, 30420, 33409,
+    34949, 40283, 40493, 40549, 47282, 49146, 50257, 50359, 50360, 50361
+]
+NON_SPEECH_TOKENS_MULTI = [
+    1, 2, 7, 8, 9, 10, 14, 25,
+    26, 27, 28, 29, 31, 58, 59, 60, 61, 62,
+    63, 90, 91, 92, 93, 359, 503, 522, 542, 873,
+    893, 902, 918, 922, 931, 1350, 1853, 1982, 2460, 2627,
+    3246, 3253, 3268, 3536, 3846, 3961, 4183, 4667, 6585, 6647,
+    7273, 9061, 9383, 10428, 10929, 11938, 12033, 12331, 12562, 13793,
+    14157, 14635, 15265, 15618, 16553, 16604, 18362, 18956, 20075, 21675,
+    22520, 26130, 26161, 26435, 28279, 29464, 31650, 32302, 32470, 36865,
+    42863, 47425, 49870, 50254, 50258, 50360, 50361, 50362
+]
+# fmt: on
+
+
+class WhisperConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`WhisperModel`]. It is used to instantiate a
+    Whisper model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Whisper
+    [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 51865):
+            Vocabulary size of the Whisper model. Defines the number of different tokens that can be represented by the
+            `decoder_input_ids` passed when calling [`WhisperModel`]
+        num_mel_bins (`int`, *optional*, defaults to 80):
+            Number of mel features used per input features. Should correspond to the value used in the
+            `WhisperProcessor` class.
+        encoder_layers (`int`, *optional*, defaults to 4):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 4):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 6):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 6):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 1536):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in encoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 1536):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_start_token_id (`int`, *optional*, defaults to 50257):
+            Corresponds to the "<|startoftranscript|>" token, which is automatically used when no `decoder_input_ids`
+            are provided to the `generate` function. It is used to guide the model`s generation process depending on
+            the task.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        is_encoder_decoder (`bool`, *optional*, defaults to `True`):
+            Whether the model is used as an encoder/decoder or not.
+        activation_function (`str`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        d_model (`int`, *optional*, defaults to 384):
+            Dimensionality of the layers.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        scale_embedding (`bool`, *optional*, defaults to False):
+            Scale embeddings by diving by sqrt(d_model).
+        max_source_positions (`int`, *optional*, defaults to 1500):
+            The maximum sequence length of log-mel filter-bank features that this model might ever be used with.
+        max_target_positions (`int`, *optional*, defaults to 448):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        pad_token_id (`int`, *optional*, defaults to 50256):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 50256):
+            Begin of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 50256):
+            End of stream token id.
+        suppress_tokens (`List[int]`, *optional*):
+            A list containing the non-speech tokens that will be used by the logit processor in the `generate`
+            function. NON_SPEECH_TOKENS and NON_SPEECH_TOKENS_MULTI each correspond to the `english-only` and the
+            `multilingual` model.
+        begin_suppress_tokens (`List[int]`, *optional*, defaults to `[220,50256]`):
+            A list containing tokens that will be supressed at the beginning of the sampling process. Initialized as
+            the token for `" "` (`blank_token_id`) and the `eos_token_id`
+        use_weighted_layer_sum (`bool`, *optional*, defaults to `False`):
+            Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an
+            instance of [`WhisperForAudioClassification`].
+        classifier_proj_size (`int`, *optional*, defaults to 256):
+            Dimensionality of the projection before token mean-pooling for classification. Only relevant when using an
+            instance of [`WhisperForAudioClassification`].
+        apply_spec_augment (`bool`, *optional*, defaults to `False`):
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
+            [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
+            Recognition](https://arxiv.org/abs/1904.08779).
+        mask_time_prob (`float`, *optional*, defaults to 0.05):
+            Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking
+            procecure generates `mask_time_prob*len(time_axis)/mask_time_length` independent masks over the axis. If
+            reasoning from the propability of each feature vector to be chosen as the start of the vector span to be
+            masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the
+            actual percentage of masked vectors. This is only relevant if `apply_spec_augment == True`.
+        mask_time_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the time axis.
+        mask_time_min_masks (`int`, *optional*, defaults to 2),:
+            The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,
+            irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <
+            mask_time_min_masks''
+        mask_feature_prob (`float`, *optional*, defaults to 0.0):
+            Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The
+            masking procecure generates `mask_feature_prob*len(feature_axis)/mask_time_length` independent masks over
+            the axis. If reasoning from the propability of each feature vector to be chosen as the start of the vector
+            span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap
+            may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is
+            True`.
+        mask_feature_length (`int`, *optional*, defaults to 10):
+            Length of vector span along the feature axis.
+        mask_feature_min_masks (`int`, *optional*, defaults to 0),:
+            The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time
+            step, irrespectively of `mask_feature_prob`. Only relevant if
+            `mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks`.
+        median_filter_width (`int`, *optional*, defaults to 7):
+            Width of the median filter used to smoothen to cross-attention outputs when computing token timestamps.
+            Should be an odd number.
+
+    Example:
+
+    ```python
+    >>> from transformers import WhisperConfig, WhisperModel
+
+    >>> # Initializing a Whisper tiny style configuration
+    >>> configuration = WhisperConfig()
+
+    >>> # Initializing a model (with random weights) from the tiny style configuration
+    >>> model = WhisperModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "whisper"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_key_value_heads": "encoder_attention_heads",
+        "num_attention_heads": "encoder_attention_heads",
+        "hidden_size": "d_model",
+    }
+
+    def __init__(
+        self,
+        vocab_size=51865,
+        num_mel_bins=80,
+        encoder_layers=4,
+        encoder_attention_heads=6,
+        decoder_layers=4,
+        decoder_attention_heads=6,
+        decoder_ffn_dim=1536,
+        encoder_ffn_dim=1536,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        decoder_start_token_id=50257,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=384,
+        dropout=0.0,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        scale_embedding=False,
+        max_source_positions=1500,
+        max_target_positions=448,
+        pad_token_id=50256,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        suppress_tokens=None,
+        begin_suppress_tokens=[220, 50256],
+        use_weighted_layer_sum=False,
+        classifier_proj_size=256,
+        apply_spec_augment=False,
+        mask_time_prob=0.05,
+        mask_time_length=10,
+        mask_time_min_masks=2,
+        mask_feature_prob=0.0,
+        mask_feature_length=10,
+        mask_feature_min_masks=0,
+        median_filter_width=7,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.num_mel_bins = num_mel_bins
+        self.d_model = d_model
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+        self.max_source_positions = max_source_positions
+        self.max_target_positions = max_target_positions
+
+        # Audio Classification-specific parameters. Feel free to ignore for other classes.
+        self.classifier_proj_size = classifier_proj_size
+        self.use_weighted_layer_sum = use_weighted_layer_sum
+
+        # fine-tuning config parameters for SpecAugment: https://arxiv.org/abs/1904.08779
+        self.apply_spec_augment = apply_spec_augment
+        self.mask_time_prob = mask_time_prob
+        self.mask_time_length = mask_time_length
+        self.mask_time_min_masks = mask_time_min_masks
+        self.mask_feature_prob = mask_feature_prob
+        self.mask_feature_length = mask_feature_length
+        self.mask_feature_min_masks = mask_feature_min_masks
+
+        self.median_filter_width = median_filter_width
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            suppress_tokens=suppress_tokens,
+            begin_suppress_tokens=begin_suppress_tokens,
+            **kwargs,
+        )
+
+
+class WhisperOnnxConfig(OnnxSeq2SeqConfigWithPast):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_inputs = OrderedDict(
+            [
+                ("input_features", {0: "batch", 1: "feature_size", 2: "encoder_sequence"}),
+            ]
+        )
+        if self.use_past:
+            common_inputs["decoder_input_ids"] = {0: "batch"}
+        else:
+            common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
+
+        if self.use_past:
+            self.fill_with_past_key_values_(common_inputs, direction="inputs")
+
+        return common_inputs
+
+    def generate_dummy_inputs(
+        self,
+        preprocessor: Union["PreTrainedTokenizerBase", "FeatureExtractionMixin"],
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional["TensorType"] = None,
+        sampling_rate: int = 22050,
+        time_duration: float = 5.0,
+        frequency: int = 220,
+    ) -> Mapping[str, Any]:
+        dummy_inputs = OrderedDict()
+        encoder_inputs = OnnxConfig.generate_dummy_inputs(
+            self,
+            preprocessor=preprocessor.feature_extractor,
+            batch_size=batch_size,
+            framework=framework,
+            sampling_rate=sampling_rate,
+            time_duration=time_duration,
+            frequency=frequency,
+        )
+        encoder_sequence_length = encoder_inputs["input_features"].shape[2]
+        seq_length = encoder_sequence_length // 2 if self.use_past else seq_length
+
+        decoder_inputs = super().generate_dummy_inputs(
+            preprocessor.tokenizer, batch_size, seq_length, is_pair, framework
+        )
+
+        dummy_inputs["input_features"] = encoder_inputs.pop("input_features")
+        dummy_inputs["decoder_input_ids"] = decoder_inputs.pop("decoder_input_ids")
+
+        if "past_key_values" in decoder_inputs:
+            dummy_inputs["past_key_values"] = decoder_inputs.pop("past_key_values")
+
+        return dummy_inputs
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-3

feature_extraction_whisper (1).py

@@ -0,0 +1,324 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Feature extractor class for Whisper
+"""
+
+from typing import List, Optional, Union
+
+import numpy as np
+
+from ... import is_torch_available
+from ...audio_utils import mel_filter_bank, spectrogram, window_function
+from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
+from ...feature_extraction_utils import BatchFeature
+from ...utils import TensorType, logging
+
+
+if is_torch_available():
+    import torch
+
+logger = logging.get_logger(__name__)
+
+
+class WhisperFeatureExtractor(SequenceFeatureExtractor):
+    r"""
+    Constructs a Whisper feature extractor.
+
+    This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
+    most of the main methods. Users should refer to this superclass for more information regarding those methods.
+
+    This class extracts mel-filter bank features from raw speech using a custom numpy implementation of the `Short Time
+    Fourier Transform` which should match pytorch's `torch.stft` equivalent.
+
+    Args:
+        feature_size (`int`, *optional*, defaults to 80):
+            The feature dimension of the extracted features.
+        sampling_rate (`int`, *optional*, defaults to 16000):
+            The sampling rate at which the audio files should be digitalized expressed in hertz (Hz).
+        hop_length (`int`, *optional*, defaults to 160):
+            Length of the overlaping windows for the STFT used to obtain the Mel Frequency coefficients.
+        chunk_length (`int`, *optional*, defaults to 30):
+            The maximum number of chuncks of `sampling_rate` samples used to trim and pad longer or shorter audio
+            sequences.
+        n_fft (`int`, *optional*, defaults to 400):
+            Size of the Fourier transform.
+        padding_value (`float`, *optional*, defaults to 0.0):
+            Padding value used to pad the audio. Should correspond to silences.
+    """
+
+    model_input_names = ["input_features"]
+
+    def __init__(
+        self,
+        feature_size=80,
+        sampling_rate=16000,
+        hop_length=160,
+        chunk_length=30,
+        n_fft=400,
+        padding_value=0.0,
+        return_attention_mask=False,  # pad inputs to max length with silence token (zero) and no attention mask
+        **kwargs,
+    ):
+        super().__init__(
+            feature_size=feature_size,
+            sampling_rate=sampling_rate,
+            padding_value=padding_value,
+            return_attention_mask=return_attention_mask,
+            **kwargs,
+        )
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.chunk_length = chunk_length
+        self.n_samples = chunk_length * sampling_rate
+        self.nb_max_frames = self.n_samples // hop_length
+        self.sampling_rate = sampling_rate
+        self.mel_filters = mel_filter_bank(
+            num_frequency_bins=1 + n_fft // 2,
+            num_mel_filters=feature_size,
+            min_frequency=0.0,
+            max_frequency=8000.0,
+            sampling_rate=sampling_rate,
+            norm="slaney",
+            mel_scale="slaney",
+        )
+
+    def _np_extract_fbank_features(self, waveform_batch: np.array, device: str) -> np.ndarray:
+        """
+        Compute the log-mel spectrogram of the provided audio, gives similar results to Whisper's original torch
+        implementation with 1e-5 tolerance.
+        """
+        if device != "cpu":
+            raise ValueError(
+                f"Got device `{device}` for feature extraction, but feature extraction on CUDA accelerator "
+                "devices requires torch, which is not installed. Either set `device='cpu'`, or "
+                "install torch according to the official instructions: https://pytorch.org/get-started/locally/"
+            )
+        log_spec_batch = []
+        for waveform in waveform_batch:
+            log_spec = spectrogram(
+                waveform,
+                window_function(self.n_fft, "hann"),
+                frame_length=self.n_fft,
+                hop_length=self.hop_length,
+                power=2.0,
+                mel_filters=self.mel_filters,
+                log_mel="log10",
+            )
+            log_spec = log_spec[:, :-1]
+            log_spec = np.maximum(log_spec, log_spec.max() - 8.0)
+            log_spec = (log_spec + 4.0) / 4.0
+            log_spec_batch.append(log_spec)
+        log_spec_batch = np.array(log_spec_batch)
+        return log_spec_batch
+
+    def _torch_extract_fbank_features(self, waveform: np.array, device: str = "cpu") -> np.ndarray:
+        """
+        Compute the log-mel spectrogram of the audio using PyTorch's GPU-accelerated STFT implementation with batching,
+        yielding results similar to cpu computing with 1e-5 tolerance.
+        """
+        waveform = torch.from_numpy(waveform).type(torch.float32)
+
+        window = torch.hann_window(self.n_fft)
+        if device != "cpu":
+            waveform = waveform.to(device)
+            window = window.to(device)
+        stft = torch.stft(waveform, self.n_fft, self.hop_length, window=window, return_complex=True)
+        magnitudes = stft[..., :-1].abs() ** 2
+
+        mel_filters = torch.from_numpy(self.mel_filters).type(torch.float32)
+        if device != "cpu":
+            mel_filters = mel_filters.to(device)
+        mel_spec = mel_filters.T @ magnitudes
+
+        log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+        if waveform.dim() == 2:
+            max_val = log_spec.max(dim=2, keepdim=True)[0].max(dim=1, keepdim=True)[0]
+            log_spec = torch.maximum(log_spec, max_val - 8.0)
+        else:
+            log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+        log_spec = (log_spec + 4.0) / 4.0
+        if device != "cpu":
+            log_spec = log_spec.detach().cpu()
+        return log_spec.numpy()
+
+    @staticmethod
+    # Copied from transformers.models.wav2vec2.feature_extraction_wav2vec2.Wav2Vec2FeatureExtractor.zero_mean_unit_var_norm
+    def zero_mean_unit_var_norm(
+        input_values: List[np.ndarray], attention_mask: List[np.ndarray], padding_value: float = 0.0
+    ) -> List[np.ndarray]:
+        """
+        Every array in the list is normalized to have zero mean and unit variance
+        """
+        if attention_mask is not None:
+            attention_mask = np.array(attention_mask, np.int32)
+            normed_input_values = []
+
+            for vector, length in zip(input_values, attention_mask.sum(-1)):
+                normed_slice = (vector - vector[:length].mean()) / np.sqrt(vector[:length].var() + 1e-7)
+                if length < normed_slice.shape[0]:
+                    normed_slice[length:] = padding_value
+
+                normed_input_values.append(normed_slice)
+        else:
+            normed_input_values = [(x - x.mean()) / np.sqrt(x.var() + 1e-7) for x in input_values]
+
+        return normed_input_values
+
+    def __call__(
+        self,
+        raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
+        truncation: bool = True,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        return_attention_mask: Optional[bool] = None,
+        padding: Optional[str] = "max_length",
+        max_length: Optional[int] = None,
+        sampling_rate: Optional[int] = None,
+        do_normalize: Optional[bool] = None,
+        device: Optional[str] = "cpu",
+        return_token_timestamps: Optional[bool] = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Main method to featurize and prepare for the model one or several sequence(s). Implementation uses PyTorch for
+        the STFT computation if available, otherwise a slower NumPy based one.
+
+        Args:
+            raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
+                The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
+                values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not
+                stereo, i.e. single float per timestep.
+            truncation (`bool`, *optional*, default to `True`):
+                Activates truncation to cut input sequences longer than *max_length* to *max_length*.
+            pad_to_multiple_of (`int`, *optional*, defaults to None):
+                If set will pad the sequence to a multiple of the provided value.
+
+                This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
+                `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128.
+            return_attention_mask (`bool`, *optional*):
+                Whether to return the attention mask. If left to the default, will return the attention mask according
+                to the specific feature_extractor's default.
+
+                [What are attention masks?](../glossary#attention-mask)
+
+                <Tip>
+
+                For Whisper models, `attention_mask` should always be passed for batched inference, to avoid subtle
+                bugs.
+
+                </Tip>
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors instead of list of python integers. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return Numpy `np.ndarray` objects.
+            sampling_rate (`int`, *optional*):
+                The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
+                `sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition
+                pipeline.
+            padding_value (`float`, *optional*, defaults to 0.0):
+                The value that is used to fill the padding values / vectors.
+            do_normalize (`bool`, *optional*, defaults to `False`):
+                Whether or not to zero-mean unit-variance normalize the input. Normalizing can help to significantly
+                improve the performance of the model.
+            device (`str`, *optional*, defaults to `'cpu'`):
+                Specifies the device for computation of the log-mel spectrogram of audio signals in the
+                `_torch_extract_fbank_features` method. (e.g., "cpu", "cuda")
+            return_token_timestamps (`bool`, *optional*, defaults to `None`):
+                Whether or not to return the number of frames of the input raw_speech.
+                These num_frames can be used by the model to compute word level timestamps.
+        """
+
+        if sampling_rate is not None:
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(
+                    f"The model corresponding to this feature extractor: {self.__class__.__name__} was trained using a"
+                    f" sampling rate of {self.sampling_rate}. Please make sure that the provided `raw_speech` input"
+                    f" was sampled with {self.sampling_rate} and not {sampling_rate}."
+                )
+        else:
+            logger.warning(
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
+                "Failing to do so can result in silent errors that might be hard to debug."
+            )
+
+        is_batched_numpy = isinstance(raw_speech, np.ndarray) and len(raw_speech.shape) > 1
+        if is_batched_numpy and len(raw_speech.shape) > 2:
+            raise ValueError(f"Only mono-channel audio is supported for input to {self}")
+        is_batched = is_batched_numpy or (
+            isinstance(raw_speech, (list, tuple)) and (isinstance(raw_speech[0], (np.ndarray, tuple, list)))
+        )
+
+        if is_batched:
+            raw_speech = [np.asarray([speech], dtype=np.float32).T for speech in raw_speech]
+        elif not is_batched and not isinstance(raw_speech, np.ndarray):
+            raw_speech = np.asarray(raw_speech, dtype=np.float32)
+        elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
+            raw_speech = raw_speech.astype(np.float32)
+
+        # always return batch
+        if not is_batched:
+            raw_speech = [np.asarray([raw_speech]).T]
+
+        batched_speech = BatchFeature({"input_features": raw_speech})
+
+        # convert into correct format for padding
+
+        padded_inputs = self.pad(
+            batched_speech,
+            padding=padding,
+            max_length=max_length if max_length else self.n_samples,
+            truncation=truncation,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask or do_normalize,
+        )
+
+        # zero-mean and unit-variance normalization
+        if do_normalize:
+            padded_inputs["input_features"] = self.zero_mean_unit_var_norm(
+                padded_inputs["input_features"],
+                attention_mask=padded_inputs["attention_mask"],
+                padding_value=self.padding_value,
+            )
+            padded_inputs["input_features"] = np.stack(padded_inputs["input_features"], axis=0)
+
+        # make sure list is in array format
+        input_features = padded_inputs.get("input_features").transpose(2, 0, 1)
+
+        extract_fbank_features = (
+            self._torch_extract_fbank_features if is_torch_available() else self._np_extract_fbank_features
+        )
+        input_features = extract_fbank_features(input_features[0], device)
+
+        if isinstance(input_features[0], List):
+            padded_inputs["input_features"] = [np.asarray(feature, dtype=np.float32) for feature in input_features]
+
+        else:
+            padded_inputs["input_features"] = input_features
+
+        if return_attention_mask:
+            # rescale from sample (48000) to feature (3000)
+            padded_inputs["attention_mask"] = padded_inputs["attention_mask"][:, :: self.hop_length]
+
+        if return_token_timestamps is not None:
+            padded_inputs["num_frames"] = [len(raw_speech_i) // self.hop_length for raw_speech_i in raw_speech]
+
+        if return_tensors is not None:
+            padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
+
+        return padded_inputs

feature_extraction_whisper.py

@@ -0,0 +1,324 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Feature extractor class for Whisper
+"""
+
+from typing import List, Optional, Union
+
+import numpy as np
+
+from ... import is_torch_available
+from ...audio_utils import mel_filter_bank, spectrogram, window_function
+from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
+from ...feature_extraction_utils import BatchFeature
+from ...utils import TensorType, logging
+
+
+if is_torch_available():
+    import torch
+
+logger = logging.get_logger(__name__)
+
+
+class WhisperFeatureExtractor(SequenceFeatureExtractor):
+    r"""
+    Constructs a Whisper feature extractor.
+
+    This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
+    most of the main methods. Users should refer to this superclass for more information regarding those methods.
+
+    This class extracts mel-filter bank features from raw speech using a custom numpy implementation of the `Short Time
+    Fourier Transform` which should match pytorch's `torch.stft` equivalent.
+
+    Args:
+        feature_size (`int`, *optional*, defaults to 80):
+            The feature dimension of the extracted features.
+        sampling_rate (`int`, *optional*, defaults to 16000):
+            The sampling rate at which the audio files should be digitalized expressed in hertz (Hz).
+        hop_length (`int`, *optional*, defaults to 160):
+            Length of the overlaping windows for the STFT used to obtain the Mel Frequency coefficients.
+        chunk_length (`int`, *optional*, defaults to 30):
+            The maximum number of chuncks of `sampling_rate` samples used to trim and pad longer or shorter audio
+            sequences.
+        n_fft (`int`, *optional*, defaults to 400):
+            Size of the Fourier transform.
+        padding_value (`float`, *optional*, defaults to 0.0):
+            Padding value used to pad the audio. Should correspond to silences.
+    """
+
+    model_input_names = ["input_features"]
+
+    def __init__(
+        self,
+        feature_size=80,
+        sampling_rate=16000,
+        hop_length=160,
+        chunk_length=30,
+        n_fft=400,
+        padding_value=0.0,
+        return_attention_mask=False,  # pad inputs to max length with silence token (zero) and no attention mask
+        **kwargs,
+    ):
+        super().__init__(
+            feature_size=feature_size,
+            sampling_rate=sampling_rate,
+            padding_value=padding_value,
+            return_attention_mask=return_attention_mask,
+            **kwargs,
+        )
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.chunk_length = chunk_length
+        self.n_samples = chunk_length * sampling_rate
+        self.nb_max_frames = self.n_samples // hop_length
+        self.sampling_rate = sampling_rate
+        self.mel_filters = mel_filter_bank(
+            num_frequency_bins=1 + n_fft // 2,
+            num_mel_filters=feature_size,
+            min_frequency=0.0,
+            max_frequency=8000.0,
+            sampling_rate=sampling_rate,
+            norm="slaney",
+            mel_scale="slaney",
+        )
+
+    def _np_extract_fbank_features(self, waveform_batch: np.array, device: str) -> np.ndarray:
+        """
+        Compute the log-mel spectrogram of the provided audio, gives similar results to Whisper's original torch
+        implementation with 1e-5 tolerance.
+        """
+        if device != "cpu":
+            raise ValueError(
+                f"Got device `{device}` for feature extraction, but feature extraction on CUDA accelerator "
+                "devices requires torch, which is not installed. Either set `device='cpu'`, or "
+                "install torch according to the official instructions: https://pytorch.org/get-started/locally/"
+            )
+        log_spec_batch = []
+        for waveform in waveform_batch:
+            log_spec = spectrogram(
+                waveform,
+                window_function(self.n_fft, "hann"),
+                frame_length=self.n_fft,
+                hop_length=self.hop_length,
+                power=2.0,
+                mel_filters=self.mel_filters,
+                log_mel="log10",
+            )
+            log_spec = log_spec[:, :-1]
+            log_spec = np.maximum(log_spec, log_spec.max() - 8.0)
+            log_spec = (log_spec + 4.0) / 4.0
+            log_spec_batch.append(log_spec)
+        log_spec_batch = np.array(log_spec_batch)
+        return log_spec_batch
+
+    def _torch_extract_fbank_features(self, waveform: np.array, device: str = "cpu") -> np.ndarray:
+        """
+        Compute the log-mel spectrogram of the audio using PyTorch's GPU-accelerated STFT implementation with batching,
+        yielding results similar to cpu computing with 1e-5 tolerance.
+        """
+        waveform = torch.from_numpy(waveform).type(torch.float32)
+
+        window = torch.hann_window(self.n_fft)
+        if device != "cpu":
+            waveform = waveform.to(device)
+            window = window.to(device)
+        stft = torch.stft(waveform, self.n_fft, self.hop_length, window=window, return_complex=True)
+        magnitudes = stft[..., :-1].abs() ** 2
+
+        mel_filters = torch.from_numpy(self.mel_filters).type(torch.float32)
+        if device != "cpu":
+            mel_filters = mel_filters.to(device)
+        mel_spec = mel_filters.T @ magnitudes
+
+        log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+        if waveform.dim() == 2:
+            max_val = log_spec.max(dim=2, keepdim=True)[0].max(dim=1, keepdim=True)[0]
+            log_spec = torch.maximum(log_spec, max_val - 8.0)
+        else:
+            log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+        log_spec = (log_spec + 4.0) / 4.0
+        if device != "cpu":
+            log_spec = log_spec.detach().cpu()
+        return log_spec.numpy()
+
+    @staticmethod
+    # Copied from transformers.models.wav2vec2.feature_extraction_wav2vec2.Wav2Vec2FeatureExtractor.zero_mean_unit_var_norm
+    def zero_mean_unit_var_norm(
+        input_values: List[np.ndarray], attention_mask: List[np.ndarray], padding_value: float = 0.0
+    ) -> List[np.ndarray]:
+        """
+        Every array in the list is normalized to have zero mean and unit variance
+        """
+        if attention_mask is not None:
+            attention_mask = np.array(attention_mask, np.int32)
+            normed_input_values = []
+
+            for vector, length in zip(input_values, attention_mask.sum(-1)):
+                normed_slice = (vector - vector[:length].mean()) / np.sqrt(vector[:length].var() + 1e-7)
+                if length < normed_slice.shape[0]:
+                    normed_slice[length:] = padding_value
+
+                normed_input_values.append(normed_slice)
+        else:
+            normed_input_values = [(x - x.mean()) / np.sqrt(x.var() + 1e-7) for x in input_values]
+
+        return normed_input_values
+
+    def __call__(
+        self,
+        raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
+        truncation: bool = True,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        return_attention_mask: Optional[bool] = None,
+        padding: Optional[str] = "max_length",
+        max_length: Optional[int] = None,
+        sampling_rate: Optional[int] = None,
+        do_normalize: Optional[bool] = None,
+        device: Optional[str] = "cpu",
+        return_token_timestamps: Optional[bool] = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Main method to featurize and prepare for the model one or several sequence(s). Implementation uses PyTorch for
+        the STFT computation if available, otherwise a slower NumPy based one.
+
+        Args:
+            raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
+                The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
+                values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not
+                stereo, i.e. single float per timestep.
+            truncation (`bool`, *optional*, default to `True`):
+                Activates truncation to cut input sequences longer than *max_length* to *max_length*.
+            pad_to_multiple_of (`int`, *optional*, defaults to None):
+                If set will pad the sequence to a multiple of the provided value.
+
+                This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
+                `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128.
+            return_attention_mask (`bool`, *optional*):
+                Whether to return the attention mask. If left to the default, will return the attention mask according
+                to the specific feature_extractor's default.
+
+                [What are attention masks?](../glossary#attention-mask)
+
+                <Tip>
+
+                For Whisper models, `attention_mask` should always be passed for batched inference, to avoid subtle
+                bugs.
+
+                </Tip>
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors instead of list of python integers. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return Numpy `np.ndarray` objects.
+            sampling_rate (`int`, *optional*):
+                The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
+                `sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition
+                pipeline.
+            padding_value (`float`, *optional*, defaults to 0.0):
+                The value that is used to fill the padding values / vectors.
+            do_normalize (`bool`, *optional*, defaults to `False`):
+                Whether or not to zero-mean unit-variance normalize the input. Normalizing can help to significantly
+                improve the performance of the model.
+            device (`str`, *optional*, defaults to `'cpu'`):
+                Specifies the device for computation of the log-mel spectrogram of audio signals in the
+                `_torch_extract_fbank_features` method. (e.g., "cpu", "cuda")
+            return_token_timestamps (`bool`, *optional*, defaults to `None`):
+                Whether or not to return the number of frames of the input raw_speech.
+                These num_frames can be used by the model to compute word level timestamps.
+        """
+
+        if sampling_rate is not None:
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(
+                    f"The model corresponding to this feature extractor: {self.__class__.__name__} was trained using a"
+                    f" sampling rate of {self.sampling_rate}. Please make sure that the provided `raw_speech` input"
+                    f" was sampled with {self.sampling_rate} and not {sampling_rate}."
+                )
+        else:
+            logger.warning(
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
+                "Failing to do so can result in silent errors that might be hard to debug."
+            )
+
+        is_batched_numpy = isinstance(raw_speech, np.ndarray) and len(raw_speech.shape) > 1
+        if is_batched_numpy and len(raw_speech.shape) > 2:
+            raise ValueError(f"Only mono-channel audio is supported for input to {self}")
+        is_batched = is_batched_numpy or (
+            isinstance(raw_speech, (list, tuple)) and (isinstance(raw_speech[0], (np.ndarray, tuple, list)))
+        )
+
+        if is_batched:
+            raw_speech = [np.asarray([speech], dtype=np.float32).T for speech in raw_speech]
+        elif not is_batched and not isinstance(raw_speech, np.ndarray):
+            raw_speech = np.asarray(raw_speech, dtype=np.float32)
+        elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
+            raw_speech = raw_speech.astype(np.float32)
+
+        # always return batch
+        if not is_batched:
+            raw_speech = [np.asarray([raw_speech]).T]
+
+        batched_speech = BatchFeature({"input_features": raw_speech})
+
+        # convert into correct format for padding
+
+        padded_inputs = self.pad(
+            batched_speech,
+            padding=padding,
+            max_length=max_length if max_length else self.n_samples,
+            truncation=truncation,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask or do_normalize,
+        )
+
+        # zero-mean and unit-variance normalization
+        if do_normalize:
+            padded_inputs["input_features"] = self.zero_mean_unit_var_norm(
+                padded_inputs["input_features"],
+                attention_mask=padded_inputs["attention_mask"],
+                padding_value=self.padding_value,
+            )
+            padded_inputs["input_features"] = np.stack(padded_inputs["input_features"], axis=0)
+
+        # make sure list is in array format
+        input_features = padded_inputs.get("input_features").transpose(2, 0, 1)
+
+        extract_fbank_features = (
+            self._torch_extract_fbank_features if is_torch_available() else self._np_extract_fbank_features
+        )
+        input_features = extract_fbank_features(input_features[0], device)
+
+        if isinstance(input_features[0], List):
+            padded_inputs["input_features"] = [np.asarray(feature, dtype=np.float32) for feature in input_features]
+
+        else:
+            padded_inputs["input_features"] = input_features
+
+        if return_attention_mask:
+            # rescale from sample (48000) to feature (3000)
+            padded_inputs["attention_mask"] = padded_inputs["attention_mask"][:, :: self.hop_length]
+
+        if return_token_timestamps is not None:
+            padded_inputs["num_frames"] = [len(raw_speech_i) // self.hop_length for raw_speech_i in raw_speech]
+
+        if return_tensors is not None:
+            padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
+
+        return padded_inputs

find-corrupt-whisper-files (1).py

@@ -0,0 +1,81 @@
+#!C:\Python312\python.exe
+# encoding: utf-8
+"""Find and (optionally) delete corrupt Whisper data files"""
+
+import argparse
+import os
+import sys
+import logging
+
+try:
+    import whisper
+except ImportError:
+    raise SystemExit("[ERROR] Please make sure Whisper is installed properly")
+
+
+def setup_logging(verbose=False):
+    """Configure logging."""
+    logging.basicConfig(
+        level=logging.DEBUG if verbose else logging.INFO,
+        format="%(asctime)s [%(levelname)s]: %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    )
+
+
+def walk_dir(base_dir, delete_corrupt=False, backup_corrupt=False):
+    """Walk through directories to find and handle corrupt Whisper files."""
+    total_files = 0
+    corrupt_files = 0
+    deleted_files = 0
+
+    for dirpath, _, filenames in os.walk(base_dir):
+        logging.info("Scanning %s...", dirpath)
+
+        whisper_files = (os.path.join(dirpath, f) for f in filenames if f.endswith(".wsp"))
+        for f in whisper_files:
+            total_files += 1
+            try:
+                info = whisper.info(f)
+                logging.debug("%s: %d points", f, sum(i["points"] for i in info.get("archives", {})))
+            except whisper.CorruptWhisperFile:
+                corrupt_files += 1
+                if backup_corrupt:
+                    backup_path = f + ".bak"
+                    try:
+                        os.rename(f, backup_path)
+                        logging.warning("Backed up corrupt file: %s -> %s", f, backup_path)
+                    except OSError as e:
+                        logging.error("Failed to back up %s: %s", f, e)
+                        continue
+
+                if delete_corrupt:
+                    try:
+                        os.unlink(f)
+                        deleted_files += 1
+                        logging.warning("Deleted corrupt file: %s", f)
+                    except OSError as e:
+                        logging.error("Failed to delete %s: %s", f, e)
+                else:
+                    logging.error("Corrupt Whisper file: %s", f)
+
+    logging.info("Summary: Scanned %d files, Found %d corrupt, Deleted %d", total_files, corrupt_files, deleted_files)
+    return total_files, corrupt_files, deleted_files
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__.strip())
+    parser.add_argument("--delete-corrupt", action="store_true", help="Delete reported corrupt files")
+    parser.add_argument("--backup-corrupt", action="store_true", help="Back up corrupt files before deletion")
+    parser.add_argument("--verbose", action="store_true", help="Display detailed progress")
+    parser.add_argument("directories", type=str, nargs="+", metavar="WHISPER_DIR", help="Directory containing Whisper files")
+    args = parser.parse_args()
+
+    setup_logging(verbose=args.verbose)
+
+    for d in args.directories:
+        d = os.path.realpath(d)
+        if not os.path.isdir(d):
+            logging.error("%s is not a directory!", d)
+            continue
+
+        walk_dir(d, delete_corrupt=args.delete_corrupt, backup_corrupt=args.backup_corrupt)

find-corrupt-whisper-files (2).py

@@ -0,0 +1,81 @@
+#!C:\Python312\python.exe
+# encoding: utf-8
+"""Find and (optionally) delete corrupt Whisper data files"""
+
+import argparse
+import os
+import sys
+import logging
+
+try:
+    import whisper
+except ImportError:
+    raise SystemExit("[ERROR] Please make sure Whisper is installed properly")
+
+
+def setup_logging(verbose=False):
+    """Configure logging."""
+    logging.basicConfig(
+        level=logging.DEBUG if verbose else logging.INFO,
+        format="%(asctime)s [%(levelname)s]: %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    )
+
+
+def walk_dir(base_dir, delete_corrupt=False, backup_corrupt=False):
+    """Walk through directories to find and handle corrupt Whisper files."""
+    total_files = 0
+    corrupt_files = 0
+    deleted_files = 0
+
+    for dirpath, _, filenames in os.walk(base_dir):
+        logging.info("Scanning %s...", dirpath)
+
+        whisper_files = (os.path.join(dirpath, f) for f in filenames if f.endswith(".wsp"))
+        for f in whisper_files:
+            total_files += 1
+            try:
+                info = whisper.info(f)
+                logging.debug("%s: %d points", f, sum(i["points"] for i in info.get("archives", {})))
+            except whisper.CorruptWhisperFile:
+                corrupt_files += 1
+                if backup_corrupt:
+                    backup_path = f + ".bak"
+                    try:
+                        os.rename(f, backup_path)
+                        logging.warning("Backed up corrupt file: %s -> %s", f, backup_path)
+                    except OSError as e:
+                        logging.error("Failed to back up %s: %s", f, e)
+                        continue
+
+                if delete_corrupt:
+                    try:
+                        os.unlink(f)
+                        deleted_files += 1
+                        logging.warning("Deleted corrupt file: %s", f)
+                    except OSError as e:
+                        logging.error("Failed to delete %s: %s", f, e)
+                else:
+                    logging.error("Corrupt Whisper file: %s", f)
+
+    logging.info("Summary: Scanned %d files, Found %d corrupt, Deleted %d", total_files, corrupt_files, deleted_files)
+    return total_files, corrupt_files, deleted_files
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__.strip())
+    parser.add_argument("--delete-corrupt", action="store_true", help="Delete reported corrupt files")
+    parser.add_argument("--backup-corrupt", action="store_true", help="Back up corrupt files before deletion")
+    parser.add_argument("--verbose", action="store_true", help="Display detailed progress")
+    parser.add_argument("directories", type=str, nargs="+", metavar="WHISPER_DIR", help="Directory containing Whisper files")
+    args = parser.parse_args()
+
+    setup_logging(verbose=args.verbose)
+
+    for d in args.directories:
+        d = os.path.realpath(d)
+        if not os.path.isdir(d):
+            logging.error("%s is not a directory!", d)
+            continue
+
+        walk_dir(d, delete_corrupt=args.delete_corrupt, backup_corrupt=args.backup_corrupt)

find-corrupt-whisper-files.py

@@ -0,0 +1,81 @@
+#!C:\Python312\python.exe
+# encoding: utf-8
+"""Find and (optionally) delete corrupt Whisper data files"""
+
+import argparse
+import os
+import sys
+import logging
+
+try:
+    import whisper
+except ImportError:
+    raise SystemExit("[ERROR] Please make sure Whisper is installed properly")
+
+
+def setup_logging(verbose=False):
+    """Configure logging."""
+    logging.basicConfig(
+        level=logging.DEBUG if verbose else logging.INFO,
+        format="%(asctime)s [%(levelname)s]: %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    )
+
+
+def walk_dir(base_dir, delete_corrupt=False, backup_corrupt=False):
+    """Walk through directories to find and handle corrupt Whisper files."""
+    total_files = 0
+    corrupt_files = 0
+    deleted_files = 0
+
+    for dirpath, _, filenames in os.walk(base_dir):
+        logging.info("Scanning %s...", dirpath)
+
+        whisper_files = (os.path.join(dirpath, f) for f in filenames if f.endswith(".wsp"))
+        for f in whisper_files:
+            total_files += 1
+            try:
+                info = whisper.info(f)
+                logging.debug("%s: %d points", f, sum(i["points"] for i in info.get("archives", {})))
+            except whisper.CorruptWhisperFile:
+                corrupt_files += 1
+                if backup_corrupt:
+                    backup_path = f + ".bak"
+                    try:
+                        os.rename(f, backup_path)
+                        logging.warning("Backed up corrupt file: %s -> %s", f, backup_path)
+                    except OSError as e:
+                        logging.error("Failed to back up %s: %s", f, e)
+                        continue
+
+                if delete_corrupt:
+                    try:
+                        os.unlink(f)
+                        deleted_files += 1
+                        logging.warning("Deleted corrupt file: %s", f)
+                    except OSError as e:
+                        logging.error("Failed to delete %s: %s", f, e)
+                else:
+                    logging.error("Corrupt Whisper file: %s", f)
+
+    logging.info("Summary: Scanned %d files, Found %d corrupt, Deleted %d", total_files, corrupt_files, deleted_files)
+    return total_files, corrupt_files, deleted_files
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__.strip())
+    parser.add_argument("--delete-corrupt", action="store_true", help="Delete reported corrupt files")
+    parser.add_argument("--backup-corrupt", action="store_true", help="Back up corrupt files before deletion")
+    parser.add_argument("--verbose", action="store_true", help="Display detailed progress")
+    parser.add_argument("directories", type=str, nargs="+", metavar="WHISPER_DIR", help="Directory containing Whisper files")
+    args = parser.parse_args()
+
+    setup_logging(verbose=args.verbose)
+
+    for d in args.directories:
+        d = os.path.realpath(d)
+        if not os.path.isdir(d):
+            logging.error("%s is not a directory!", d)
+            continue
+
+        walk_dir(d, delete_corrupt=args.delete_corrupt, backup_corrupt=args.backup_corrupt)

generation_whisper (1).py

@@ -0,0 +1,1881 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import copy
+import math
+import warnings
+import zlib
+from typing import Callable, Iterator, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.cache_utils import EncoderDecoderCache
+
+from ...generation import GenerationConfig, GenerationMixin
+from ...generation.logits_process import (
+    LogitsProcessorList,
+    SuppressTokensAtBeginLogitsProcessor,
+    SuppressTokensLogitsProcessor,
+    WhisperNoSpeechDetection,
+    WhisperTimeStampLogitsProcessor,
+)
+from ...generation.stopping_criteria import StoppingCriteriaList
+from ...modeling_outputs import BaseModelOutput
+from ...utils import logging
+from .tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
+
+
+logger = logging.get_logger(__name__)
+
+
+def _median_filter(inputs: torch.Tensor, filter_width: int) -> torch.Tensor:
+    """
+    Applies a median filter of width `filter_width` along the last dimension of the input.
+
+    The `inputs` tensor is assumed to be 3- or 4-dimensional.
+    """
+    if filter_width <= 0 or filter_width % 2 != 1:
+        raise ValueError("`filter_width` should be an odd number")
+
+    pad_width = filter_width // 2
+    if inputs.shape[-1] <= pad_width:
+        return inputs
+
+    # Pad the left and right edges.
+    inputs = nn.functional.pad(inputs, (pad_width, pad_width, 0, 0), mode="reflect")
+
+    # sort() is faster than torch.median (https://github.com/pytorch/pytorch/issues/51450)
+    result = inputs.unfold(-1, filter_width, 1).sort()[0][..., pad_width]
+    return result
+
+
+def _dynamic_time_warping(matrix: np.ndarray):
+    """
+    Measures similarity between two temporal sequences: the input audio and the output tokens. Used to generate
+    token-level timestamps.
+    """
+    output_length, input_length = matrix.shape
+    cost = np.ones((output_length + 1, input_length + 1), dtype=np.float32) * np.inf
+    trace = -np.ones((output_length + 1, input_length + 1), dtype=np.float32)
+
+    cost[0, 0] = 0
+    for j in range(1, input_length + 1):
+        for i in range(1, output_length + 1):
+            c0 = cost[i - 1, j - 1]
+            c1 = cost[i - 1, j]
+            c2 = cost[i, j - 1]
+
+            if c0 < c1 and c0 < c2:
+                c, t = c0, 0
+            elif c1 < c0 and c1 < c2:
+                c, t = c1, 1
+            else:
+                c, t = c2, 2
+
+            cost[i, j] = matrix[i - 1, j - 1] + c
+            trace[i, j] = t
+
+    # backtrace
+    i = trace.shape[0] - 1
+    j = trace.shape[1] - 1
+    trace[0, :] = 2
+    trace[:, 0] = 1
+
+    text_indices = []
+    time_indices = []
+    while i > 0 or j > 0:
+        text_indices.append(i - 1)
+        time_indices.append(j - 1)
+        if trace[i, j] == 0:
+            i -= 1
+            j -= 1
+        elif trace[i, j] == 1:
+            i -= 1
+        elif trace[i, j] == 2:
+            j -= 1
+        else:
+            raise RuntimeError(
+                f"Internal error in dynamic time warping. Unexpected trace[{i}, {j}]. Please file a bug report."
+            )
+
+    text_indices = np.array(text_indices)[::-1]
+    time_indices = np.array(time_indices)[::-1]
+    return text_indices, time_indices
+
+
+def _get_attr_from_logit_processors(logits_processor, logit_processor_class, attribute_name):
+    if logits_processor is not None:
+        logit_processor = next((cls for cls in logits_processor if isinstance(cls, logit_processor_class)), None)
+        if logit_processor:
+            return getattr(logit_processor, attribute_name, None)
+    return None
+
+
+def _pad_to_max_length(
+    current_segments,
+    pad_token_id,
+    device,
+    padding_side="right",
+    padding="longest",
+    bos_token_tensor=None,
+    cut_off_length=None,
+):
+    max_total_length = 0
+    sequences = []
+
+    if padding_side not in ["right", "left"]:
+        raise ValueError(f"`padding_side` must be either 'right' or 'left', not {padding_side}")
+
+    if padding not in ["longest", "max_length"]:
+        raise ValueError(f"`padding` must be either 'longest' or 'max_length', not {padding}")
+    elif padding == "max_length" and cut_off_length is None:
+        raise ValueError("`cut_off_length` must be specified when `padding='max_length'`")
+
+    for current_segment_list in current_segments:
+        if current_segment_list is not None and len([d["tokens"] for d in current_segment_list]) > 0:
+            sequence = torch.cat([d["tokens"] for d in current_segment_list], dim=-1)
+
+            if cut_off_length is not None:
+                sequence = sequence[-cut_off_length:]
+
+            if bos_token_tensor is not None:
+                sequence = torch.cat([bos_token_tensor, sequence])
+
+            sequences.append(sequence)
+            max_total_length = max(max_total_length, len(sequences[-1]))
+        elif bos_token_tensor is not None:
+            sequences.append(bos_token_tensor)
+        else:
+            sequences.append(torch.tensor([], device=device))
+
+    max_total_length = cut_off_length + 1 if padding == "max_length" else max_total_length
+    for i in range(len(current_segments)):
+        pad_length = max_total_length - len(sequences[i])
+        pad = (0, pad_length) if padding_side == "right" else (pad_length, 0)
+        sequences[i] = F.pad(sequences[i], pad=pad, value=pad_token_id)
+
+    sequences = torch.stack(sequences, dim=0)
+    return sequences
+
+
+class WhisperGenerationMixin(GenerationMixin):
+    def _extract_token_timestamps(
+        self, generate_outputs, alignment_heads, time_precision=0.02, num_frames=None, num_input_ids=None
+    ):
+        """
+        Calculates token-level timestamps using the encoder-decoder cross-attentions and dynamic time-warping (DTW) to
+        map each output token to a position in the input audio. If `num_frames` is specified, the encoder-decoder
+        cross-attentions will be cropped before applying DTW.
+
+        Returns:
+            tensor containing the timestamps in seconds for each predicted token
+        """
+        # Create a list with `decoder_layers` elements, each a tensor of shape
+        # (batch size, attention_heads, output length, input length).
+        cross_attentions = []
+        for i in range(self.config.decoder_layers):
+            cross_attentions.append(torch.cat([x[i] for x in generate_outputs.cross_attentions], dim=2))
+
+        # Select specific cross-attention layers and heads. This is a tensor
+        # of shape (batch size, num selected, output length, input length).
+        weights = torch.stack([cross_attentions[l][:, h] for l, h in alignment_heads])
+        weights = weights.permute([1, 0, 2, 3])
+
+        weight_length = None
+
+        if "beam_indices" in generate_outputs:
+            # If beam search has been used, the output sequences may have been generated for more timesteps than their sequence_lengths
+            # since the beam search strategy chooses the most probable sequences at the end of the search.
+            # In that case, the cross_attentions weights are too long and we have to make sure that they have the right output_length
+            weight_length = (generate_outputs.beam_indices != -1).sum(-1).max()
+            weight_length = weight_length if num_input_ids is None else weight_length + num_input_ids
+
+            # beam search takes `decoder_input_ids` into account in the `beam_indices` length
+            # but forgot to shift the beam_indices by the number of `decoder_input_ids`
+            beam_indices = torch.zeros_like(generate_outputs.beam_indices[:, :weight_length])
+            # we actually shif the beam indices here
+            beam_indices[:, num_input_ids:] = generate_outputs.beam_indices[:, : weight_length - num_input_ids]
+
+            weights = weights[:, :, :weight_length]
+
+            # If beam index is still -1, it means that the associated token id is EOS
+            # We need to replace the index with 0 since index_select gives an error if any of the indexes is -1.
+            beam_indices = beam_indices.masked_fill(beam_indices == -1, 0)
+
+            # Select the cross attention from the right beam for each output sequences
+            weights = torch.stack(
+                [
+                    torch.index_select(weights[:, :, i, :], dim=0, index=beam_indices[:, i])
+                    for i in range(beam_indices.shape[1])
+                ],
+                dim=2,
+            )
+
+        # make sure timestamps are as long as weights
+        input_length = weight_length or cross_attentions[0].shape[2]
+        batch_size = generate_outputs.sequences.shape[0]
+        timestamps = torch.zeros(
+            (batch_size, input_length + 1), dtype=torch.float32, device=generate_outputs.sequences.device
+        )
+
+        if num_frames is not None:
+            # two cases:
+            # 1. num_frames is the same for each sample -> compute the DTW matrix for each sample in parallel
+            # 2. num_frames is different, compute the DTW matrix for each sample sequentially
+
+            # we're using np.unique because num_frames can be int/list/tuple
+            if isinstance(num_frames, int):
+                weights = weights[..., : num_frames // 2]
+
+            elif isinstance(num_frames, (list, tuple, np.ndarray)) and len(np.unique(num_frames)) == 1:
+                weights = weights[..., : num_frames[0] // 2]
+
+            elif isinstance(num_frames, (torch.Tensor)) and len(torch.unique(num_frames)) == 1:
+                weights = weights[..., : num_frames[0] // 2]
+
+            else:
+                # num_frames is of shape (batch_size,) whereas batch_size is truely batch_size*num_return_sequences
+                repeat_time = batch_size if isinstance(num_frames, int) else batch_size // len(num_frames)
+                num_frames = num_frames.cpu() if isinstance(num_frames, (torch.Tensor)) else num_frames
+                num_frames = np.repeat(num_frames, repeat_time)
+
+        if num_frames is None or isinstance(num_frames, int):
+            # Normalize and smoothen the weights.
+            std = torch.std(weights, dim=-2, keepdim=True, unbiased=False)
+            mean = torch.mean(weights, dim=-2, keepdim=True)
+            weights = (weights - mean) / std
+            weights = _median_filter(weights, self.config.median_filter_width)
+
+            # Average the different cross-attention heads.
+            weights = weights.mean(dim=1)
+
+        # Perform dynamic time warping on each element of the batch.
+        for batch_idx in range(batch_size):
+            if num_frames is not None and isinstance(num_frames, (tuple, list, np.ndarray, torch.Tensor)):
+                matrix = weights[batch_idx, ..., : num_frames[batch_idx] // 2]
+
+                # Normalize and smoothen the weights.
+                std = torch.std(matrix, dim=-2, keepdim=True, unbiased=False)
+                mean = torch.mean(matrix, dim=-2, keepdim=True)
+                matrix = (matrix - mean) / std
+                matrix = _median_filter(matrix, self.config.median_filter_width)
+
+                # Average the different cross-attention heads.
+                matrix = matrix.mean(dim=0)
+            else:
+                matrix = weights[batch_idx]
+
+            text_indices, time_indices = _dynamic_time_warping(-matrix.cpu().double().numpy())
+            jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
+            jump_times = time_indices[jumps] * time_precision
+            timestamps[batch_idx, 1:] = torch.tensor(jump_times)
+
+        return timestamps
+
+    def generate(
+        self,
+        input_features: Optional[torch.Tensor] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        logits_processor: Optional[LogitsProcessorList] = None,
+        stopping_criteria: Optional[StoppingCriteriaList] = None,
+        prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+        synced_gpus: bool = False,
+        return_timestamps: Optional[bool] = None,
+        task: Optional[str] = None,
+        language: Optional[Union[str, List[str]]] = None,
+        is_multilingual: Optional[bool] = None,
+        prompt_ids: Optional[torch.Tensor] = None,
+        prompt_condition_type: Optional[str] = None,  # first-segment, all-segments
+        condition_on_prev_tokens: Optional[bool] = None,
+        temperature: Optional[Union[float, Tuple[float, ...]]] = None,
+        compression_ratio_threshold: Optional[float] = None,
+        logprob_threshold: Optional[float] = None,
+        no_speech_threshold: Optional[float] = None,
+        num_segment_frames: Optional[int] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        time_precision: float = 0.02,
+        time_precision_features: float = 0.01,
+        return_token_timestamps: Optional[bool] = None,
+        return_segments: bool = False,
+        return_dict_in_generate: Optional[bool] = None,
+        **kwargs,
+    ):
+        """
+        Transcribes or translates log-mel input features to a sequence of auto-regressively generated token ids.
+
+        <Tip warning={true}>
+
+        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
+        model's default generation configuration. You can override any `generation_config` by passing the corresponding
+        parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
+
+        For an overview of generation strategies and code examples, check out the [following
+        guide](./generation_strategies).
+
+        </Tip>
+
+        Parameters:
+            input_features (`torch.Tensor` of shape `(batch_size, feature_size, sequence_length)`, *optional*):
+                Float values of log-mel features extracted from the raw speech waveform. The raw speech waveform can be obtained by
+                loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+                the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+                [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+                tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] for details.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            logits_processor (`LogitsProcessorList`, *optional*):
+                Custom logits processors that complement the default logits processors built from arguments and
+                generation config. If a logit processor is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            stopping_criteria (`StoppingCriteriaList`, *optional*):
+                Custom stopping criteria that complement the default stopping criteria built from arguments and a
+                generation config. If a stopping criteria is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
+                If provided, this function constraints the beam search to allowed tokens only at each step. If not
+                provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
+                `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
+                on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
+                for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
+                Retrieval](https://arxiv.org/abs/2010.00904).
+            synced_gpus (`bool`, *optional*, defaults to `False`):
+                Whether to continue running the while loop until max_length (needed to avoid deadlocking with
+                `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3).
+            return_timestamps (`bool`, *optional*):
+                Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
+            task (`str`, *optional*):
+                Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
+                will be updated accordingly.
+            language (`str` or list of `str`, *optional*):
+                Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. For
+                batched generation, a list of language tokens can be passed. You can find all the possible language
+                tokens in the `model.generation_config.lang_to_id` dictionary.
+            is_multilingual (`bool`, *optional*):
+                Whether or not the model is multilingual.
+            prompt_ids (`torch.Tensor`, *optional*):
+                Rank-1 tensor of token IDs created by passing text to [`~WhisperProcessor.get_prompt_ids`] that is
+                provided as a prompt to each chunk. This can be used to provide or "prompt-engineer" a context for
+                transcription, e.g. custom vocabularies or proper nouns to make it more likely to predict those words
+                correctly. It cannot be used in conjunction with `decoder_start_token_id` as it overwrites this value.
+            prompt_condition_type (`str`, *optional*):
+                Only relevant for long-form transcription. Condition type of `prompt_ids`. 'first-segment' means only the first segment is conditioned on `prompt_ids`. 'all-segments' means each segment is conditioned on `prompt_ids`. Make sure to enable `condition_on_prev_tokens` for 'all-segments'.
+                Defaults to 'first-segment'. For short-term transcription only 'first-segment' is possible.
+            condition_on_prev_tokens (`bool`, *optional*):
+                Only relevant for long-form transcription. Whether to condition each segment on the previous segment.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            temperature (`float` or list of `float`, *optional*):
+                The temperature to be used for generation. Passing a single `float` value and `do_sample=True` activates
+                generation using sampling. For long-form transcription, temperature fallback can be activated by passing
+                a list of float values such as (0.0, 0.2, 0.4, 0.6, 0.8, 1.0). As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            compression_ratio_threshold (`float`, *optional*):
+                Only relevant for long-form transcription. If defined, the zlib compression rate of each segment will be computed. If the compression rate of
+                a segment is higher than `compression_ratio_threshold`, temperature fallback is activated: the generated segment is discarded and the generation is
+                repeated using a higher temperature. The intuition behind this feature is that segments with very high compression rates
+                suffer from a lot of repetition. The unwanted repetition can be reduced by injecting more randomness by increasing the temperature. If `compression_ratio_threshold` is defined
+                make sure that `temperature` is a list of values. A common value for `compression_ratio_threshold` is 1.35.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            logprob_threshold (`float`, *optional*):
+                Only relevant for long-form transcription. If defined, the average log-probability of each segment will be computed. If the log-probability of
+                a given segment is lower than `logprob_threshold`, temperature fallback is activated: the generated segment is discarded and the generation is
+                repeated using a higher temperature. The intuition behind this feature is that segments of low log-probability
+                can be improved by injecting more randomness by increasing the temperature. If `logprob_threshold` is defined
+                make sure that `temperature` is a list of values. A common value for `logprob_threshold` is -1.0.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            no_speech_threshold (`float`, *optional*):
+                Only relevant for long-form transcription. If defined, the "no-speech" token combined with the `logprob_threshold`
+                is used to determine whether a segment contains only silence. In this case, the transcription for this segment
+                is skipped.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            num_segment_frames (`int`, *optional*):
+                The number of frames a single segment is made of. If not defined, `num_segment_frames` defaults to the model's stride
+                times the maximum input length.
+            attention_mask (`torch.Tensor`, *optional*):
+                `attention_mask` needs to be passed when doing long-form transcription using a batch size > 1.
+            time_precision (`int`, *optional*, defaults to 0.02):
+                The duration of output token in seconds. *E.g.* 0.02 means that a generated token on average accounts
+                for 20 ms.
+            time_precision_features (`int`, *optional*, defaults to 0.01):
+                The duration represented by a feature frame in seconds.
+            return_token_timestamps (`bool`, *optional*):
+                Whether to return token-level timestamps with the text. This can be used with or without the
+                `return_timestamps` option. To get word-level timestamps, use the tokenizer to group the tokens into
+                words.
+            return_segments (`bool`, *optional*, defaults to `False`):
+                Whether to additionally return a list of all segments. Note that this option can only be enabled
+                when doing long-form transcription.
+            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
+                Whether or not to return a [`~utils.ModelOutput`] instead of just returning the generated tokens.
+                Note that when doing long-form transcription, `return_dict_in_generate` can only be enabled when
+                `return_segments` is set True. In this case the generation outputs of each segment is added to each
+                segment.
+            kwargs (`Dict[str, Any]`, *optional*):
+                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
+                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
+                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
+
+        Return:
+            [`~utils.ModelOutput`] or `torch.LongTensor` or `Dict[str, Any]`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
+            or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor` or a dict of segments when `return_segments=True`.
+
+                If the passed input is > 30 seconds / > 3000 mel input features and `return_segments=True` then a dictionary of generated sequence ids, called `sequences` and a list of each generated segment is returned.
+
+                else if the passed input is <= 30 seconds / >= 3000 mel input features, the possible [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.GenerateEncoderDecoderOutput`],
+                    - [`~generation.GenerateBeamEncoderDecoderOutput`]
+
+                else only the generated output sequence ids are returned.
+
+        Example:
+
+        - *Longform transcription*: To transcribe or translate audios longer than 30 seconds, process the audio files without truncation and pass all mel features at once to generate.
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset, Audio
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+        >>> model.cuda()  # doctest: +IGNORE_RESULT
+
+        >>> # load audios > 30 seconds
+        >>> ds = load_dataset("distil-whisper/meanwhile", "default")["test"]
+        >>> # resample to 16kHz
+        >>> ds = ds.cast_column("audio", Audio(sampling_rate=16000))
+        >>> # take first 8 audios and retrieve array
+        >>> audio = ds[:8]["audio"]
+        >>> audio = [x["array"] for x in audio]
+
+        >>> # make sure to NOT truncate the input audio, to return the `attention_mask` and to pad to the longest audio
+        >>> inputs = processor(audio, return_tensors="pt", truncation=False, padding="longest", return_attention_mask=True, sampling_rate=16_000)
+        >>> inputs = inputs.to("cuda", torch.float32)
+
+        >>> # transcribe audio to ids
+        >>> generated_ids = model.generate(**inputs)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        >>> transcription[0]
+        " Folks, if you watch the show, you know, I spent a lot of time right over there. Patiently and astutely scrutinizing the boxwood and mahogany chest set of the day's biggest stories developing the central headline pawns, definitely maneuvering an oso topical night to F6, fainting a classic Sicilian, nade door variation on the news, all the while seeing eight moves deep and patiently marshalling the latest press releases into a fisher's shows in Lip Nitsky attack that culminates in the elegant lethal slow-played, all-passant checkmate that is my nightly monologue. But sometimes, sometimes, folks, I. CHEERING AND APPLAUSE Sometimes I startle away, cubside down in the monkey bars of a condemned playground on a super fun site. Get all hept up on goofballs. Rummage that were discarded tag bag of defective toys. Yank out a fist bowl of disembodied doll limbs, toss them on a stained kid's place mat from a defunct dennies. set up a table inside a rusty cargo container down by the Wharf and challenged toothless drifters to the godless bughouse blitz of tournament that is my segment. Meanwhile."
+        ```
+
+        - *Shortform transcription*: If passed mel input features are < 30 seconds, the whole audio will be transcribed with a single call to generate.
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(inputs=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```
+
+        """
+        # 0. deprecate old inputs
+        if "inputs" in kwargs:
+            input_features = kwargs.pop("inputs")
+            warnings.warn(
+                "The input name `inputs` is deprecated. Please make sure to use `input_features` instead.",
+                FutureWarning,
+            )
+
+        # 1. prepare generation config
+        generation_config, kwargs = self._prepare_generation_config(generation_config, **kwargs)
+
+        # 2. set global generate variables
+        input_stride = self.model.encoder.conv1.stride[0] * self.model.encoder.conv2.stride[0]
+        num_segment_frames = input_stride * self.config.max_source_positions
+        batch_size, total_input_frames = self._retrieve_total_input_frames(
+            input_features=input_features, input_stride=input_stride, kwargs=kwargs
+        )
+        is_shortform = total_input_frames <= num_segment_frames
+
+        # 3. Make sure generation config is correctly set
+        # Make sure the generation config is correctly set depending on whether timestamps are to be returned or not
+        return_dict_in_generate = self._set_return_outputs(
+            return_dict_in_generate=return_dict_in_generate,
+            return_token_timestamps=return_token_timestamps,
+            logprob_threshold=logprob_threshold,
+            generation_config=generation_config,
+        )
+        timestamp_begin = self._set_return_timestamps(
+            return_timestamps=return_timestamps, is_shortform=is_shortform, generation_config=generation_config
+        )
+        self._set_language_and_task(
+            language=language, task=task, is_multilingual=is_multilingual, generation_config=generation_config
+        )
+        self._set_num_frames(
+            return_token_timestamps=return_token_timestamps, generation_config=generation_config, kwargs=kwargs
+        )
+        self._set_thresholds_and_condition(
+            generation_config=generation_config,
+            logprob_threshold=logprob_threshold,
+            compression_ratio_threshold=compression_ratio_threshold,
+            no_speech_threshold=no_speech_threshold,
+            condition_on_prev_tokens=condition_on_prev_tokens,
+        )
+        self._set_prompt_condition_type(
+            generation_config=generation_config,
+            prompt_condition_type=prompt_condition_type,
+        )
+
+        # pass self.config for backward compatibility
+        init_tokens = self._retrieve_init_tokens(
+            input_features,
+            batch_size=batch_size,
+            generation_config=generation_config,
+            config=self.config,
+            num_segment_frames=num_segment_frames,
+            kwargs=kwargs,
+        )
+        # passing `decoder_input_ids` is deprecated - the only exception is for assisted generation
+        # where the input ids are handled explicitly by the generate method
+        self._check_decoder_input_ids(kwargs=kwargs)
+
+        # 3. Retrieve logits processors
+        device = kwargs["encoder_outputs"][0].device if "encoder_outputs" in kwargs else input_features.device
+        begin_index = init_tokens.shape[1]
+        logits_processor = self._retrieve_logit_processors(
+            generation_config=generation_config,
+            logits_processor=logits_processor,
+            begin_index=begin_index,  # begin index is index of first generated decoder token
+            num_beams=kwargs.get("num_beams", 1),
+            device=device,
+        )
+
+        # 4 Set and retrieve global generation variables
+        self._set_condition_on_prev_tokens(
+            condition_on_prev_tokens=condition_on_prev_tokens, generation_config=generation_config
+        )
+
+        temperatures = [temperature] if not isinstance(temperature, (list, tuple)) else temperature
+        temperature = temperatures[0]
+
+        max_frames, seek = self._retrieve_max_frames_and_seek(
+            batch_size=batch_size,
+            attention_mask=attention_mask,
+            total_input_frames=total_input_frames,
+            is_shortform=is_shortform,
+        )
+
+        # 5 Prepare running variables, list for generation
+        num_return_sequences = generation_config.num_return_sequences
+        (
+            batch_idx_map,
+            cur_bsz,
+            input_features,
+            seek,
+            max_frames,
+            init_tokens,
+            do_condition_on_prev_tokens,
+        ) = self._expand_variables_for_generation(
+            input_features=input_features,
+            seek=seek,
+            max_frames=max_frames,
+            init_tokens=init_tokens,
+            batch_size=batch_size,
+            condition_on_prev_tokens=condition_on_prev_tokens,
+            generation_config=generation_config,
+        )
+
+        current_segments = self._prepare_segments(
+            prompt_ids=prompt_ids,
+            batch_size=cur_bsz,
+            generation_config=generation_config,
+        )
+
+        # 6 Transcribe audio until we reach the end of all input audios
+        while (seek < max_frames).any():
+            # 6.1 NOTE: When in longform transcription mode and batch size > 1 we need to dynamically reduce the batch size during the loop
+            # in case one audio finished earlier than another one. Thus, we need to keep a table of "previous-index-2-current-index" in order
+            # to know which original audio is being decoded
+            # Set updated index map, duration of previously decoded chunks and number of max frames of current decoding chunk
+            input_features, cur_bsz, batch_idx_map = self._maybe_reduce_batch(
+                input_features=input_features,
+                seek=seek,
+                max_frames=max_frames,
+                cur_bsz=cur_bsz,
+                batch_idx_map=batch_idx_map,
+            )
+            time_offset = (
+                seek.to(torch.float32 if device.type == "mps" else torch.float64) * time_precision / input_stride
+            )
+            seek_num_frames = (max_frames - seek).clamp(max=num_segment_frames)
+
+            # 6.2 cut out next 30s segment from input features
+            segment_input = self._get_input_segment(
+                input_features=input_features,
+                seek=seek,
+                seek_num_frames=seek_num_frames,
+                num_segment_frames=num_segment_frames,
+                cur_bsz=cur_bsz,
+                batch_idx_map=batch_idx_map,
+            )
+
+            # 6.3 prepare decoder input ids
+            suppress_tokens = _get_attr_from_logit_processors(
+                logits_processor, SuppressTokensLogitsProcessor, "suppress_tokens"
+            )
+
+            decoder_input_ids, kwargs = self._prepare_decoder_input_ids(
+                cur_bsz=cur_bsz,
+                init_tokens=init_tokens,
+                current_segments=current_segments,
+                batch_idx_map=batch_idx_map,
+                do_condition_on_prev_tokens=do_condition_on_prev_tokens,
+                prompt_ids=prompt_ids,
+                generation_config=generation_config,
+                config=self.config,
+                device=init_tokens.device,
+                suppress_tokens=suppress_tokens,
+                timestamp_begin=timestamp_begin,
+                kwargs=kwargs,
+            )
+
+            # 6.4 set max new tokens or max length
+            self._set_max_new_tokens_and_length(
+                config=self.config,
+                decoder_input_ids=decoder_input_ids,
+                generation_config=generation_config,
+            )
+
+            # 6.5 Set current `begin_index` for all logit processors
+            if logits_processor is not None:
+                for proc in logits_processor:
+                    if hasattr(proc, "set_begin_index"):
+                        proc.set_begin_index(decoder_input_ids.shape[-1])
+
+            # 6.6 Run generate with fallback
+            (
+                seek_sequences,
+                seek_outputs,
+                should_skip,
+                do_condition_on_prev_tokens,
+                model_output_type,
+            ) = self.generate_with_fallback(
+                segment_input=segment_input,
+                decoder_input_ids=decoder_input_ids,
+                cur_bsz=cur_bsz,
+                batch_idx_map=batch_idx_map,
+                seek=seek,
+                num_segment_frames=num_segment_frames,
+                max_frames=max_frames,
+                temperatures=temperatures,
+                generation_config=generation_config,
+                logits_processor=logits_processor,
+                stopping_criteria=stopping_criteria,
+                prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+                synced_gpus=synced_gpus,
+                return_token_timestamps=return_token_timestamps,
+                do_condition_on_prev_tokens=do_condition_on_prev_tokens,
+                is_shortform=is_shortform,
+                batch_size=batch_size,
+                attention_mask=attention_mask,
+                kwargs=kwargs,
+            )
+
+            # 6.7 In every generated sequence, split by timestamp tokens and extract segments
+            for i, seek_sequence in enumerate(seek_sequences):
+                prev_i = batch_idx_map[i]
+
+                if should_skip[i]:
+                    seek[prev_i] += seek_num_frames[prev_i]
+                    continue
+
+                segments, segment_offset = self._retrieve_segment(
+                    seek_sequence=seek_sequence,
+                    seek_outputs=seek_outputs,
+                    time_offset=time_offset,
+                    timestamp_begin=timestamp_begin,
+                    seek_num_frames=seek_num_frames,
+                    time_precision=time_precision,
+                    time_precision_features=time_precision_features,
+                    input_stride=input_stride,
+                    prev_idx=prev_i,
+                    idx=i,
+                    return_token_timestamps=return_token_timestamps,
+                )
+
+                current_segments[prev_i] += segments
+
+                if is_shortform:
+                    seek[prev_i] += max_frames[i]
+                else:
+                    seek[prev_i] += segment_offset
+
+        # 7. Once all segments are added to the list of all segments, called `current_segments`, we extract the predicted
+        # output tokens from the list of dicts. If we use batch size > 1, we make sure to pad the output
+        final_segments = (
+            [x[1:] for x in current_segments]
+            if (prompt_ids is not None and generation_config.prompt_condition_type == "first-segment")
+            else current_segments
+        )
+
+        sequences = _pad_to_max_length(
+            final_segments, generation_config.pad_token_id, device=self.device, padding_side="right"
+        )
+
+        # 8. If we return all segments, the predicted output sequences are put under `"sequences"`.
+        if return_segments:
+            return {"sequences": sequences, "segments": final_segments}
+
+        if is_shortform:
+            # add eos token:
+            if generation_config.max_new_tokens is None and generation_config.max_length is None:
+                eos_tokens = torch.full((sequences.shape[0], 1), generation_config.eos_token_id)
+                sequences = torch.cat([sequences, eos_tokens], dim=-1)
+
+            if return_token_timestamps:
+                outputs = {}
+                outputs["sequences"] = sequences
+                outputs["token_timestamps"] = torch.stack([d["token_timestamps"] for d in seek_outputs], dim=0)
+            else:
+                outputs = sequences
+
+            if return_dict_in_generate and generation_config.return_dict_in_generate:
+                dict_outputs = self._stack_split_outputs(seek_outputs, model_output_type, sequences.device, kwargs)
+
+                if num_return_sequences > 1:
+                    if hasattr(dict_outputs, "encoder_attentions") and dict_outputs.encoder_attentions is not None:
+                        dict_outputs.encoder_attentions = tuple(
+                            dict_outputs.encoder_attentions[i][::num_return_sequences]
+                            for i in range(len(dict_outputs.encoder_attentions))
+                        )
+                    if (
+                        hasattr(dict_outputs, "encoder_hidden_states")
+                        and dict_outputs.encoder_hidden_states is not None
+                    ):
+                        dict_outputs.encoder_hidden_states = tuple(
+                            dict_outputs.encoder_hidden_states[i][::num_return_sequences]
+                            for i in range(len(dict_outputs.encoder_hidden_states))
+                        )
+                if return_token_timestamps:
+                    dict_outputs["token_timestamps"] = outputs["token_timestamps"]
+                return dict_outputs
+
+            return outputs
+
+        return sequences
+
+    def generate_with_fallback(
+        self,
+        segment_input,
+        decoder_input_ids,
+        cur_bsz,
+        batch_idx_map,
+        seek,
+        num_segment_frames,
+        max_frames,
+        temperatures,
+        generation_config,
+        logits_processor,
+        stopping_criteria,
+        prefix_allowed_tokens_fn,
+        synced_gpus,
+        return_token_timestamps,
+        do_condition_on_prev_tokens,
+        is_shortform,
+        batch_size,
+        attention_mask,
+        kwargs,
+    ):
+        kwargs = copy.copy(kwargs)
+
+        # 6.6 Batch generate current chunk
+        seek_sequence_list = [None for _ in range(cur_bsz)]
+        seek_outputs_list = [None for _ in range(cur_bsz)]
+        needs_fallback = [False for _ in range(cur_bsz)]
+        should_skip = [False for _ in range(cur_bsz)]
+        fallback_index_map = list(range(cur_bsz))
+        if generation_config.no_speech_threshold is not None:
+            self._setup_no_speech_detection(logits_processor, segment_input, decoder_input_ids, kwargs)
+
+        for fallback_idx, temperature in enumerate(temperatures):
+            generation_config.do_sample = temperature is not None and temperature > 0.0
+            generation_config.temperature = temperature if generation_config.do_sample else 1.0
+            if generation_config.do_sample:
+                generation_config.num_beams = 1
+
+            generate_kwargs = copy.copy(kwargs)
+            for key in ["do_sample", "temperature", "num_beams"]:
+                if key in generate_kwargs:
+                    del generate_kwargs[key]
+
+            cur_bsz = decoder_input_ids.shape[0]
+            if generation_config.cache_implementation == "static" and cur_bsz < batch_size:
+                segment_input = F.pad(segment_input, (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0)
+                decoder_input_ids = F.pad(
+                    decoder_input_ids, (0, 0, 0, batch_size - cur_bsz), value=generation_config.pad_token_id
+                )
+                if generate_kwargs.get("decoder_attention_mask") is not None:
+                    generate_kwargs["decoder_attention_mask"] = F.pad(
+                        generate_kwargs["decoder_attention_mask"], (0, 0, 0, batch_size - cur_bsz), value=True
+                    )
+                if generate_kwargs.get("encoder_outputs") is not None:
+                    generate_kwargs["encoder_outputs"] = F.pad(
+                        generate_kwargs["encoder_outputs"], (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0
+                    )
+
+            seek_outputs = super().generate(
+                segment_input,
+                generation_config=generation_config,
+                logits_processor=logits_processor,
+                stopping_criteria=stopping_criteria,
+                prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+                synced_gpus=synced_gpus,
+                decoder_input_ids=decoder_input_ids,
+                attention_mask=attention_mask,
+                **generate_kwargs,
+            )
+
+            model_output_type = type(seek_outputs)
+
+            # post-process sequence tokens and outputs to be in list form
+            seek_sequences, seek_outputs = self._postprocess_outputs(
+                seek_outputs=seek_outputs,
+                decoder_input_ids=decoder_input_ids,
+                return_token_timestamps=return_token_timestamps,
+                generation_config=generation_config,
+                is_shortform=is_shortform,
+            )
+
+            if cur_bsz < batch_size:
+                seek_sequences = seek_sequences[:cur_bsz]
+                seek_outputs = seek_outputs[:cur_bsz]
+
+            # 6.7 Extract cut sequences from every sequence and check if fallback should be applied
+            # Loop over each decoded audio individually as each decoding can be of a different length
+            new_fallback_index_map = []
+            new_segment_input = []
+            new_decoder_input_ids = []
+            new_decoder_attention_mask = []
+
+            for i, seek_sequence in enumerate(seek_sequences):
+                # make sure we cut a predicted EOS token if we are not finished with the generation yet
+                prev_i = batch_idx_map[fallback_index_map[i]]
+                is_not_final = (seek[prev_i] + num_segment_frames) < max_frames[prev_i]
+
+                # remove eos token id
+                if is_not_final and seek_sequence[-1] == generation_config.eos_token_id:
+                    seek_sequence = seek_sequence[:-1]
+                    if return_token_timestamps and not is_shortform:
+                        seek_outputs[i]["token_timestamps"] = seek_outputs[i]["token_timestamps"][:-1]
+
+                # remove all padding tokens
+                if seek_sequence[-1] == generation_config.pad_token_id:
+                    num_paddings = (seek_sequence == generation_config.pad_token_id).sum()
+                    seek_sequence = seek_sequence[:-num_paddings]
+                    if return_token_timestamps and not is_shortform:
+                        seek_outputs[i]["token_timestamps"] = seek_outputs[i]["token_timestamps"][:-num_paddings]
+
+                # check which sequences in batch need fallback & which should be skipped
+                needs_fallback[i], should_skip[i] = self._need_fallback(
+                    seek_sequence,
+                    seek_outputs,
+                    i,
+                    logits_processor,
+                    generation_config,
+                    self.config.vocab_size,
+                    temperature,
+                )
+
+                seek_sequence_list[fallback_index_map[i]] = seek_sequence
+                seek_outputs_list[fallback_index_map[i]] = seek_outputs[i]
+                is_low_temperature = temperature is None or temperature < 0.5
+                do_condition_on_prev_tokens[fallback_index_map[i]] = (
+                    generation_config.condition_on_prev_tokens and is_low_temperature
+                )
+
+                if needs_fallback[i]:
+                    new_fallback_index_map.append(fallback_index_map[i])
+                    new_segment_input.append(segment_input[i])
+                    new_decoder_input_ids.append(decoder_input_ids[i])
+                    if "decoder_attention_mask" in kwargs:
+                        new_decoder_attention_mask.append(kwargs["decoder_attention_mask"][i])
+
+            fallback_index_map = new_fallback_index_map
+
+            # if no sequence needs to be run with temperature fallback, we're finished
+            if len(fallback_index_map) == 0 or fallback_idx == len(temperatures) - 1:
+                seek_sequences = seek_sequence_list
+                seek_outputs = seek_outputs_list
+                break
+
+            # if we're still in the loop, make sure that decoder_input_ids and segment inputs are tensors
+            decoder_input_ids = torch.stack(new_decoder_input_ids)
+            segment_input = torch.stack(new_segment_input)
+            if "decoder_attention_mask" in kwargs:
+                kwargs["decoder_attention_mask"] = torch.stack(new_decoder_attention_mask)
+
+        return seek_sequences, seek_outputs, should_skip, do_condition_on_prev_tokens, model_output_type
+
+    @staticmethod
+    def _prepare_segments(prompt_ids, batch_size, generation_config):
+        if prompt_ids is not None and generation_config.prompt_condition_type == "first-segment":
+            prev_sot_token_id = getattr(generation_config, "prev_sot_token_id", None)
+            prompt_ids = prompt_ids[1:] if prompt_ids[0] == prev_sot_token_id else prompt_ids
+            current_segments = [[{"tokens": prompt_ids}] for _ in range(batch_size)]
+        else:
+            current_segments = [[] for _ in range(batch_size)]
+
+        return current_segments
+
+    def _postprocess_outputs(
+        self, seek_outputs, decoder_input_ids, return_token_timestamps, generation_config, is_shortform
+    ):
+        # remove all previously passed decoder input ids
+        start_idx = decoder_input_ids.shape[-1] if not is_shortform else torch.tensor(0)
+
+        if isinstance(seek_outputs, torch.Tensor):
+            seek_outputs = seek_outputs[:, start_idx:]
+            return seek_outputs, seek_outputs
+
+        if return_token_timestamps and hasattr(generation_config, "alignment_heads"):
+            num_frames = getattr(generation_config, "num_frames", None)
+            seek_outputs["token_timestamps"] = self._extract_token_timestamps(
+                seek_outputs,
+                generation_config.alignment_heads,
+                num_frames=num_frames,
+                num_input_ids=decoder_input_ids.shape[-1],
+            )
+            seek_outputs["token_timestamps"] = seek_outputs["token_timestamps"][:, start_idx:]
+
+        seek_outputs["sequences"] = seek_outputs["sequences"][:, start_idx:]
+
+        def split_by_batch_index(values, key, batch_idx, is_shortform, beam_indices=None):
+            if beam_indices is not None and key == "scores":
+                return [v[beam_idx].cpu() for (v, beam_idx) in zip(values, beam_indices[batch_idx][: len(values)])]
+            if key in ["scores", "encoder_attentions", "encoder_hidden_states", "logits"]:
+                return [v[batch_idx].cpu() for v in values]
+            if key in ["decoder_attentions", "decoder_hidden_states", "cross_attentions"]:
+                return tuple(tuple(w[batch_idx][None].cpu() for w in v) for v in values)
+            elif key == "past_key_values":
+                if not is_shortform:
+                    # we don't save `past_key_values` as this is too costly for longform
+                    return None
+                elif isinstance(values, EncoderDecoderCache):
+                    all_past_key_values = []
+                    for layer_idx in range(self.config.decoder_layers):
+                        layer_past_key_values = []
+                        for cache_cls in [values.self_attention_cache, values.cross_attention_cache]:
+                            for v in [cache_cls.key_cache, cache_cls.value_cache]:
+                                layer_past_key_values.append(v[layer_idx][batch_idx][None].cpu())
+                        all_past_key_values.append(tuple(layer_past_key_values))
+                    return tuple(all_past_key_values)
+                else:
+                    all_past_key_values = []
+                    for v in range(len(values)):
+                        layer_past_key_values = []
+                        for w in values[v]:
+                            if len(w) != 0:
+                                layer_past_key_values.append(w[batch_idx][None].cpu())
+                            else:
+                                layer_past_key_values.append(w)
+                        all_past_key_values.append(tuple(layer_past_key_values))
+                    return tuple(all_past_key_values)
+
+            return values[batch_idx].cpu()
+
+        sequence_tokens = seek_outputs["sequences"]
+        seek_outputs = [
+            {
+                k: split_by_batch_index(v, k, i, is_shortform, beam_indices=seek_outputs.get("beam_indices"))
+                for k, v in seek_outputs.items()
+            }
+            for i in range(sequence_tokens.shape[0])
+        ]
+
+        return sequence_tokens, seek_outputs
+
+    def _stack_split_outputs(self, seek_outputs, model_output_type, device, kwargs):
+        # Stack back seek_outputs tensors after splitting them with the split_by_batch_index method
+        outputs = {}
+        for key in seek_outputs[0].keys():
+            if key in ["sequences", "beam_indices"]:
+                outputs[key] = torch.stack([v[key] for v in seek_outputs], dim=0).to(device)
+            elif key in ["scores", "encoder_attentions", "encoder_hidden_states", "logits"]:
+                outputs[key] = tuple(
+                    torch.stack([v[key][i] for v in seek_outputs]).to(device) for i in range(len(seek_outputs[0][key]))
+                )
+            elif key == "sequences_scores":
+                outputs[key] = torch.stack([v[key] for v in seek_outputs], dim=0).to(device)
+            elif key in ["decoder_attentions", "decoder_hidden_states", "cross_attentions"]:
+                outputs[key] = tuple(
+                    tuple(
+                        torch.stack([v[key][i][j] for v in seek_outputs]).squeeze(1).to(device)
+                        for j in range(len(seek_outputs[0][key][0]))
+                    )
+                    for i in range(len(seek_outputs[0][key]))
+                )
+            elif key == "past_key_values":
+                past_key_value_type = kwargs.get("past_key_values")
+                if seek_outputs[0][key] is not None:
+                    outputs[key] = tuple(
+                        tuple(
+                            torch.stack([v[key][i][j] for v in seek_outputs]).squeeze(1).to(device)
+                            for j in range(len(seek_outputs[0][key][0]))
+                        )
+                        for i in range(len(seek_outputs[0][key]))
+                    )
+                    if past_key_value_type is not None and isinstance(past_key_value_type, EncoderDecoderCache):
+                        outputs[key] = past_key_value_type.from_legacy_cache(outputs[key])
+                else:
+                    outputs[key] = None
+
+        return model_output_type(**outputs)
+
+    def _need_fallback(
+        self,
+        seek_sequence,
+        seek_outputs,
+        index,
+        logits_processor,
+        generation_config,
+        vocab_size,
+        temperature,
+    ):
+        needs_fallback = False
+        should_skip = False
+        if generation_config.compression_ratio_threshold is not None:
+            compression_ratio = self._retrieve_compression_ratio(seek_sequence, vocab_size)
+
+            if compression_ratio > generation_config.compression_ratio_threshold:
+                needs_fallback = True
+
+        if generation_config.logprob_threshold is not None:
+            if hasattr(seek_outputs[0], "sequences_scores"):
+                logprobs = [s["sequences_scores"] for s in seek_outputs][index]
+            else:
+                scores = seek_outputs[index]["scores"]
+                logprobs = self._retrieve_avg_logprobs(
+                    scores, seek_sequence, generation_config.eos_token_id, temperature
+                )
+
+            if logprobs < generation_config.logprob_threshold:
+                needs_fallback = True
+
+        if generation_config.no_speech_threshold is not None:
+            no_speech_prob = _get_attr_from_logit_processors(
+                logits_processor, WhisperNoSpeechDetection, "no_speech_prob"
+            )
+
+            if (
+                logprobs < generation_config.logprob_threshold
+                and no_speech_prob[index] > generation_config.no_speech_threshold
+            ):
+                needs_fallback = False
+                should_skip = True
+
+        return needs_fallback, should_skip
+
+    def _expand_variables_for_generation(
+        self, input_features, seek, max_frames, init_tokens, batch_size, condition_on_prev_tokens, generation_config
+    ):
+        if generation_config.num_return_sequences is not None and generation_config.num_return_sequences > 1:
+            batch_idx_map = list(range(batch_size * generation_config.num_return_sequences))
+            cur_bsz = len(batch_idx_map)
+            do_condition_on_prev_tokens = [condition_on_prev_tokens for _ in range(len(batch_idx_map))]
+            input_features = input_features.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            seek = seek.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            max_frames = max_frames.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            init_tokens = init_tokens.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            generation_config.num_return_sequences = 1
+        else:
+            cur_bsz = batch_size
+            batch_idx_map = list(range(cur_bsz))
+            do_condition_on_prev_tokens = [condition_on_prev_tokens for _ in range(cur_bsz)]
+
+        return (
+            batch_idx_map,
+            cur_bsz,
+            input_features,
+            seek,
+            max_frames,
+            init_tokens,
+            do_condition_on_prev_tokens,
+        )
+
+    @staticmethod
+    def _setup_no_speech_detection(logits_processor, segment_input, decoder_input_ids, kwargs):
+        set_inputs = _get_attr_from_logit_processors(logits_processor, WhisperNoSpeechDetection, "set_inputs")
+        extra_kwargs = {k: v for k, v in kwargs.items() if torch.is_tensor(v)}
+        set_inputs({"inputs": segment_input, "decoder_input_ids": decoder_input_ids, **extra_kwargs})
+
+    @staticmethod
+    def _retrieve_total_input_frames(input_features, input_stride, kwargs):
+        if input_features is not None:
+            return input_features.shape[0], input_features.shape[-1]
+
+        if "encoder_outputs" in kwargs:
+            encoder_outputs_shape = (
+                kwargs["encoder_outputs"][0].shape
+                if isinstance(kwargs["encoder_outputs"], BaseModelOutput)
+                else kwargs["encoder_outputs"].shape
+            )
+            return encoder_outputs_shape[0], encoder_outputs_shape[1] * input_stride
+
+        raise ValueError("Make sure to provide either `input_features` or `encoder_outputs` to `generate`.")
+
+    @staticmethod
+    def _maybe_warn_unused_inputs(
+        condition_on_prev_tokens,
+        temperature,
+        compression_ratio_threshold,
+        logprob_threshold,
+        no_speech_threshold,
+        total_input_frames,
+    ):
+        warning_prefix = (
+            f"Audio input consists of only {total_input_frames}. "
+            "Short-form transcription is activated."
+            "{}, but will be ignored."
+        )
+        if condition_on_prev_tokens is not None:
+            logger.warning(warning_prefix.format(f"condition_on_prev_tokens is set to {condition_on_prev_tokens}"))
+
+        if compression_ratio_threshold is not None:
+            logger.warning(
+                warning_prefix.format(f"compression_ratio_threshold is set to {compression_ratio_threshold}")
+            )
+
+        if logprob_threshold is not None:
+            logger.warning(warning_prefix.format(f"logprob_threshold is set to {logprob_threshold}"))
+
+        if no_speech_threshold is not None:
+            logger.warning(warning_prefix.format(f"no_speech_threshold is set to {no_speech_threshold}"))
+
+        # when passing temperature as a list it cannot just be ignored => throw error in this case
+        if isinstance(temperature, (list, tuple)):
+            raise ValueError(
+                f"Audio input consists of only {total_input_frames}. Short-form transcription is activated."
+                f"temperature cannot be set to {temperature} which can only be used for temperature fallback for long-form generation. Make sure to set `temperature` to a float value or `None` for short-form generation."
+            )
+
+    @staticmethod
+    def _set_return_outputs(return_dict_in_generate, return_token_timestamps, logprob_threshold, generation_config):
+        if return_dict_in_generate is None:
+            return_dict_in_generate = generation_config.return_dict_in_generate
+        else:
+            generation_config.return_dict_in_generate = return_dict_in_generate
+
+        generation_config.return_token_timestamps = return_token_timestamps
+        if return_token_timestamps:
+            generation_config.return_dict_in_generate = True
+            generation_config.output_attentions = True
+            generation_config.output_scores = True
+
+        if logprob_threshold is not None:
+            generation_config.return_dict_in_generate = True
+            generation_config.output_scores = True
+
+        return return_dict_in_generate
+
+    def _set_return_timestamps(self, return_timestamps, is_shortform, generation_config):
+        if return_timestamps is None and hasattr(generation_config, "return_timestamps"):
+            return_timestamps = generation_config.return_timestamps
+
+        if not is_shortform:
+            if return_timestamps is False:
+                raise ValueError(
+                    "You have passed more than 3000 mel input features (> 30 seconds) which automatically enables long-form generation which "
+                    "requires the model to predict timestamp tokens. Please either pass `return_timestamps=True` or make sure to pass no more than 3000 mel input features."
+                )
+
+            logger.info("Setting `return_timestamps=True` for long-form generation.")
+            return_timestamps = True
+
+        if return_timestamps and not hasattr(generation_config, "no_timestamps_token_id"):
+            raise ValueError(
+                "You are trying to return timestamps, but the generation config is not properly set. "
+                "Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`. "
+                "For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
+            )
+
+        generation_config.return_timestamps = return_timestamps
+
+        if hasattr(generation_config, "no_timestamps_token_id"):
+            timestamp_begin = generation_config.no_timestamps_token_id + 1
+        else:
+            # BC for models missing the `no_timestamps_token_id` in the generation config when generating short-form with no timestamps
+            # We set the timestamp begin token larger than the vocab size, such that the timestamp condition is never met in the decoding loop
+            timestamp_begin = self.config.vocab_size + 1
+
+        return timestamp_begin
+
+    @staticmethod
+    def _set_language_and_task(language, task, is_multilingual, generation_config):
+        if is_multilingual is not None:
+            if not hasattr(generation_config, "is_multilingual"):
+                raise ValueError(
+                    "The generation config is outdated and is thus not compatible with the `is_multilingual` argument "
+                    "to `generate`. Please update the generation config as per the instructions "
+                    "https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
+                )
+            generation_config.is_multilingual = is_multilingual
+
+        if hasattr(generation_config, "is_multilingual") and not generation_config.is_multilingual:
+            if task is not None or language is not None:
+                raise ValueError(
+                    "Cannot specify `task` or `language` for an English-only model. If the model is intended to be "
+                    "multilingual, pass `is_multilingual=True` to generate, or update the generation config."
+                )
+
+        if language is not None:
+            if not hasattr(generation_config, "lang_to_id"):
+                raise ValueError(
+                    "The generation config is outdated and is thus not compatible with the `language` argument "
+                    "to `generate`. Either set the language using the `forced_decoder_ids` in the model config, "
+                    "or update the generation config as per the instructions https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
+                )
+            generation_config.language = language
+
+        if task is not None:
+            if not hasattr(generation_config, "task_to_id"):
+                raise ValueError(
+                    "The generation config is outdated and is thus not compatible with the `task` argument "
+                    "to `generate`. Either set the task using the `forced_decoder_ids` in the model config, "
+                    "or update the generation config as per the instructions https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
+                )
+            generation_config.task = task
+
+    def _retrieve_init_tokens(self, input_features, batch_size, generation_config, config, num_segment_frames, kwargs):
+        def replace_or_add(lst: List[int], num: int, itr: Iterator[int]):
+            """short function to replace num with a itr in lst"""
+            found = any(i in lst for i in itr)
+            if found:
+                lst = [num if i in itr else i for i in lst]
+            else:
+                lst.append(num)
+            return lst
+
+        def language_to_id(language: str) -> int:
+            language = language.lower()
+            if language in generation_config.lang_to_id.keys():
+                language_token = language
+            elif language in TO_LANGUAGE_CODE.keys():
+                language_token = f"<|{TO_LANGUAGE_CODE[language]}|>"
+            elif language in TO_LANGUAGE_CODE.values():
+                language_token = f"<|{language}|>"
+            else:
+                is_language_code = len(language) == 2
+                raise ValueError(
+                    f"Unsupported language: {language}. Language should be one of:"
+                    f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
+                )
+            if language_token not in generation_config.lang_to_id:
+                raise ValueError(
+                    f"{language_token} is not supported by this specific model as it is not in the `generation_config.lang_to_id`."
+                    "(You should just add it to the generation config)"
+                )
+
+            return generation_config.lang_to_id[language_token]
+
+        task = getattr(generation_config, "task", None)
+        language = getattr(generation_config, "language", None)
+
+        forced_decoder_ids = generation_config.forced_decoder_ids
+        if forced_decoder_ids is not None:
+            if language is None and task is None and forced_decoder_ids[0][1] is None:
+                logger.warning_once(
+                    "Due to a bug fix in https://github.com/huggingface/transformers/pull/28687 transcription using a multilingual Whisper will default to language detection followed by transcription instead of translation to English."
+                    "This might be a breaking change for your use case. If you want to instead always translate your audio to English, make sure to pass `language='en'`."
+                )
+        elif hasattr(config, "forced_decoder_ids") and config.forced_decoder_ids is not None:
+            forced_decoder_ids = config.forced_decoder_ids
+
+        if forced_decoder_ids is not None and task is not None:
+            logger.warning_once(
+                f"You have passed task={task}, but also have set `forced_decoder_ids` to {forced_decoder_ids} which creates a conflict. `forced_decoder_ids` will be ignored in favor of task={task}."
+            )
+            forced_decoder_ids = None
+        elif forced_decoder_ids is not None and language is not None:
+            logger.warning_once(
+                f"You have passed language={language}, but also have set `forced_decoder_ids` to {forced_decoder_ids} which creates a conflict. `forced_decoder_ids` will be ignored in favor of language={language}."
+            )
+            forced_decoder_ids = None
+
+        init_tokens = [generation_config.decoder_start_token_id]
+        if forced_decoder_ids is not None and forced_decoder_ids[0][0] == 1:
+            i = 1
+            while len(forced_decoder_ids) > 0 and forced_decoder_ids[0][0] == i:
+                init_tokens += [forced_decoder_ids[0][1]]
+                forced_decoder_ids = forced_decoder_ids[1:]
+                i += 1
+
+            if len(forced_decoder_ids) > 0:
+                raise ValueError(
+                    f"You are using token ids in `forced_decoder_ids` that do not seem to correctly follow the prompt pattern of Whisper. Make sure that {forced_decoder_ids} has an entry for all indices >= 1 and < {forced_decoder_ids[0][0]}.",
+                )
+
+        # from v4.39 the forced decoder ids are always None in favour of decoder input ids
+        generation_config.forced_decoder_ids = None
+
+        is_lang_id_undefined = len(init_tokens) <= 1 or (len(init_tokens) > 1 and init_tokens[1] is None)
+
+        # Make sure language is a list of strings of the correct length
+        if isinstance(language, (list, tuple)):
+            if any(l is None for l in language):
+                raise TypeError(
+                    "Expected `language` to be `None`, a single string (e.g. `'en'`), or a list of strings with length equal to the batch size (e.g. `('en', 'fr')` for a batch size of 2). Got a list containing `None`."
+                )
+            if len(language) != batch_size:
+                raise ValueError(
+                    "When passing a list of languages, the length of the list must match the batch size. "
+                    f"Expected length of {batch_size}, but got {len(language)} languages."
+                )
+            languages = language
+        elif language is None:
+            # Language will be detected for each item in batch
+            languages = [None] * batch_size
+        else:
+            languages = [language]  # Use a length-1 list now, broadcast later
+
+        # Separate init_tokens for each language
+        init_tokens = [copy.copy(init_tokens) for _ in languages]
+
+        # Update init_tokens with languages
+        lang_ids = None
+        if language is not None:
+            lang_ids = [language_to_id(l) for l in languages]
+        elif hasattr(generation_config, "lang_to_id") and is_lang_id_undefined:
+            # language is not defined or intentially set to `None` to trigger language detection
+            lang_ids = self.detect_language(
+                input_features=input_features,
+                encoder_outputs=kwargs.get("encoder_outputs", None),
+                generation_config=generation_config,
+                num_segment_frames=num_segment_frames,
+            ).tolist()
+        if lang_ids is not None:
+            # append or replace lang_ids to init_tokens
+            for i in range(len(init_tokens)):
+                if len(init_tokens[i]) > 1:
+                    init_tokens[i][1] = lang_ids[i]
+                else:
+                    init_tokens[i].append(lang_ids[i])
+        del languages
+
+        # Update init_tokens with task
+        for i in range(len(init_tokens)):
+            if task is not None:
+                if task in TASK_IDS:
+                    init_tokens[i].append(generation_config.task_to_id[generation_config.task])
+                    task_id = generation_config.task_to_id[generation_config.task]
+
+                    # if task is defined it'll overwrite task ids that might have already been defined via the generation_config
+                    replace_or_add(init_tokens[i], task_id, generation_config.task_to_id.values())
+                else:
+                    raise ValueError(f"The `{task}`task is not supported. The task should be one of `{TASK_IDS}`")
+            elif language is not None and hasattr(generation_config, "task_to_id"):
+                # if language is defined, but no task id is in `init_tokens`, default to transcribe
+                if not any(ti in init_tokens[i] for ti in generation_config.task_to_id.values()):
+                    init_tokens[i].append(generation_config.task_to_id["transcribe"])
+
+            if (
+                not generation_config.return_timestamps
+                and hasattr(generation_config, "no_timestamps_token_id")
+                and init_tokens[i][-1] != generation_config.no_timestamps_token_id
+            ):
+                init_tokens[i].append(generation_config.no_timestamps_token_id)
+            elif (
+                generation_config.return_timestamps and init_tokens[i][-1] == generation_config.no_timestamps_token_id
+            ):
+                logger.info(
+                    "<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `'True'`."
+                )
+                init_tokens[i] = init_tokens[i][:-1]
+
+            # let's make sure we don't pass `None` tokens as prompt tokens
+            init_tokens[i] = [t for t in init_tokens[i] if t is not None]
+
+        return torch.as_tensor(init_tokens, dtype=torch.long, device=self.device).expand(batch_size, -1)
+
+    def detect_language(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        encoder_outputs: Optional[Union[torch.FloatTensor, BaseModelOutput]] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        num_segment_frames: int = 3000,
+    ) -> torch.Tensor:
+        """
+        Detects language from log-mel input features or encoder_outputs
+
+        Parameters:
+            input_features (`torch.Tensor` of shape `(batch_size, feature_size, sequence_length)`, *optional*):
+                Float values of log-mel features extracted from the raw speech waveform. The raw speech waveform can be obtained by
+                loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+                the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+                [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+                tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] for details.
+            encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+                Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+                `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+                hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            num_segment_frames (`int`, *optional*, defaults to 3000):
+                The number of log-mel frames the model expects
+
+        Return:
+            A `torch.LongTensor` representing the detected language ids.
+        """
+        if input_features is None and encoder_outputs is None:
+            raise ValueError("You have to specify either `input_features` or `encoder_outputs`")
+        elif input_features is not None and encoder_outputs is not None:
+            raise ValueError("Make sure to specificy only one of `input_features` or `encoder_outputs` - not both!")
+        elif input_features is not None:
+            inputs = {"input_features": input_features[:, :, :num_segment_frames]}
+            batch_size = input_features.shape[0]
+        elif encoder_outputs is not None:
+            inputs = {"encoder_outputs": encoder_outputs}
+            batch_size = (
+                encoder_outputs[0].shape[0] if isinstance(encoder_outputs, BaseModelOutput) else encoder_outputs[0]
+            )
+
+        generation_config = generation_config or self.generation_config
+        decoder_input_ids = (
+            torch.ones((batch_size, 1), device=self.device, dtype=torch.long)
+            * generation_config.decoder_start_token_id
+        )
+
+        with torch.no_grad():
+            logits = self(**inputs, decoder_input_ids=decoder_input_ids).logits[:, -1]
+
+        non_lang_mask = torch.ones_like(logits[0], dtype=torch.bool)
+        non_lang_mask[list(generation_config.lang_to_id.values())] = False
+
+        logits[:, non_lang_mask] = -np.inf
+
+        lang_ids = logits.argmax(-1)
+
+        return lang_ids
+
+    @staticmethod
+    def _check_decoder_input_ids(kwargs):
+        decoder_input_ids = kwargs.get("decoder_input_ids", None)
+        assistant_model = kwargs.get("assistant_model", None)
+        if decoder_input_ids is not None and assistant_model is not None:
+            raise ValueError(
+                "Passing `decoder_input_ids` is deprecated. Consider passing `prompt_ids` instead.",
+            )
+
+    @staticmethod
+    def _set_num_frames(return_token_timestamps, generation_config, kwargs):
+        if return_token_timestamps:
+            if getattr(generation_config, "task", None) == "translate":
+                logger.warning("Token-level timestamps may not be reliable for task 'translate'.")
+            if not hasattr(generation_config, "alignment_heads"):
+                raise ValueError(
+                    "Model generation config has no `alignment_heads`, token-level timestamps not available. "
+                    "See https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a on how to add this property to the generation config."
+                )
+            generation_config.num_frames = kwargs.pop("num_frames", None)
+
+    @staticmethod
+    def _set_thresholds_and_condition(
+        generation_config,
+        logprob_threshold,
+        compression_ratio_threshold,
+        no_speech_threshold,
+        condition_on_prev_tokens,
+    ):
+        generation_config.logprob_threshold = (
+            logprob_threshold
+            if logprob_threshold is not None
+            else getattr(generation_config, "logprob_threshold", None)
+        )
+        generation_config.compression_ratio_threshold = (
+            compression_ratio_threshold
+            if compression_ratio_threshold is not None
+            else getattr(generation_config, "compression_ratio_threshold", None)
+        )
+        generation_config.no_speech_threshold = (
+            no_speech_threshold
+            if no_speech_threshold is not None
+            else getattr(generation_config, "no_speech_threshold", None)
+        )
+        generation_config.condition_on_prev_tokens = (
+            condition_on_prev_tokens
+            if condition_on_prev_tokens is not None
+            else getattr(generation_config, "condition_on_prev_tokens", None)
+        )
+
+    @staticmethod
+    def _set_prompt_condition_type(generation_config, prompt_condition_type):
+        allowed_cond_types = ["first-segment", "all-segments"]
+
+        # default to "first-segment"
+        prompt_condition_type = prompt_condition_type or allowed_cond_types[0]
+
+        if prompt_condition_type not in allowed_cond_types:
+            raise ValueError(
+                f"`prompt_condition_type={prompt_condition_type} does not exist. Make sure to set `prompt_condition_type` to one of {', '.join(allowed_cond_types)}"
+            )
+
+        if generation_config.condition_on_prev_tokens is not True and prompt_condition_type == "all-segments":
+            raise ValueError(
+                "Make sure to set `condition_on_prev_tokens=True` when setting `prompt_condition_type='all-segments'`."
+            )
+
+        generation_config.prompt_condition_type = prompt_condition_type
+
+    @staticmethod
+    def _set_condition_on_prev_tokens(condition_on_prev_tokens, generation_config):
+        condition_on_prev_tokens = (
+            condition_on_prev_tokens
+            if condition_on_prev_tokens is not None
+            else getattr(generation_config, "condition_on_prev_tokens", False)
+        )
+        generation_config.condition_on_prev_tokens = condition_on_prev_tokens
+
+    @staticmethod
+    def _retrieve_max_frames_and_seek(batch_size, attention_mask, total_input_frames, is_shortform):
+        if batch_size > 1 and not is_shortform and attention_mask is None:
+            raise ValueError(
+                "When doing batched long-form audio transcription, make sure to pass an `attention_mask`. You can retrieve the `attention_mask` by doing `processor(audio, ..., return_attention_mask=True)` "
+            )
+        elif batch_size > 1 and not is_shortform:
+            max_frames = attention_mask.sum(-1).cpu().to(torch.long)
+            seek = torch.zeros((batch_size,), dtype=torch.long)
+        else:
+            max_frames = torch.ones((batch_size,), dtype=torch.long) * total_input_frames
+            seek = torch.zeros((batch_size,), dtype=torch.long)
+
+        return max_frames, seek
+
+    def _retrieve_logit_processors(self, generation_config, logits_processor, begin_index, num_beams, device):
+        if generation_config.return_timestamps is True:
+            timestamp_processor = WhisperTimeStampLogitsProcessor(generation_config, begin_index=begin_index)
+            logits_processor = (
+                [timestamp_processor] if logits_processor is None else [timestamp_processor] + logits_processor
+            )
+
+        if generation_config.suppress_tokens is not None:
+            suppress_tokens_processor = SuppressTokensLogitsProcessor(generation_config.suppress_tokens, device=device)
+            logits_processor = (
+                [suppress_tokens_processor]
+                if logits_processor is None
+                else [suppress_tokens_processor] + logits_processor
+            )
+            generation_config.suppress_tokens = None
+
+        if generation_config.begin_suppress_tokens is not None:
+            begin_suppress_processor = SuppressTokensAtBeginLogitsProcessor(
+                generation_config.begin_suppress_tokens, begin_index=begin_index, device=device
+            )
+            logits_processor = (
+                [begin_suppress_processor]
+                if logits_processor is None
+                else [begin_suppress_processor] + logits_processor
+            )
+            generation_config.begin_suppress_tokens = None
+
+        if generation_config.no_speech_threshold is not None:
+            no_speech_detector = WhisperNoSpeechDetection(
+                no_speech_token=generation_config.no_timestamps_token_id - 1,
+                begin_index=begin_index,
+                scores_is_logprobs=num_beams > 1,
+            )
+            logits_processor = (
+                [no_speech_detector] if logits_processor is None else [no_speech_detector] + logits_processor
+            )
+            no_speech_detector.set_model(self)
+
+        return logits_processor
+
+    @staticmethod
+    def _maybe_reduce_batch(input_features, seek, max_frames, cur_bsz, batch_idx_map):
+        prev_bsz = cur_bsz
+        new_batch_idx_map = []
+        for i in range(prev_bsz):
+            prev_i = batch_idx_map[i]
+            if seek[prev_i] >= max_frames[prev_i]:
+                cut_index = i + (cur_bsz - prev_bsz)
+                cur_bsz -= 1
+                input_features = torch.cat([input_features[:cut_index], input_features[cut_index + 1 :]], dim=0)
+            else:
+                # cut out index that goes away
+                new_batch_idx_map.append(prev_i)
+
+        return input_features, cur_bsz, new_batch_idx_map
+
+    @staticmethod
+    def _get_input_segment(input_features, seek, seek_num_frames, num_segment_frames, cur_bsz, batch_idx_map):
+        if input_features is None:
+            return None
+
+        segment_input = []
+        for i in range(cur_bsz):
+            prev_i = batch_idx_map[i]
+            segment_input_slice = input_features[i : i + 1, :, seek[prev_i] : seek[prev_i] + seek_num_frames[prev_i]]
+
+            if segment_input_slice.shape[-1] < num_segment_frames:
+                # pad to 3000 if necessary
+                segment_input_slice = F.pad(
+                    segment_input_slice, pad=(0, num_segment_frames - segment_input_slice.shape[-1])
+                )
+
+            segment_input.append(segment_input_slice)
+
+        segment_input = torch.cat(segment_input, dim=0)
+
+        return segment_input
+
+    @staticmethod
+    def _prepare_decoder_input_ids(
+        cur_bsz,
+        init_tokens,
+        current_segments,
+        batch_idx_map,
+        do_condition_on_prev_tokens,
+        prompt_ids,
+        generation_config,
+        config,
+        device,
+        suppress_tokens,
+        timestamp_begin,
+        kwargs,
+    ):
+        if "decoder_input_ids" in kwargs:
+            decoder_input_ids = kwargs.pop("decoder_input_ids")
+
+            return decoder_input_ids, kwargs
+
+        cut_off_length = config.max_target_positions // 2 - 1
+
+        decoder_input_ids = init_tokens[batch_idx_map]
+
+        prev_start_of_text = getattr(generation_config, "prev_sot_token_id", None)
+        if prev_start_of_text is None:
+            prev_start_of_text = suppress_tokens[-2] if suppress_tokens is not None else None
+
+        if any(do_condition_on_prev_tokens) and len(current_segments[0]) > 0:
+            # according to https://github.com/openai/whisper/blob/e58f28804528831904c3b6f2c0e473f346223433/whisper/decoding.py#L609
+            active_segments = [current_segments[i] if do_condition_on_prev_tokens[i] else None for i in batch_idx_map]
+
+            for segments in active_segments:
+                for seg in segments:
+                    if len(seg["tokens"]) > 2 and seg["tokens"][-2] >= timestamp_begin:
+                        # the segment finishes with two timestamp tokens
+                        # we need to ignore the last timestamp token
+                        # see https://github.com/huggingface/transformers/pull/34537
+                        seg["tokens"] = seg["tokens"][:-1]
+
+            if prompt_ids is not None and generation_config.prompt_condition_type == "all-segments":
+                prev_ids = prompt_ids
+            else:
+                one_tensor = torch.ones((cur_bsz, 1), device=device, dtype=torch.long)
+                prev_ids = prev_start_of_text * one_tensor[0] if prev_start_of_text is not None else None
+
+            padding = "max_length" if generation_config.cache_implementation == "static" else "longest"
+
+            prev_tokens = _pad_to_max_length(
+                active_segments,
+                generation_config.pad_token_id,
+                device=device,
+                padding_side="left",
+                padding=padding,
+                bos_token_tensor=prev_ids,
+                cut_off_length=cut_off_length,
+            )
+            decoder_input_ids = torch.cat([prev_tokens, decoder_input_ids], dim=-1)
+
+            kwargs["decoder_attention_mask"] = decoder_input_ids != generation_config.pad_token_id
+        elif prompt_ids is not None:
+            prev_tokens = prompt_ids[None].repeat(decoder_input_ids.shape[0], 1)
+            decoder_input_ids = torch.cat([prev_tokens, decoder_input_ids], dim=-1)
+            # make sure `"decoder_attention_mask"` is not passed to forward
+            kwargs.pop("decoder_attention_mask", None)
+        else:
+            # make sure `"decoder_attention_mask"` is not passed to forward
+            kwargs.pop("decoder_attention_mask", None)
+
+        return decoder_input_ids, kwargs
+
+    def _set_max_new_tokens_and_length(self, config, decoder_input_ids, generation_config):
+        max_new_tokens = generation_config.max_new_tokens if generation_config.max_new_tokens is not None else 0
+        if max_new_tokens + decoder_input_ids.shape[-1] > self.config.max_target_positions:
+            raise ValueError(
+                f"The length of `decoder_input_ids`, including special start tokens, prompt tokens, and previous tokens, is {decoder_input_ids.shape[-1]}, "
+                f" and `max_new_tokens` is {max_new_tokens}. Thus, the combined length of "
+                f"`decoder_input_ids` and `max_new_tokens` is: {max_new_tokens + decoder_input_ids.shape[-1]}. This exceeds the "
+                f"`max_target_positions` of the Whisper model: {self.config.max_target_positions}. "
+                "You should either reduce the length of your prompt, or reduce the value of `max_new_tokens`, "
+                f"so that their combined length is less than {self.config.max_target_positions}."
+            )
+
+        num_initial_tokens = min(config.max_target_positions // 2 - 1, decoder_input_ids.shape[-1] - 1)
+
+        # Make sure we don't get larger than `max_length`
+        if generation_config.max_length is not None and generation_config.max_new_tokens is None:
+            max_length = min(generation_config.max_length + num_initial_tokens, config.max_target_positions)
+            logger.info(
+                f"Increase max_length from {generation_config.max_length} to {max_length} since input is conditioned on previous segment."
+            )
+        elif (
+            generation_config.max_new_tokens is not None
+            and generation_config.max_new_tokens + decoder_input_ids.shape[-1] > config.max_target_positions
+        ):
+            max_new_tokens = config.max_target_positions - decoder_input_ids.shape[-1]
+            generation_config.max_new_tokens = max_new_tokens
+
+    @staticmethod
+    def _retrieve_compression_ratio(tokens, vocab_size):
+        """Compute byte length of zlib compressed token bytes vs. byte length of raw token bytes"""
+        length = int(math.log2(vocab_size) / 8) + 1
+        token_bytes = b"".join([t.to_bytes(length, "little") for t in tokens.tolist()])
+        compression_ratio = len(token_bytes) / len(zlib.compress(token_bytes))
+
+        return compression_ratio
+
+    @staticmethod
+    def _retrieve_avg_logprobs(scores, tokens, eos_token_id, temperature):
+        rescale_temperature = temperature if temperature > 0.0 else 1
+        scores = torch.stack(scores).to(tokens.device)
+
+        if scores.shape[0] > tokens.shape[0]:
+            scores = scores[: tokens.shape[0]]
+        else:
+            tokens = tokens[-scores.shape[0] :]
+
+        logprobs = F.log_softmax((scores * rescale_temperature).float(), dim=-1).to(scores.dtype)
+
+        # retrieve logprob of selected tokens and sum
+        sum_logprobs = sum((logprobs[i][tokens[i]] * (tokens[i] != eos_token_id)) for i in range(logprobs.shape[0]))
+        length = (tokens != eos_token_id).sum(-1) if eos_token_id is not None else tokens.shape[0]
+
+        avg_logprobs = sum_logprobs / (length + 1)
+        return avg_logprobs
+
+    @staticmethod
+    def _retrieve_segment(
+        seek_sequence,
+        seek_outputs,
+        time_offset,
+        timestamp_begin,
+        seek_num_frames,
+        time_precision,
+        time_precision_features,
+        input_stride,
+        prev_idx,
+        idx,
+        return_token_timestamps,
+    ):
+        # find the predicted "end of segment" predictions of Whisper
+        # "end of segment" predictions occur whenever Whisper predicts a timestamp token
+        timestamp_tokens: torch.Tensor = seek_sequence.ge(timestamp_begin)
+        single_timestamp_ending = timestamp_tokens[-2:].tolist() == [False, True]
+        timestamp_segment_indices = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0]
+        timestamp_segment_indices.add_(1)
+        token_timestamps = seek_outputs[idx]["token_timestamps"] if return_token_timestamps else []
+        device = seek_sequence.device
+
+        # If whisper predicted a "end of segment" via a timestep token, let's go ever each
+        # "end of segment" prediction and slice the decoding into segments accordingly
+        if len(timestamp_segment_indices) > 0:
+            # if the output contains two consecutive timestamp tokens
+            slices = timestamp_segment_indices.tolist()
+            segments = []
+            if single_timestamp_ending:
+                slices.append(len(seek_sequence))
+            else:
+                # we want to include the last timestamp token in the last segment to know it was no single ending
+                slices[-1] += 1
+
+            last_slice = 0
+            # Add each segment to list of all segments
+            for i, current_slice in enumerate(slices):
+                is_last_slice = i == len(slices) - 1
+                sliced_tokens = seek_sequence[last_slice:current_slice]
+                start_timestamp_pos = sliced_tokens[0] - timestamp_begin
+                idx_sliced_tokens = -1 if not is_last_slice or single_timestamp_ending else -2
+                end_timestamp_pos = sliced_tokens[idx_sliced_tokens] - timestamp_begin
+                segments.append(
+                    {
+                        "start": time_offset[prev_idx]
+                        + start_timestamp_pos.to(torch.float32 if device.type == "mps" else torch.float64)
+                        * time_precision,
+                        "end": time_offset[prev_idx]
+                        + end_timestamp_pos.to(torch.float32 if device.type == "mps" else torch.float64)
+                        * time_precision,
+                        "tokens": sliced_tokens,
+                        "result": seek_outputs[idx],
+                    }
+                )
+                if return_token_timestamps:
+                    segments[-1]["token_timestamps"] = (
+                        token_timestamps[last_slice:current_slice] + time_offset[prev_idx]
+                    )
+                last_slice = current_slice
+
+            if single_timestamp_ending:
+                # single timestamp at the end means no speech after the last timestamp.
+                segment_offset = seek_num_frames[prev_idx]
+            else:
+                # otherwise, ignore the unfinished segment and seek to the last timestamp
+                # here we throw away all predictions after the last predicted "end of segment"
+                # since we are cutting right in the middle of an audio
+                last_timestamp_pos = seek_sequence[last_slice - 2].item() - timestamp_begin
+                segment_offset = last_timestamp_pos * input_stride
+        else:
+            # If whisper does not predict any "end of segment" token, then
+            # the whole decoding is considered a segment and we add it to the list of segments
+            timestamps = seek_sequence[timestamp_tokens.nonzero().flatten()]
+            last_timestamp_pos = int(seek_num_frames[prev_idx] * time_precision_features / time_precision)
+            if timestamps.numel() > 0 and timestamps[-1] != timestamp_begin:
+                # no consecutive timestamps but it has a timestamp; use the last one.
+                last_timestamp_pos = (timestamps[-1] - timestamp_begin).to(
+                    torch.float32 if device.type == "mps" else torch.float64
+                )
+            segments = [
+                {
+                    "start": time_offset[prev_idx],
+                    "end": time_offset[prev_idx] + last_timestamp_pos * time_precision,
+                    "tokens": seek_sequence,
+                    "result": seek_outputs[idx],
+                }
+            ]
+            if return_token_timestamps:
+                segments[-1]["token_timestamps"] = token_timestamps + time_offset[prev_idx]
+            segment_offset = seek_num_frames[prev_idx]
+
+        return segments, segment_offset

generation_whisper.py

@@ -0,0 +1,1881 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import copy
+import math
+import warnings
+import zlib
+from typing import Callable, Iterator, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.cache_utils import EncoderDecoderCache
+
+from ...generation import GenerationConfig, GenerationMixin
+from ...generation.logits_process import (
+    LogitsProcessorList,
+    SuppressTokensAtBeginLogitsProcessor,
+    SuppressTokensLogitsProcessor,
+    WhisperNoSpeechDetection,
+    WhisperTimeStampLogitsProcessor,
+)
+from ...generation.stopping_criteria import StoppingCriteriaList
+from ...modeling_outputs import BaseModelOutput
+from ...utils import logging
+from .tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
+
+
+logger = logging.get_logger(__name__)
+
+
+def _median_filter(inputs: torch.Tensor, filter_width: int) -> torch.Tensor:
+    """
+    Applies a median filter of width `filter_width` along the last dimension of the input.
+
+    The `inputs` tensor is assumed to be 3- or 4-dimensional.
+    """
+    if filter_width <= 0 or filter_width % 2 != 1:
+        raise ValueError("`filter_width` should be an odd number")
+
+    pad_width = filter_width // 2
+    if inputs.shape[-1] <= pad_width:
+        return inputs
+
+    # Pad the left and right edges.
+    inputs = nn.functional.pad(inputs, (pad_width, pad_width, 0, 0), mode="reflect")
+
+    # sort() is faster than torch.median (https://github.com/pytorch/pytorch/issues/51450)
+    result = inputs.unfold(-1, filter_width, 1).sort()[0][..., pad_width]
+    return result
+
+
+def _dynamic_time_warping(matrix: np.ndarray):
+    """
+    Measures similarity between two temporal sequences: the input audio and the output tokens. Used to generate
+    token-level timestamps.
+    """
+    output_length, input_length = matrix.shape
+    cost = np.ones((output_length + 1, input_length + 1), dtype=np.float32) * np.inf
+    trace = -np.ones((output_length + 1, input_length + 1), dtype=np.float32)
+
+    cost[0, 0] = 0
+    for j in range(1, input_length + 1):
+        for i in range(1, output_length + 1):
+            c0 = cost[i - 1, j - 1]
+            c1 = cost[i - 1, j]
+            c2 = cost[i, j - 1]
+
+            if c0 < c1 and c0 < c2:
+                c, t = c0, 0
+            elif c1 < c0 and c1 < c2:
+                c, t = c1, 1
+            else:
+                c, t = c2, 2
+
+            cost[i, j] = matrix[i - 1, j - 1] + c
+            trace[i, j] = t
+
+    # backtrace
+    i = trace.shape[0] - 1
+    j = trace.shape[1] - 1
+    trace[0, :] = 2
+    trace[:, 0] = 1
+
+    text_indices = []
+    time_indices = []
+    while i > 0 or j > 0:
+        text_indices.append(i - 1)
+        time_indices.append(j - 1)
+        if trace[i, j] == 0:
+            i -= 1
+            j -= 1
+        elif trace[i, j] == 1:
+            i -= 1
+        elif trace[i, j] == 2:
+            j -= 1
+        else:
+            raise RuntimeError(
+                f"Internal error in dynamic time warping. Unexpected trace[{i}, {j}]. Please file a bug report."
+            )
+
+    text_indices = np.array(text_indices)[::-1]
+    time_indices = np.array(time_indices)[::-1]
+    return text_indices, time_indices
+
+
+def _get_attr_from_logit_processors(logits_processor, logit_processor_class, attribute_name):
+    if logits_processor is not None:
+        logit_processor = next((cls for cls in logits_processor if isinstance(cls, logit_processor_class)), None)
+        if logit_processor:
+            return getattr(logit_processor, attribute_name, None)
+    return None
+
+
+def _pad_to_max_length(
+    current_segments,
+    pad_token_id,
+    device,
+    padding_side="right",
+    padding="longest",
+    bos_token_tensor=None,
+    cut_off_length=None,
+):
+    max_total_length = 0
+    sequences = []
+
+    if padding_side not in ["right", "left"]:
+        raise ValueError(f"`padding_side` must be either 'right' or 'left', not {padding_side}")
+
+    if padding not in ["longest", "max_length"]:
+        raise ValueError(f"`padding` must be either 'longest' or 'max_length', not {padding}")
+    elif padding == "max_length" and cut_off_length is None:
+        raise ValueError("`cut_off_length` must be specified when `padding='max_length'`")
+
+    for current_segment_list in current_segments:
+        if current_segment_list is not None and len([d["tokens"] for d in current_segment_list]) > 0:
+            sequence = torch.cat([d["tokens"] for d in current_segment_list], dim=-1)
+
+            if cut_off_length is not None:
+                sequence = sequence[-cut_off_length:]
+
+            if bos_token_tensor is not None:
+                sequence = torch.cat([bos_token_tensor, sequence])
+
+            sequences.append(sequence)
+            max_total_length = max(max_total_length, len(sequences[-1]))
+        elif bos_token_tensor is not None:
+            sequences.append(bos_token_tensor)
+        else:
+            sequences.append(torch.tensor([], device=device))
+
+    max_total_length = cut_off_length + 1 if padding == "max_length" else max_total_length
+    for i in range(len(current_segments)):
+        pad_length = max_total_length - len(sequences[i])
+        pad = (0, pad_length) if padding_side == "right" else (pad_length, 0)
+        sequences[i] = F.pad(sequences[i], pad=pad, value=pad_token_id)
+
+    sequences = torch.stack(sequences, dim=0)
+    return sequences
+
+
+class WhisperGenerationMixin(GenerationMixin):
+    def _extract_token_timestamps(
+        self, generate_outputs, alignment_heads, time_precision=0.02, num_frames=None, num_input_ids=None
+    ):
+        """
+        Calculates token-level timestamps using the encoder-decoder cross-attentions and dynamic time-warping (DTW) to
+        map each output token to a position in the input audio. If `num_frames` is specified, the encoder-decoder
+        cross-attentions will be cropped before applying DTW.
+
+        Returns:
+            tensor containing the timestamps in seconds for each predicted token
+        """
+        # Create a list with `decoder_layers` elements, each a tensor of shape
+        # (batch size, attention_heads, output length, input length).
+        cross_attentions = []
+        for i in range(self.config.decoder_layers):
+            cross_attentions.append(torch.cat([x[i] for x in generate_outputs.cross_attentions], dim=2))
+
+        # Select specific cross-attention layers and heads. This is a tensor
+        # of shape (batch size, num selected, output length, input length).
+        weights = torch.stack([cross_attentions[l][:, h] for l, h in alignment_heads])
+        weights = weights.permute([1, 0, 2, 3])
+
+        weight_length = None
+
+        if "beam_indices" in generate_outputs:
+            # If beam search has been used, the output sequences may have been generated for more timesteps than their sequence_lengths
+            # since the beam search strategy chooses the most probable sequences at the end of the search.
+            # In that case, the cross_attentions weights are too long and we have to make sure that they have the right output_length
+            weight_length = (generate_outputs.beam_indices != -1).sum(-1).max()
+            weight_length = weight_length if num_input_ids is None else weight_length + num_input_ids
+
+            # beam search takes `decoder_input_ids` into account in the `beam_indices` length
+            # but forgot to shift the beam_indices by the number of `decoder_input_ids`
+            beam_indices = torch.zeros_like(generate_outputs.beam_indices[:, :weight_length])
+            # we actually shif the beam indices here
+            beam_indices[:, num_input_ids:] = generate_outputs.beam_indices[:, : weight_length - num_input_ids]
+
+            weights = weights[:, :, :weight_length]
+
+            # If beam index is still -1, it means that the associated token id is EOS
+            # We need to replace the index with 0 since index_select gives an error if any of the indexes is -1.
+            beam_indices = beam_indices.masked_fill(beam_indices == -1, 0)
+
+            # Select the cross attention from the right beam for each output sequences
+            weights = torch.stack(
+                [
+                    torch.index_select(weights[:, :, i, :], dim=0, index=beam_indices[:, i])
+                    for i in range(beam_indices.shape[1])
+                ],
+                dim=2,
+            )
+
+        # make sure timestamps are as long as weights
+        input_length = weight_length or cross_attentions[0].shape[2]
+        batch_size = generate_outputs.sequences.shape[0]
+        timestamps = torch.zeros(
+            (batch_size, input_length + 1), dtype=torch.float32, device=generate_outputs.sequences.device
+        )
+
+        if num_frames is not None:
+            # two cases:
+            # 1. num_frames is the same for each sample -> compute the DTW matrix for each sample in parallel
+            # 2. num_frames is different, compute the DTW matrix for each sample sequentially
+
+            # we're using np.unique because num_frames can be int/list/tuple
+            if isinstance(num_frames, int):
+                weights = weights[..., : num_frames // 2]
+
+            elif isinstance(num_frames, (list, tuple, np.ndarray)) and len(np.unique(num_frames)) == 1:
+                weights = weights[..., : num_frames[0] // 2]
+
+            elif isinstance(num_frames, (torch.Tensor)) and len(torch.unique(num_frames)) == 1:
+                weights = weights[..., : num_frames[0] // 2]
+
+            else:
+                # num_frames is of shape (batch_size,) whereas batch_size is truely batch_size*num_return_sequences
+                repeat_time = batch_size if isinstance(num_frames, int) else batch_size // len(num_frames)
+                num_frames = num_frames.cpu() if isinstance(num_frames, (torch.Tensor)) else num_frames
+                num_frames = np.repeat(num_frames, repeat_time)
+
+        if num_frames is None or isinstance(num_frames, int):
+            # Normalize and smoothen the weights.
+            std = torch.std(weights, dim=-2, keepdim=True, unbiased=False)
+            mean = torch.mean(weights, dim=-2, keepdim=True)
+            weights = (weights - mean) / std
+            weights = _median_filter(weights, self.config.median_filter_width)
+
+            # Average the different cross-attention heads.
+            weights = weights.mean(dim=1)
+
+        # Perform dynamic time warping on each element of the batch.
+        for batch_idx in range(batch_size):
+            if num_frames is not None and isinstance(num_frames, (tuple, list, np.ndarray, torch.Tensor)):
+                matrix = weights[batch_idx, ..., : num_frames[batch_idx] // 2]
+
+                # Normalize and smoothen the weights.
+                std = torch.std(matrix, dim=-2, keepdim=True, unbiased=False)
+                mean = torch.mean(matrix, dim=-2, keepdim=True)
+                matrix = (matrix - mean) / std
+                matrix = _median_filter(matrix, self.config.median_filter_width)
+
+                # Average the different cross-attention heads.
+                matrix = matrix.mean(dim=0)
+            else:
+                matrix = weights[batch_idx]
+
+            text_indices, time_indices = _dynamic_time_warping(-matrix.cpu().double().numpy())
+            jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
+            jump_times = time_indices[jumps] * time_precision
+            timestamps[batch_idx, 1:] = torch.tensor(jump_times)
+
+        return timestamps
+
+    def generate(
+        self,
+        input_features: Optional[torch.Tensor] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        logits_processor: Optional[LogitsProcessorList] = None,
+        stopping_criteria: Optional[StoppingCriteriaList] = None,
+        prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+        synced_gpus: bool = False,
+        return_timestamps: Optional[bool] = None,
+        task: Optional[str] = None,
+        language: Optional[Union[str, List[str]]] = None,
+        is_multilingual: Optional[bool] = None,
+        prompt_ids: Optional[torch.Tensor] = None,
+        prompt_condition_type: Optional[str] = None,  # first-segment, all-segments
+        condition_on_prev_tokens: Optional[bool] = None,
+        temperature: Optional[Union[float, Tuple[float, ...]]] = None,
+        compression_ratio_threshold: Optional[float] = None,
+        logprob_threshold: Optional[float] = None,
+        no_speech_threshold: Optional[float] = None,
+        num_segment_frames: Optional[int] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        time_precision: float = 0.02,
+        time_precision_features: float = 0.01,
+        return_token_timestamps: Optional[bool] = None,
+        return_segments: bool = False,
+        return_dict_in_generate: Optional[bool] = None,
+        **kwargs,
+    ):
+        """
+        Transcribes or translates log-mel input features to a sequence of auto-regressively generated token ids.
+
+        <Tip warning={true}>
+
+        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
+        model's default generation configuration. You can override any `generation_config` by passing the corresponding
+        parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
+
+        For an overview of generation strategies and code examples, check out the [following
+        guide](./generation_strategies).
+
+        </Tip>
+
+        Parameters:
+            input_features (`torch.Tensor` of shape `(batch_size, feature_size, sequence_length)`, *optional*):
+                Float values of log-mel features extracted from the raw speech waveform. The raw speech waveform can be obtained by
+                loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+                the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+                [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+                tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] for details.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            logits_processor (`LogitsProcessorList`, *optional*):
+                Custom logits processors that complement the default logits processors built from arguments and
+                generation config. If a logit processor is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            stopping_criteria (`StoppingCriteriaList`, *optional*):
+                Custom stopping criteria that complement the default stopping criteria built from arguments and a
+                generation config. If a stopping criteria is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
+                If provided, this function constraints the beam search to allowed tokens only at each step. If not
+                provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
+                `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
+                on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
+                for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
+                Retrieval](https://arxiv.org/abs/2010.00904).
+            synced_gpus (`bool`, *optional*, defaults to `False`):
+                Whether to continue running the while loop until max_length (needed to avoid deadlocking with
+                `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3).
+            return_timestamps (`bool`, *optional*):
+                Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
+            task (`str`, *optional*):
+                Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
+                will be updated accordingly.
+            language (`str` or list of `str`, *optional*):
+                Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. For
+                batched generation, a list of language tokens can be passed. You can find all the possible language
+                tokens in the `model.generation_config.lang_to_id` dictionary.
+            is_multilingual (`bool`, *optional*):
+                Whether or not the model is multilingual.
+            prompt_ids (`torch.Tensor`, *optional*):
+                Rank-1 tensor of token IDs created by passing text to [`~WhisperProcessor.get_prompt_ids`] that is
+                provided as a prompt to each chunk. This can be used to provide or "prompt-engineer" a context for
+                transcription, e.g. custom vocabularies or proper nouns to make it more likely to predict those words
+                correctly. It cannot be used in conjunction with `decoder_start_token_id` as it overwrites this value.
+            prompt_condition_type (`str`, *optional*):
+                Only relevant for long-form transcription. Condition type of `prompt_ids`. 'first-segment' means only the first segment is conditioned on `prompt_ids`. 'all-segments' means each segment is conditioned on `prompt_ids`. Make sure to enable `condition_on_prev_tokens` for 'all-segments'.
+                Defaults to 'first-segment'. For short-term transcription only 'first-segment' is possible.
+            condition_on_prev_tokens (`bool`, *optional*):
+                Only relevant for long-form transcription. Whether to condition each segment on the previous segment.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            temperature (`float` or list of `float`, *optional*):
+                The temperature to be used for generation. Passing a single `float` value and `do_sample=True` activates
+                generation using sampling. For long-form transcription, temperature fallback can be activated by passing
+                a list of float values such as (0.0, 0.2, 0.4, 0.6, 0.8, 1.0). As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            compression_ratio_threshold (`float`, *optional*):
+                Only relevant for long-form transcription. If defined, the zlib compression rate of each segment will be computed. If the compression rate of
+                a segment is higher than `compression_ratio_threshold`, temperature fallback is activated: the generated segment is discarded and the generation is
+                repeated using a higher temperature. The intuition behind this feature is that segments with very high compression rates
+                suffer from a lot of repetition. The unwanted repetition can be reduced by injecting more randomness by increasing the temperature. If `compression_ratio_threshold` is defined
+                make sure that `temperature` is a list of values. A common value for `compression_ratio_threshold` is 1.35.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            logprob_threshold (`float`, *optional*):
+                Only relevant for long-form transcription. If defined, the average log-probability of each segment will be computed. If the log-probability of
+                a given segment is lower than `logprob_threshold`, temperature fallback is activated: the generated segment is discarded and the generation is
+                repeated using a higher temperature. The intuition behind this feature is that segments of low log-probability
+                can be improved by injecting more randomness by increasing the temperature. If `logprob_threshold` is defined
+                make sure that `temperature` is a list of values. A common value for `logprob_threshold` is -1.0.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            no_speech_threshold (`float`, *optional*):
+                Only relevant for long-form transcription. If defined, the "no-speech" token combined with the `logprob_threshold`
+                is used to determine whether a segment contains only silence. In this case, the transcription for this segment
+                is skipped.
+                As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
+                performance.
+            num_segment_frames (`int`, *optional*):
+                The number of frames a single segment is made of. If not defined, `num_segment_frames` defaults to the model's stride
+                times the maximum input length.
+            attention_mask (`torch.Tensor`, *optional*):
+                `attention_mask` needs to be passed when doing long-form transcription using a batch size > 1.
+            time_precision (`int`, *optional*, defaults to 0.02):
+                The duration of output token in seconds. *E.g.* 0.02 means that a generated token on average accounts
+                for 20 ms.
+            time_precision_features (`int`, *optional*, defaults to 0.01):
+                The duration represented by a feature frame in seconds.
+            return_token_timestamps (`bool`, *optional*):
+                Whether to return token-level timestamps with the text. This can be used with or without the
+                `return_timestamps` option. To get word-level timestamps, use the tokenizer to group the tokens into
+                words.
+            return_segments (`bool`, *optional*, defaults to `False`):
+                Whether to additionally return a list of all segments. Note that this option can only be enabled
+                when doing long-form transcription.
+            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
+                Whether or not to return a [`~utils.ModelOutput`] instead of just returning the generated tokens.
+                Note that when doing long-form transcription, `return_dict_in_generate` can only be enabled when
+                `return_segments` is set True. In this case the generation outputs of each segment is added to each
+                segment.
+            kwargs (`Dict[str, Any]`, *optional*):
+                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
+                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
+                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
+
+        Return:
+            [`~utils.ModelOutput`] or `torch.LongTensor` or `Dict[str, Any]`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
+            or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor` or a dict of segments when `return_segments=True`.
+
+                If the passed input is > 30 seconds / > 3000 mel input features and `return_segments=True` then a dictionary of generated sequence ids, called `sequences` and a list of each generated segment is returned.
+
+                else if the passed input is <= 30 seconds / >= 3000 mel input features, the possible [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.GenerateEncoderDecoderOutput`],
+                    - [`~generation.GenerateBeamEncoderDecoderOutput`]
+
+                else only the generated output sequence ids are returned.
+
+        Example:
+
+        - *Longform transcription*: To transcribe or translate audios longer than 30 seconds, process the audio files without truncation and pass all mel features at once to generate.
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset, Audio
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+        >>> model.cuda()  # doctest: +IGNORE_RESULT
+
+        >>> # load audios > 30 seconds
+        >>> ds = load_dataset("distil-whisper/meanwhile", "default")["test"]
+        >>> # resample to 16kHz
+        >>> ds = ds.cast_column("audio", Audio(sampling_rate=16000))
+        >>> # take first 8 audios and retrieve array
+        >>> audio = ds[:8]["audio"]
+        >>> audio = [x["array"] for x in audio]
+
+        >>> # make sure to NOT truncate the input audio, to return the `attention_mask` and to pad to the longest audio
+        >>> inputs = processor(audio, return_tensors="pt", truncation=False, padding="longest", return_attention_mask=True, sampling_rate=16_000)
+        >>> inputs = inputs.to("cuda", torch.float32)
+
+        >>> # transcribe audio to ids
+        >>> generated_ids = model.generate(**inputs)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        >>> transcription[0]
+        " Folks, if you watch the show, you know, I spent a lot of time right over there. Patiently and astutely scrutinizing the boxwood and mahogany chest set of the day's biggest stories developing the central headline pawns, definitely maneuvering an oso topical night to F6, fainting a classic Sicilian, nade door variation on the news, all the while seeing eight moves deep and patiently marshalling the latest press releases into a fisher's shows in Lip Nitsky attack that culminates in the elegant lethal slow-played, all-passant checkmate that is my nightly monologue. But sometimes, sometimes, folks, I. CHEERING AND APPLAUSE Sometimes I startle away, cubside down in the monkey bars of a condemned playground on a super fun site. Get all hept up on goofballs. Rummage that were discarded tag bag of defective toys. Yank out a fist bowl of disembodied doll limbs, toss them on a stained kid's place mat from a defunct dennies. set up a table inside a rusty cargo container down by the Wharf and challenged toothless drifters to the godless bughouse blitz of tournament that is my segment. Meanwhile."
+        ```
+
+        - *Shortform transcription*: If passed mel input features are < 30 seconds, the whole audio will be transcribed with a single call to generate.
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(inputs=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```
+
+        """
+        # 0. deprecate old inputs
+        if "inputs" in kwargs:
+            input_features = kwargs.pop("inputs")
+            warnings.warn(
+                "The input name `inputs` is deprecated. Please make sure to use `input_features` instead.",
+                FutureWarning,
+            )
+
+        # 1. prepare generation config
+        generation_config, kwargs = self._prepare_generation_config(generation_config, **kwargs)
+
+        # 2. set global generate variables
+        input_stride = self.model.encoder.conv1.stride[0] * self.model.encoder.conv2.stride[0]
+        num_segment_frames = input_stride * self.config.max_source_positions
+        batch_size, total_input_frames = self._retrieve_total_input_frames(
+            input_features=input_features, input_stride=input_stride, kwargs=kwargs
+        )
+        is_shortform = total_input_frames <= num_segment_frames
+
+        # 3. Make sure generation config is correctly set
+        # Make sure the generation config is correctly set depending on whether timestamps are to be returned or not
+        return_dict_in_generate = self._set_return_outputs(
+            return_dict_in_generate=return_dict_in_generate,
+            return_token_timestamps=return_token_timestamps,
+            logprob_threshold=logprob_threshold,
+            generation_config=generation_config,
+        )
+        timestamp_begin = self._set_return_timestamps(
+            return_timestamps=return_timestamps, is_shortform=is_shortform, generation_config=generation_config
+        )
+        self._set_language_and_task(
+            language=language, task=task, is_multilingual=is_multilingual, generation_config=generation_config
+        )
+        self._set_num_frames(
+            return_token_timestamps=return_token_timestamps, generation_config=generation_config, kwargs=kwargs
+        )
+        self._set_thresholds_and_condition(
+            generation_config=generation_config,
+            logprob_threshold=logprob_threshold,
+            compression_ratio_threshold=compression_ratio_threshold,
+            no_speech_threshold=no_speech_threshold,
+            condition_on_prev_tokens=condition_on_prev_tokens,
+        )
+        self._set_prompt_condition_type(
+            generation_config=generation_config,
+            prompt_condition_type=prompt_condition_type,
+        )
+
+        # pass self.config for backward compatibility
+        init_tokens = self._retrieve_init_tokens(
+            input_features,
+            batch_size=batch_size,
+            generation_config=generation_config,
+            config=self.config,
+            num_segment_frames=num_segment_frames,
+            kwargs=kwargs,
+        )
+        # passing `decoder_input_ids` is deprecated - the only exception is for assisted generation
+        # where the input ids are handled explicitly by the generate method
+        self._check_decoder_input_ids(kwargs=kwargs)
+
+        # 3. Retrieve logits processors
+        device = kwargs["encoder_outputs"][0].device if "encoder_outputs" in kwargs else input_features.device
+        begin_index = init_tokens.shape[1]
+        logits_processor = self._retrieve_logit_processors(
+            generation_config=generation_config,
+            logits_processor=logits_processor,
+            begin_index=begin_index,  # begin index is index of first generated decoder token
+            num_beams=kwargs.get("num_beams", 1),
+            device=device,
+        )
+
+        # 4 Set and retrieve global generation variables
+        self._set_condition_on_prev_tokens(
+            condition_on_prev_tokens=condition_on_prev_tokens, generation_config=generation_config
+        )
+
+        temperatures = [temperature] if not isinstance(temperature, (list, tuple)) else temperature
+        temperature = temperatures[0]
+
+        max_frames, seek = self._retrieve_max_frames_and_seek(
+            batch_size=batch_size,
+            attention_mask=attention_mask,
+            total_input_frames=total_input_frames,
+            is_shortform=is_shortform,
+        )
+
+        # 5 Prepare running variables, list for generation
+        num_return_sequences = generation_config.num_return_sequences
+        (
+            batch_idx_map,
+            cur_bsz,
+            input_features,
+            seek,
+            max_frames,
+            init_tokens,
+            do_condition_on_prev_tokens,
+        ) = self._expand_variables_for_generation(
+            input_features=input_features,
+            seek=seek,
+            max_frames=max_frames,
+            init_tokens=init_tokens,
+            batch_size=batch_size,
+            condition_on_prev_tokens=condition_on_prev_tokens,
+            generation_config=generation_config,
+        )
+
+        current_segments = self._prepare_segments(
+            prompt_ids=prompt_ids,
+            batch_size=cur_bsz,
+            generation_config=generation_config,
+        )
+
+        # 6 Transcribe audio until we reach the end of all input audios
+        while (seek < max_frames).any():
+            # 6.1 NOTE: When in longform transcription mode and batch size > 1 we need to dynamically reduce the batch size during the loop
+            # in case one audio finished earlier than another one. Thus, we need to keep a table of "previous-index-2-current-index" in order
+            # to know which original audio is being decoded
+            # Set updated index map, duration of previously decoded chunks and number of max frames of current decoding chunk
+            input_features, cur_bsz, batch_idx_map = self._maybe_reduce_batch(
+                input_features=input_features,
+                seek=seek,
+                max_frames=max_frames,
+                cur_bsz=cur_bsz,
+                batch_idx_map=batch_idx_map,
+            )
+            time_offset = (
+                seek.to(torch.float32 if device.type == "mps" else torch.float64) * time_precision / input_stride
+            )
+            seek_num_frames = (max_frames - seek).clamp(max=num_segment_frames)
+
+            # 6.2 cut out next 30s segment from input features
+            segment_input = self._get_input_segment(
+                input_features=input_features,
+                seek=seek,
+                seek_num_frames=seek_num_frames,
+                num_segment_frames=num_segment_frames,
+                cur_bsz=cur_bsz,
+                batch_idx_map=batch_idx_map,
+            )
+
+            # 6.3 prepare decoder input ids
+            suppress_tokens = _get_attr_from_logit_processors(
+                logits_processor, SuppressTokensLogitsProcessor, "suppress_tokens"
+            )
+
+            decoder_input_ids, kwargs = self._prepare_decoder_input_ids(
+                cur_bsz=cur_bsz,
+                init_tokens=init_tokens,
+                current_segments=current_segments,
+                batch_idx_map=batch_idx_map,
+                do_condition_on_prev_tokens=do_condition_on_prev_tokens,
+                prompt_ids=prompt_ids,
+                generation_config=generation_config,
+                config=self.config,
+                device=init_tokens.device,
+                suppress_tokens=suppress_tokens,
+                timestamp_begin=timestamp_begin,
+                kwargs=kwargs,
+            )
+
+            # 6.4 set max new tokens or max length
+            self._set_max_new_tokens_and_length(
+                config=self.config,
+                decoder_input_ids=decoder_input_ids,
+                generation_config=generation_config,
+            )
+
+            # 6.5 Set current `begin_index` for all logit processors
+            if logits_processor is not None:
+                for proc in logits_processor:
+                    if hasattr(proc, "set_begin_index"):
+                        proc.set_begin_index(decoder_input_ids.shape[-1])
+
+            # 6.6 Run generate with fallback
+            (
+                seek_sequences,
+                seek_outputs,
+                should_skip,
+                do_condition_on_prev_tokens,
+                model_output_type,
+            ) = self.generate_with_fallback(
+                segment_input=segment_input,
+                decoder_input_ids=decoder_input_ids,
+                cur_bsz=cur_bsz,
+                batch_idx_map=batch_idx_map,
+                seek=seek,
+                num_segment_frames=num_segment_frames,
+                max_frames=max_frames,
+                temperatures=temperatures,
+                generation_config=generation_config,
+                logits_processor=logits_processor,
+                stopping_criteria=stopping_criteria,
+                prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+                synced_gpus=synced_gpus,
+                return_token_timestamps=return_token_timestamps,
+                do_condition_on_prev_tokens=do_condition_on_prev_tokens,
+                is_shortform=is_shortform,
+                batch_size=batch_size,
+                attention_mask=attention_mask,
+                kwargs=kwargs,
+            )
+
+            # 6.7 In every generated sequence, split by timestamp tokens and extract segments
+            for i, seek_sequence in enumerate(seek_sequences):
+                prev_i = batch_idx_map[i]
+
+                if should_skip[i]:
+                    seek[prev_i] += seek_num_frames[prev_i]
+                    continue
+
+                segments, segment_offset = self._retrieve_segment(
+                    seek_sequence=seek_sequence,
+                    seek_outputs=seek_outputs,
+                    time_offset=time_offset,
+                    timestamp_begin=timestamp_begin,
+                    seek_num_frames=seek_num_frames,
+                    time_precision=time_precision,
+                    time_precision_features=time_precision_features,
+                    input_stride=input_stride,
+                    prev_idx=prev_i,
+                    idx=i,
+                    return_token_timestamps=return_token_timestamps,
+                )
+
+                current_segments[prev_i] += segments
+
+                if is_shortform:
+                    seek[prev_i] += max_frames[i]
+                else:
+                    seek[prev_i] += segment_offset
+
+        # 7. Once all segments are added to the list of all segments, called `current_segments`, we extract the predicted
+        # output tokens from the list of dicts. If we use batch size > 1, we make sure to pad the output
+        final_segments = (
+            [x[1:] for x in current_segments]
+            if (prompt_ids is not None and generation_config.prompt_condition_type == "first-segment")
+            else current_segments
+        )
+
+        sequences = _pad_to_max_length(
+            final_segments, generation_config.pad_token_id, device=self.device, padding_side="right"
+        )
+
+        # 8. If we return all segments, the predicted output sequences are put under `"sequences"`.
+        if return_segments:
+            return {"sequences": sequences, "segments": final_segments}
+
+        if is_shortform:
+            # add eos token:
+            if generation_config.max_new_tokens is None and generation_config.max_length is None:
+                eos_tokens = torch.full((sequences.shape[0], 1), generation_config.eos_token_id)
+                sequences = torch.cat([sequences, eos_tokens], dim=-1)
+
+            if return_token_timestamps:
+                outputs = {}
+                outputs["sequences"] = sequences
+                outputs["token_timestamps"] = torch.stack([d["token_timestamps"] for d in seek_outputs], dim=0)
+            else:
+                outputs = sequences
+
+            if return_dict_in_generate and generation_config.return_dict_in_generate:
+                dict_outputs = self._stack_split_outputs(seek_outputs, model_output_type, sequences.device, kwargs)
+
+                if num_return_sequences > 1:
+                    if hasattr(dict_outputs, "encoder_attentions") and dict_outputs.encoder_attentions is not None:
+                        dict_outputs.encoder_attentions = tuple(
+                            dict_outputs.encoder_attentions[i][::num_return_sequences]
+                            for i in range(len(dict_outputs.encoder_attentions))
+                        )
+                    if (
+                        hasattr(dict_outputs, "encoder_hidden_states")
+                        and dict_outputs.encoder_hidden_states is not None
+                    ):
+                        dict_outputs.encoder_hidden_states = tuple(
+                            dict_outputs.encoder_hidden_states[i][::num_return_sequences]
+                            for i in range(len(dict_outputs.encoder_hidden_states))
+                        )
+                if return_token_timestamps:
+                    dict_outputs["token_timestamps"] = outputs["token_timestamps"]
+                return dict_outputs
+
+            return outputs
+
+        return sequences
+
+    def generate_with_fallback(
+        self,
+        segment_input,
+        decoder_input_ids,
+        cur_bsz,
+        batch_idx_map,
+        seek,
+        num_segment_frames,
+        max_frames,
+        temperatures,
+        generation_config,
+        logits_processor,
+        stopping_criteria,
+        prefix_allowed_tokens_fn,
+        synced_gpus,
+        return_token_timestamps,
+        do_condition_on_prev_tokens,
+        is_shortform,
+        batch_size,
+        attention_mask,
+        kwargs,
+    ):
+        kwargs = copy.copy(kwargs)
+
+        # 6.6 Batch generate current chunk
+        seek_sequence_list = [None for _ in range(cur_bsz)]
+        seek_outputs_list = [None for _ in range(cur_bsz)]
+        needs_fallback = [False for _ in range(cur_bsz)]
+        should_skip = [False for _ in range(cur_bsz)]
+        fallback_index_map = list(range(cur_bsz))
+        if generation_config.no_speech_threshold is not None:
+            self._setup_no_speech_detection(logits_processor, segment_input, decoder_input_ids, kwargs)
+
+        for fallback_idx, temperature in enumerate(temperatures):
+            generation_config.do_sample = temperature is not None and temperature > 0.0
+            generation_config.temperature = temperature if generation_config.do_sample else 1.0
+            if generation_config.do_sample:
+                generation_config.num_beams = 1
+
+            generate_kwargs = copy.copy(kwargs)
+            for key in ["do_sample", "temperature", "num_beams"]:
+                if key in generate_kwargs:
+                    del generate_kwargs[key]
+
+            cur_bsz = decoder_input_ids.shape[0]
+            if generation_config.cache_implementation == "static" and cur_bsz < batch_size:
+                segment_input = F.pad(segment_input, (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0)
+                decoder_input_ids = F.pad(
+                    decoder_input_ids, (0, 0, 0, batch_size - cur_bsz), value=generation_config.pad_token_id
+                )
+                if generate_kwargs.get("decoder_attention_mask") is not None:
+                    generate_kwargs["decoder_attention_mask"] = F.pad(
+                        generate_kwargs["decoder_attention_mask"], (0, 0, 0, batch_size - cur_bsz), value=True
+                    )
+                if generate_kwargs.get("encoder_outputs") is not None:
+                    generate_kwargs["encoder_outputs"] = F.pad(
+                        generate_kwargs["encoder_outputs"], (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0
+                    )
+
+            seek_outputs = super().generate(
+                segment_input,
+                generation_config=generation_config,
+                logits_processor=logits_processor,
+                stopping_criteria=stopping_criteria,
+                prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+                synced_gpus=synced_gpus,
+                decoder_input_ids=decoder_input_ids,
+                attention_mask=attention_mask,
+                **generate_kwargs,
+            )
+
+            model_output_type = type(seek_outputs)
+
+            # post-process sequence tokens and outputs to be in list form
+            seek_sequences, seek_outputs = self._postprocess_outputs(
+                seek_outputs=seek_outputs,
+                decoder_input_ids=decoder_input_ids,
+                return_token_timestamps=return_token_timestamps,
+                generation_config=generation_config,
+                is_shortform=is_shortform,
+            )
+
+            if cur_bsz < batch_size:
+                seek_sequences = seek_sequences[:cur_bsz]
+                seek_outputs = seek_outputs[:cur_bsz]
+
+            # 6.7 Extract cut sequences from every sequence and check if fallback should be applied
+            # Loop over each decoded audio individually as each decoding can be of a different length
+            new_fallback_index_map = []
+            new_segment_input = []
+            new_decoder_input_ids = []
+            new_decoder_attention_mask = []
+
+            for i, seek_sequence in enumerate(seek_sequences):
+                # make sure we cut a predicted EOS token if we are not finished with the generation yet
+                prev_i = batch_idx_map[fallback_index_map[i]]
+                is_not_final = (seek[prev_i] + num_segment_frames) < max_frames[prev_i]
+
+                # remove eos token id
+                if is_not_final and seek_sequence[-1] == generation_config.eos_token_id:
+                    seek_sequence = seek_sequence[:-1]
+                    if return_token_timestamps and not is_shortform:
+                        seek_outputs[i]["token_timestamps"] = seek_outputs[i]["token_timestamps"][:-1]
+
+                # remove all padding tokens
+                if seek_sequence[-1] == generation_config.pad_token_id:
+                    num_paddings = (seek_sequence == generation_config.pad_token_id).sum()
+                    seek_sequence = seek_sequence[:-num_paddings]
+                    if return_token_timestamps and not is_shortform:
+                        seek_outputs[i]["token_timestamps"] = seek_outputs[i]["token_timestamps"][:-num_paddings]
+
+                # check which sequences in batch need fallback & which should be skipped
+                needs_fallback[i], should_skip[i] = self._need_fallback(
+                    seek_sequence,
+                    seek_outputs,
+                    i,
+                    logits_processor,
+                    generation_config,
+                    self.config.vocab_size,
+                    temperature,
+                )
+
+                seek_sequence_list[fallback_index_map[i]] = seek_sequence
+                seek_outputs_list[fallback_index_map[i]] = seek_outputs[i]
+                is_low_temperature = temperature is None or temperature < 0.5
+                do_condition_on_prev_tokens[fallback_index_map[i]] = (
+                    generation_config.condition_on_prev_tokens and is_low_temperature
+                )
+
+                if needs_fallback[i]:
+                    new_fallback_index_map.append(fallback_index_map[i])
+                    new_segment_input.append(segment_input[i])
+                    new_decoder_input_ids.append(decoder_input_ids[i])
+                    if "decoder_attention_mask" in kwargs:
+                        new_decoder_attention_mask.append(kwargs["decoder_attention_mask"][i])
+
+            fallback_index_map = new_fallback_index_map
+
+            # if no sequence needs to be run with temperature fallback, we're finished
+            if len(fallback_index_map) == 0 or fallback_idx == len(temperatures) - 1:
+                seek_sequences = seek_sequence_list
+                seek_outputs = seek_outputs_list
+                break
+
+            # if we're still in the loop, make sure that decoder_input_ids and segment inputs are tensors
+            decoder_input_ids = torch.stack(new_decoder_input_ids)
+            segment_input = torch.stack(new_segment_input)
+            if "decoder_attention_mask" in kwargs:
+                kwargs["decoder_attention_mask"] = torch.stack(new_decoder_attention_mask)
+
+        return seek_sequences, seek_outputs, should_skip, do_condition_on_prev_tokens, model_output_type
+
+    @staticmethod
+    def _prepare_segments(prompt_ids, batch_size, generation_config):
+        if prompt_ids is not None and generation_config.prompt_condition_type == "first-segment":
+            prev_sot_token_id = getattr(generation_config, "prev_sot_token_id", None)
+            prompt_ids = prompt_ids[1:] if prompt_ids[0] == prev_sot_token_id else prompt_ids
+            current_segments = [[{"tokens": prompt_ids}] for _ in range(batch_size)]
+        else:
+            current_segments = [[] for _ in range(batch_size)]
+
+        return current_segments
+
+    def _postprocess_outputs(
+        self, seek_outputs, decoder_input_ids, return_token_timestamps, generation_config, is_shortform
+    ):
+        # remove all previously passed decoder input ids
+        start_idx = decoder_input_ids.shape[-1] if not is_shortform else torch.tensor(0)
+
+        if isinstance(seek_outputs, torch.Tensor):
+            seek_outputs = seek_outputs[:, start_idx:]
+            return seek_outputs, seek_outputs
+
+        if return_token_timestamps and hasattr(generation_config, "alignment_heads"):
+            num_frames = getattr(generation_config, "num_frames", None)
+            seek_outputs["token_timestamps"] = self._extract_token_timestamps(
+                seek_outputs,
+                generation_config.alignment_heads,
+                num_frames=num_frames,
+                num_input_ids=decoder_input_ids.shape[-1],
+            )
+            seek_outputs["token_timestamps"] = seek_outputs["token_timestamps"][:, start_idx:]
+
+        seek_outputs["sequences"] = seek_outputs["sequences"][:, start_idx:]
+
+        def split_by_batch_index(values, key, batch_idx, is_shortform, beam_indices=None):
+            if beam_indices is not None and key == "scores":
+                return [v[beam_idx].cpu() for (v, beam_idx) in zip(values, beam_indices[batch_idx][: len(values)])]
+            if key in ["scores", "encoder_attentions", "encoder_hidden_states", "logits"]:
+                return [v[batch_idx].cpu() for v in values]
+            if key in ["decoder_attentions", "decoder_hidden_states", "cross_attentions"]:
+                return tuple(tuple(w[batch_idx][None].cpu() for w in v) for v in values)
+            elif key == "past_key_values":
+                if not is_shortform:
+                    # we don't save `past_key_values` as this is too costly for longform
+                    return None
+                elif isinstance(values, EncoderDecoderCache):
+                    all_past_key_values = []
+                    for layer_idx in range(self.config.decoder_layers):
+                        layer_past_key_values = []
+                        for cache_cls in [values.self_attention_cache, values.cross_attention_cache]:
+                            for v in [cache_cls.key_cache, cache_cls.value_cache]:
+                                layer_past_key_values.append(v[layer_idx][batch_idx][None].cpu())
+                        all_past_key_values.append(tuple(layer_past_key_values))
+                    return tuple(all_past_key_values)
+                else:
+                    all_past_key_values = []
+                    for v in range(len(values)):
+                        layer_past_key_values = []
+                        for w in values[v]:
+                            if len(w) != 0:
+                                layer_past_key_values.append(w[batch_idx][None].cpu())
+                            else:
+                                layer_past_key_values.append(w)
+                        all_past_key_values.append(tuple(layer_past_key_values))
+                    return tuple(all_past_key_values)
+
+            return values[batch_idx].cpu()
+
+        sequence_tokens = seek_outputs["sequences"]
+        seek_outputs = [
+            {
+                k: split_by_batch_index(v, k, i, is_shortform, beam_indices=seek_outputs.get("beam_indices"))
+                for k, v in seek_outputs.items()
+            }
+            for i in range(sequence_tokens.shape[0])
+        ]
+
+        return sequence_tokens, seek_outputs
+
+    def _stack_split_outputs(self, seek_outputs, model_output_type, device, kwargs):
+        # Stack back seek_outputs tensors after splitting them with the split_by_batch_index method
+        outputs = {}
+        for key in seek_outputs[0].keys():
+            if key in ["sequences", "beam_indices"]:
+                outputs[key] = torch.stack([v[key] for v in seek_outputs], dim=0).to(device)
+            elif key in ["scores", "encoder_attentions", "encoder_hidden_states", "logits"]:
+                outputs[key] = tuple(
+                    torch.stack([v[key][i] for v in seek_outputs]).to(device) for i in range(len(seek_outputs[0][key]))
+                )
+            elif key == "sequences_scores":
+                outputs[key] = torch.stack([v[key] for v in seek_outputs], dim=0).to(device)
+            elif key in ["decoder_attentions", "decoder_hidden_states", "cross_attentions"]:
+                outputs[key] = tuple(
+                    tuple(
+                        torch.stack([v[key][i][j] for v in seek_outputs]).squeeze(1).to(device)
+                        for j in range(len(seek_outputs[0][key][0]))
+                    )
+                    for i in range(len(seek_outputs[0][key]))
+                )
+            elif key == "past_key_values":
+                past_key_value_type = kwargs.get("past_key_values")
+                if seek_outputs[0][key] is not None:
+                    outputs[key] = tuple(
+                        tuple(
+                            torch.stack([v[key][i][j] for v in seek_outputs]).squeeze(1).to(device)
+                            for j in range(len(seek_outputs[0][key][0]))
+                        )
+                        for i in range(len(seek_outputs[0][key]))
+                    )
+                    if past_key_value_type is not None and isinstance(past_key_value_type, EncoderDecoderCache):
+                        outputs[key] = past_key_value_type.from_legacy_cache(outputs[key])
+                else:
+                    outputs[key] = None
+
+        return model_output_type(**outputs)
+
+    def _need_fallback(
+        self,
+        seek_sequence,
+        seek_outputs,
+        index,
+        logits_processor,
+        generation_config,
+        vocab_size,
+        temperature,
+    ):
+        needs_fallback = False
+        should_skip = False
+        if generation_config.compression_ratio_threshold is not None:
+            compression_ratio = self._retrieve_compression_ratio(seek_sequence, vocab_size)
+
+            if compression_ratio > generation_config.compression_ratio_threshold:
+                needs_fallback = True
+
+        if generation_config.logprob_threshold is not None:
+            if hasattr(seek_outputs[0], "sequences_scores"):
+                logprobs = [s["sequences_scores"] for s in seek_outputs][index]
+            else:
+                scores = seek_outputs[index]["scores"]
+                logprobs = self._retrieve_avg_logprobs(
+                    scores, seek_sequence, generation_config.eos_token_id, temperature
+                )
+
+            if logprobs < generation_config.logprob_threshold:
+                needs_fallback = True
+
+        if generation_config.no_speech_threshold is not None:
+            no_speech_prob = _get_attr_from_logit_processors(
+                logits_processor, WhisperNoSpeechDetection, "no_speech_prob"
+            )
+
+            if (
+                logprobs < generation_config.logprob_threshold
+                and no_speech_prob[index] > generation_config.no_speech_threshold
+            ):
+                needs_fallback = False
+                should_skip = True
+
+        return needs_fallback, should_skip
+
+    def _expand_variables_for_generation(
+        self, input_features, seek, max_frames, init_tokens, batch_size, condition_on_prev_tokens, generation_config
+    ):
+        if generation_config.num_return_sequences is not None and generation_config.num_return_sequences > 1:
+            batch_idx_map = list(range(batch_size * generation_config.num_return_sequences))
+            cur_bsz = len(batch_idx_map)
+            do_condition_on_prev_tokens = [condition_on_prev_tokens for _ in range(len(batch_idx_map))]
+            input_features = input_features.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            seek = seek.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            max_frames = max_frames.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            init_tokens = init_tokens.repeat_interleave(generation_config.num_return_sequences, dim=0)
+            generation_config.num_return_sequences = 1
+        else:
+            cur_bsz = batch_size
+            batch_idx_map = list(range(cur_bsz))
+            do_condition_on_prev_tokens = [condition_on_prev_tokens for _ in range(cur_bsz)]
+
+        return (
+            batch_idx_map,
+            cur_bsz,
+            input_features,
+            seek,
+            max_frames,
+            init_tokens,
+            do_condition_on_prev_tokens,
+        )
+
+    @staticmethod
+    def _setup_no_speech_detection(logits_processor, segment_input, decoder_input_ids, kwargs):
+        set_inputs = _get_attr_from_logit_processors(logits_processor, WhisperNoSpeechDetection, "set_inputs")
+        extra_kwargs = {k: v for k, v in kwargs.items() if torch.is_tensor(v)}
+        set_inputs({"inputs": segment_input, "decoder_input_ids": decoder_input_ids, **extra_kwargs})
+
+    @staticmethod
+    def _retrieve_total_input_frames(input_features, input_stride, kwargs):
+        if input_features is not None:
+            return input_features.shape[0], input_features.shape[-1]
+
+        if "encoder_outputs" in kwargs:
+            encoder_outputs_shape = (
+                kwargs["encoder_outputs"][0].shape
+                if isinstance(kwargs["encoder_outputs"], BaseModelOutput)
+                else kwargs["encoder_outputs"].shape
+            )
+            return encoder_outputs_shape[0], encoder_outputs_shape[1] * input_stride
+
+        raise ValueError("Make sure to provide either `input_features` or `encoder_outputs` to `generate`.")
+
+    @staticmethod
+    def _maybe_warn_unused_inputs(
+        condition_on_prev_tokens,
+        temperature,
+        compression_ratio_threshold,
+        logprob_threshold,
+        no_speech_threshold,
+        total_input_frames,
+    ):
+        warning_prefix = (
+            f"Audio input consists of only {total_input_frames}. "
+            "Short-form transcription is activated."
+            "{}, but will be ignored."
+        )
+        if condition_on_prev_tokens is not None:
+            logger.warning(warning_prefix.format(f"condition_on_prev_tokens is set to {condition_on_prev_tokens}"))
+
+        if compression_ratio_threshold is not None:
+            logger.warning(
+                warning_prefix.format(f"compression_ratio_threshold is set to {compression_ratio_threshold}")
+            )
+
+        if logprob_threshold is not None:
+            logger.warning(warning_prefix.format(f"logprob_threshold is set to {logprob_threshold}"))
+
+        if no_speech_threshold is not None:
+            logger.warning(warning_prefix.format(f"no_speech_threshold is set to {no_speech_threshold}"))
+
+        # when passing temperature as a list it cannot just be ignored => throw error in this case
+        if isinstance(temperature, (list, tuple)):
+            raise ValueError(
+                f"Audio input consists of only {total_input_frames}. Short-form transcription is activated."
+                f"temperature cannot be set to {temperature} which can only be used for temperature fallback for long-form generation. Make sure to set `temperature` to a float value or `None` for short-form generation."
+            )
+
+    @staticmethod
+    def _set_return_outputs(return_dict_in_generate, return_token_timestamps, logprob_threshold, generation_config):
+        if return_dict_in_generate is None:
+            return_dict_in_generate = generation_config.return_dict_in_generate
+        else:
+            generation_config.return_dict_in_generate = return_dict_in_generate
+
+        generation_config.return_token_timestamps = return_token_timestamps
+        if return_token_timestamps:
+            generation_config.return_dict_in_generate = True
+            generation_config.output_attentions = True
+            generation_config.output_scores = True
+
+        if logprob_threshold is not None:
+            generation_config.return_dict_in_generate = True
+            generation_config.output_scores = True
+
+        return return_dict_in_generate
+
+    def _set_return_timestamps(self, return_timestamps, is_shortform, generation_config):
+        if return_timestamps is None and hasattr(generation_config, "return_timestamps"):
+            return_timestamps = generation_config.return_timestamps
+
+        if not is_shortform:
+            if return_timestamps is False:
+                raise ValueError(
+                    "You have passed more than 3000 mel input features (> 30 seconds) which automatically enables long-form generation which "
+                    "requires the model to predict timestamp tokens. Please either pass `return_timestamps=True` or make sure to pass no more than 3000 mel input features."
+                )
+
+            logger.info("Setting `return_timestamps=True` for long-form generation.")
+            return_timestamps = True
+
+        if return_timestamps and not hasattr(generation_config, "no_timestamps_token_id"):
+            raise ValueError(
+                "You are trying to return timestamps, but the generation config is not properly set. "
+                "Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`. "
+                "For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
+            )
+
+        generation_config.return_timestamps = return_timestamps
+
+        if hasattr(generation_config, "no_timestamps_token_id"):
+            timestamp_begin = generation_config.no_timestamps_token_id + 1
+        else:
+            # BC for models missing the `no_timestamps_token_id` in the generation config when generating short-form with no timestamps
+            # We set the timestamp begin token larger than the vocab size, such that the timestamp condition is never met in the decoding loop
+            timestamp_begin = self.config.vocab_size + 1
+
+        return timestamp_begin
+
+    @staticmethod
+    def _set_language_and_task(language, task, is_multilingual, generation_config):
+        if is_multilingual is not None:
+            if not hasattr(generation_config, "is_multilingual"):
+                raise ValueError(
+                    "The generation config is outdated and is thus not compatible with the `is_multilingual` argument "
+                    "to `generate`. Please update the generation config as per the instructions "
+                    "https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
+                )
+            generation_config.is_multilingual = is_multilingual
+
+        if hasattr(generation_config, "is_multilingual") and not generation_config.is_multilingual:
+            if task is not None or language is not None:
+                raise ValueError(
+                    "Cannot specify `task` or `language` for an English-only model. If the model is intended to be "
+                    "multilingual, pass `is_multilingual=True` to generate, or update the generation config."
+                )
+
+        if language is not None:
+            if not hasattr(generation_config, "lang_to_id"):
+                raise ValueError(
+                    "The generation config is outdated and is thus not compatible with the `language` argument "
+                    "to `generate`. Either set the language using the `forced_decoder_ids` in the model config, "
+                    "or update the generation config as per the instructions https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
+                )
+            generation_config.language = language
+
+        if task is not None:
+            if not hasattr(generation_config, "task_to_id"):
+                raise ValueError(
+                    "The generation config is outdated and is thus not compatible with the `task` argument "
+                    "to `generate`. Either set the task using the `forced_decoder_ids` in the model config, "
+                    "or update the generation config as per the instructions https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
+                )
+            generation_config.task = task
+
+    def _retrieve_init_tokens(self, input_features, batch_size, generation_config, config, num_segment_frames, kwargs):
+        def replace_or_add(lst: List[int], num: int, itr: Iterator[int]):
+            """short function to replace num with a itr in lst"""
+            found = any(i in lst for i in itr)
+            if found:
+                lst = [num if i in itr else i for i in lst]
+            else:
+                lst.append(num)
+            return lst
+
+        def language_to_id(language: str) -> int:
+            language = language.lower()
+            if language in generation_config.lang_to_id.keys():
+                language_token = language
+            elif language in TO_LANGUAGE_CODE.keys():
+                language_token = f"<|{TO_LANGUAGE_CODE[language]}|>"
+            elif language in TO_LANGUAGE_CODE.values():
+                language_token = f"<|{language}|>"
+            else:
+                is_language_code = len(language) == 2
+                raise ValueError(
+                    f"Unsupported language: {language}. Language should be one of:"
+                    f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
+                )
+            if language_token not in generation_config.lang_to_id:
+                raise ValueError(
+                    f"{language_token} is not supported by this specific model as it is not in the `generation_config.lang_to_id`."
+                    "(You should just add it to the generation config)"
+                )
+
+            return generation_config.lang_to_id[language_token]
+
+        task = getattr(generation_config, "task", None)
+        language = getattr(generation_config, "language", None)
+
+        forced_decoder_ids = generation_config.forced_decoder_ids
+        if forced_decoder_ids is not None:
+            if language is None and task is None and forced_decoder_ids[0][1] is None:
+                logger.warning_once(
+                    "Due to a bug fix in https://github.com/huggingface/transformers/pull/28687 transcription using a multilingual Whisper will default to language detection followed by transcription instead of translation to English."
+                    "This might be a breaking change for your use case. If you want to instead always translate your audio to English, make sure to pass `language='en'`."
+                )
+        elif hasattr(config, "forced_decoder_ids") and config.forced_decoder_ids is not None:
+            forced_decoder_ids = config.forced_decoder_ids
+
+        if forced_decoder_ids is not None and task is not None:
+            logger.warning_once(
+                f"You have passed task={task}, but also have set `forced_decoder_ids` to {forced_decoder_ids} which creates a conflict. `forced_decoder_ids` will be ignored in favor of task={task}."
+            )
+            forced_decoder_ids = None
+        elif forced_decoder_ids is not None and language is not None:
+            logger.warning_once(
+                f"You have passed language={language}, but also have set `forced_decoder_ids` to {forced_decoder_ids} which creates a conflict. `forced_decoder_ids` will be ignored in favor of language={language}."
+            )
+            forced_decoder_ids = None
+
+        init_tokens = [generation_config.decoder_start_token_id]
+        if forced_decoder_ids is not None and forced_decoder_ids[0][0] == 1:
+            i = 1
+            while len(forced_decoder_ids) > 0 and forced_decoder_ids[0][0] == i:
+                init_tokens += [forced_decoder_ids[0][1]]
+                forced_decoder_ids = forced_decoder_ids[1:]
+                i += 1
+
+            if len(forced_decoder_ids) > 0:
+                raise ValueError(
+                    f"You are using token ids in `forced_decoder_ids` that do not seem to correctly follow the prompt pattern of Whisper. Make sure that {forced_decoder_ids} has an entry for all indices >= 1 and < {forced_decoder_ids[0][0]}.",
+                )
+
+        # from v4.39 the forced decoder ids are always None in favour of decoder input ids
+        generation_config.forced_decoder_ids = None
+
+        is_lang_id_undefined = len(init_tokens) <= 1 or (len(init_tokens) > 1 and init_tokens[1] is None)
+
+        # Make sure language is a list of strings of the correct length
+        if isinstance(language, (list, tuple)):
+            if any(l is None for l in language):
+                raise TypeError(
+                    "Expected `language` to be `None`, a single string (e.g. `'en'`), or a list of strings with length equal to the batch size (e.g. `('en', 'fr')` for a batch size of 2). Got a list containing `None`."
+                )
+            if len(language) != batch_size:
+                raise ValueError(
+                    "When passing a list of languages, the length of the list must match the batch size. "
+                    f"Expected length of {batch_size}, but got {len(language)} languages."
+                )
+            languages = language
+        elif language is None:
+            # Language will be detected for each item in batch
+            languages = [None] * batch_size
+        else:
+            languages = [language]  # Use a length-1 list now, broadcast later
+
+        # Separate init_tokens for each language
+        init_tokens = [copy.copy(init_tokens) for _ in languages]
+
+        # Update init_tokens with languages
+        lang_ids = None
+        if language is not None:
+            lang_ids = [language_to_id(l) for l in languages]
+        elif hasattr(generation_config, "lang_to_id") and is_lang_id_undefined:
+            # language is not defined or intentially set to `None` to trigger language detection
+            lang_ids = self.detect_language(
+                input_features=input_features,
+                encoder_outputs=kwargs.get("encoder_outputs", None),
+                generation_config=generation_config,
+                num_segment_frames=num_segment_frames,
+            ).tolist()
+        if lang_ids is not None:
+            # append or replace lang_ids to init_tokens
+            for i in range(len(init_tokens)):
+                if len(init_tokens[i]) > 1:
+                    init_tokens[i][1] = lang_ids[i]
+                else:
+                    init_tokens[i].append(lang_ids[i])
+        del languages
+
+        # Update init_tokens with task
+        for i in range(len(init_tokens)):
+            if task is not None:
+                if task in TASK_IDS:
+                    init_tokens[i].append(generation_config.task_to_id[generation_config.task])
+                    task_id = generation_config.task_to_id[generation_config.task]
+
+                    # if task is defined it'll overwrite task ids that might have already been defined via the generation_config
+                    replace_or_add(init_tokens[i], task_id, generation_config.task_to_id.values())
+                else:
+                    raise ValueError(f"The `{task}`task is not supported. The task should be one of `{TASK_IDS}`")
+            elif language is not None and hasattr(generation_config, "task_to_id"):
+                # if language is defined, but no task id is in `init_tokens`, default to transcribe
+                if not any(ti in init_tokens[i] for ti in generation_config.task_to_id.values()):
+                    init_tokens[i].append(generation_config.task_to_id["transcribe"])
+
+            if (
+                not generation_config.return_timestamps
+                and hasattr(generation_config, "no_timestamps_token_id")
+                and init_tokens[i][-1] != generation_config.no_timestamps_token_id
+            ):
+                init_tokens[i].append(generation_config.no_timestamps_token_id)
+            elif (
+                generation_config.return_timestamps and init_tokens[i][-1] == generation_config.no_timestamps_token_id
+            ):
+                logger.info(
+                    "<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `'True'`."
+                )
+                init_tokens[i] = init_tokens[i][:-1]
+
+            # let's make sure we don't pass `None` tokens as prompt tokens
+            init_tokens[i] = [t for t in init_tokens[i] if t is not None]
+
+        return torch.as_tensor(init_tokens, dtype=torch.long, device=self.device).expand(batch_size, -1)
+
+    def detect_language(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        encoder_outputs: Optional[Union[torch.FloatTensor, BaseModelOutput]] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        num_segment_frames: int = 3000,
+    ) -> torch.Tensor:
+        """
+        Detects language from log-mel input features or encoder_outputs
+
+        Parameters:
+            input_features (`torch.Tensor` of shape `(batch_size, feature_size, sequence_length)`, *optional*):
+                Float values of log-mel features extracted from the raw speech waveform. The raw speech waveform can be obtained by
+                loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+                the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+                [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+                tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] for details.
+            encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+                Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+                `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+                hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            num_segment_frames (`int`, *optional*, defaults to 3000):
+                The number of log-mel frames the model expects
+
+        Return:
+            A `torch.LongTensor` representing the detected language ids.
+        """
+        if input_features is None and encoder_outputs is None:
+            raise ValueError("You have to specify either `input_features` or `encoder_outputs`")
+        elif input_features is not None and encoder_outputs is not None:
+            raise ValueError("Make sure to specificy only one of `input_features` or `encoder_outputs` - not both!")
+        elif input_features is not None:
+            inputs = {"input_features": input_features[:, :, :num_segment_frames]}
+            batch_size = input_features.shape[0]
+        elif encoder_outputs is not None:
+            inputs = {"encoder_outputs": encoder_outputs}
+            batch_size = (
+                encoder_outputs[0].shape[0] if isinstance(encoder_outputs, BaseModelOutput) else encoder_outputs[0]
+            )
+
+        generation_config = generation_config or self.generation_config
+        decoder_input_ids = (
+            torch.ones((batch_size, 1), device=self.device, dtype=torch.long)
+            * generation_config.decoder_start_token_id
+        )
+
+        with torch.no_grad():
+            logits = self(**inputs, decoder_input_ids=decoder_input_ids).logits[:, -1]
+
+        non_lang_mask = torch.ones_like(logits[0], dtype=torch.bool)
+        non_lang_mask[list(generation_config.lang_to_id.values())] = False
+
+        logits[:, non_lang_mask] = -np.inf
+
+        lang_ids = logits.argmax(-1)
+
+        return lang_ids
+
+    @staticmethod
+    def _check_decoder_input_ids(kwargs):
+        decoder_input_ids = kwargs.get("decoder_input_ids", None)
+        assistant_model = kwargs.get("assistant_model", None)
+        if decoder_input_ids is not None and assistant_model is not None:
+            raise ValueError(
+                "Passing `decoder_input_ids` is deprecated. Consider passing `prompt_ids` instead.",
+            )
+
+    @staticmethod
+    def _set_num_frames(return_token_timestamps, generation_config, kwargs):
+        if return_token_timestamps:
+            if getattr(generation_config, "task", None) == "translate":
+                logger.warning("Token-level timestamps may not be reliable for task 'translate'.")
+            if not hasattr(generation_config, "alignment_heads"):
+                raise ValueError(
+                    "Model generation config has no `alignment_heads`, token-level timestamps not available. "
+                    "See https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a on how to add this property to the generation config."
+                )
+            generation_config.num_frames = kwargs.pop("num_frames", None)
+
+    @staticmethod
+    def _set_thresholds_and_condition(
+        generation_config,
+        logprob_threshold,
+        compression_ratio_threshold,
+        no_speech_threshold,
+        condition_on_prev_tokens,
+    ):
+        generation_config.logprob_threshold = (
+            logprob_threshold
+            if logprob_threshold is not None
+            else getattr(generation_config, "logprob_threshold", None)
+        )
+        generation_config.compression_ratio_threshold = (
+            compression_ratio_threshold
+            if compression_ratio_threshold is not None
+            else getattr(generation_config, "compression_ratio_threshold", None)
+        )
+        generation_config.no_speech_threshold = (
+            no_speech_threshold
+            if no_speech_threshold is not None
+            else getattr(generation_config, "no_speech_threshold", None)
+        )
+        generation_config.condition_on_prev_tokens = (
+            condition_on_prev_tokens
+            if condition_on_prev_tokens is not None
+            else getattr(generation_config, "condition_on_prev_tokens", None)
+        )
+
+    @staticmethod
+    def _set_prompt_condition_type(generation_config, prompt_condition_type):
+        allowed_cond_types = ["first-segment", "all-segments"]
+
+        # default to "first-segment"
+        prompt_condition_type = prompt_condition_type or allowed_cond_types[0]
+
+        if prompt_condition_type not in allowed_cond_types:
+            raise ValueError(
+                f"`prompt_condition_type={prompt_condition_type} does not exist. Make sure to set `prompt_condition_type` to one of {', '.join(allowed_cond_types)}"
+            )
+
+        if generation_config.condition_on_prev_tokens is not True and prompt_condition_type == "all-segments":
+            raise ValueError(
+                "Make sure to set `condition_on_prev_tokens=True` when setting `prompt_condition_type='all-segments'`."
+            )
+
+        generation_config.prompt_condition_type = prompt_condition_type
+
+    @staticmethod
+    def _set_condition_on_prev_tokens(condition_on_prev_tokens, generation_config):
+        condition_on_prev_tokens = (
+            condition_on_prev_tokens
+            if condition_on_prev_tokens is not None
+            else getattr(generation_config, "condition_on_prev_tokens", False)
+        )
+        generation_config.condition_on_prev_tokens = condition_on_prev_tokens
+
+    @staticmethod
+    def _retrieve_max_frames_and_seek(batch_size, attention_mask, total_input_frames, is_shortform):
+        if batch_size > 1 and not is_shortform and attention_mask is None:
+            raise ValueError(
+                "When doing batched long-form audio transcription, make sure to pass an `attention_mask`. You can retrieve the `attention_mask` by doing `processor(audio, ..., return_attention_mask=True)` "
+            )
+        elif batch_size > 1 and not is_shortform:
+            max_frames = attention_mask.sum(-1).cpu().to(torch.long)
+            seek = torch.zeros((batch_size,), dtype=torch.long)
+        else:
+            max_frames = torch.ones((batch_size,), dtype=torch.long) * total_input_frames
+            seek = torch.zeros((batch_size,), dtype=torch.long)
+
+        return max_frames, seek
+
+    def _retrieve_logit_processors(self, generation_config, logits_processor, begin_index, num_beams, device):
+        if generation_config.return_timestamps is True:
+            timestamp_processor = WhisperTimeStampLogitsProcessor(generation_config, begin_index=begin_index)
+            logits_processor = (
+                [timestamp_processor] if logits_processor is None else [timestamp_processor] + logits_processor
+            )
+
+        if generation_config.suppress_tokens is not None:
+            suppress_tokens_processor = SuppressTokensLogitsProcessor(generation_config.suppress_tokens, device=device)
+            logits_processor = (
+                [suppress_tokens_processor]
+                if logits_processor is None
+                else [suppress_tokens_processor] + logits_processor
+            )
+            generation_config.suppress_tokens = None
+
+        if generation_config.begin_suppress_tokens is not None:
+            begin_suppress_processor = SuppressTokensAtBeginLogitsProcessor(
+                generation_config.begin_suppress_tokens, begin_index=begin_index, device=device
+            )
+            logits_processor = (
+                [begin_suppress_processor]
+                if logits_processor is None
+                else [begin_suppress_processor] + logits_processor
+            )
+            generation_config.begin_suppress_tokens = None
+
+        if generation_config.no_speech_threshold is not None:
+            no_speech_detector = WhisperNoSpeechDetection(
+                no_speech_token=generation_config.no_timestamps_token_id - 1,
+                begin_index=begin_index,
+                scores_is_logprobs=num_beams > 1,
+            )
+            logits_processor = (
+                [no_speech_detector] if logits_processor is None else [no_speech_detector] + logits_processor
+            )
+            no_speech_detector.set_model(self)
+
+        return logits_processor
+
+    @staticmethod
+    def _maybe_reduce_batch(input_features, seek, max_frames, cur_bsz, batch_idx_map):
+        prev_bsz = cur_bsz
+        new_batch_idx_map = []
+        for i in range(prev_bsz):
+            prev_i = batch_idx_map[i]
+            if seek[prev_i] >= max_frames[prev_i]:
+                cut_index = i + (cur_bsz - prev_bsz)
+                cur_bsz -= 1
+                input_features = torch.cat([input_features[:cut_index], input_features[cut_index + 1 :]], dim=0)
+            else:
+                # cut out index that goes away
+                new_batch_idx_map.append(prev_i)
+
+        return input_features, cur_bsz, new_batch_idx_map
+
+    @staticmethod
+    def _get_input_segment(input_features, seek, seek_num_frames, num_segment_frames, cur_bsz, batch_idx_map):
+        if input_features is None:
+            return None
+
+        segment_input = []
+        for i in range(cur_bsz):
+            prev_i = batch_idx_map[i]
+            segment_input_slice = input_features[i : i + 1, :, seek[prev_i] : seek[prev_i] + seek_num_frames[prev_i]]
+
+            if segment_input_slice.shape[-1] < num_segment_frames:
+                # pad to 3000 if necessary
+                segment_input_slice = F.pad(
+                    segment_input_slice, pad=(0, num_segment_frames - segment_input_slice.shape[-1])
+                )
+
+            segment_input.append(segment_input_slice)
+
+        segment_input = torch.cat(segment_input, dim=0)
+
+        return segment_input
+
+    @staticmethod
+    def _prepare_decoder_input_ids(
+        cur_bsz,
+        init_tokens,
+        current_segments,
+        batch_idx_map,
+        do_condition_on_prev_tokens,
+        prompt_ids,
+        generation_config,
+        config,
+        device,
+        suppress_tokens,
+        timestamp_begin,
+        kwargs,
+    ):
+        if "decoder_input_ids" in kwargs:
+            decoder_input_ids = kwargs.pop("decoder_input_ids")
+
+            return decoder_input_ids, kwargs
+
+        cut_off_length = config.max_target_positions // 2 - 1
+
+        decoder_input_ids = init_tokens[batch_idx_map]
+
+        prev_start_of_text = getattr(generation_config, "prev_sot_token_id", None)
+        if prev_start_of_text is None:
+            prev_start_of_text = suppress_tokens[-2] if suppress_tokens is not None else None
+
+        if any(do_condition_on_prev_tokens) and len(current_segments[0]) > 0:
+            # according to https://github.com/openai/whisper/blob/e58f28804528831904c3b6f2c0e473f346223433/whisper/decoding.py#L609
+            active_segments = [current_segments[i] if do_condition_on_prev_tokens[i] else None for i in batch_idx_map]
+
+            for segments in active_segments:
+                for seg in segments:
+                    if len(seg["tokens"]) > 2 and seg["tokens"][-2] >= timestamp_begin:
+                        # the segment finishes with two timestamp tokens
+                        # we need to ignore the last timestamp token
+                        # see https://github.com/huggingface/transformers/pull/34537
+                        seg["tokens"] = seg["tokens"][:-1]
+
+            if prompt_ids is not None and generation_config.prompt_condition_type == "all-segments":
+                prev_ids = prompt_ids
+            else:
+                one_tensor = torch.ones((cur_bsz, 1), device=device, dtype=torch.long)
+                prev_ids = prev_start_of_text * one_tensor[0] if prev_start_of_text is not None else None
+
+            padding = "max_length" if generation_config.cache_implementation == "static" else "longest"
+
+            prev_tokens = _pad_to_max_length(
+                active_segments,
+                generation_config.pad_token_id,
+                device=device,
+                padding_side="left",
+                padding=padding,
+                bos_token_tensor=prev_ids,
+                cut_off_length=cut_off_length,
+            )
+            decoder_input_ids = torch.cat([prev_tokens, decoder_input_ids], dim=-1)
+
+            kwargs["decoder_attention_mask"] = decoder_input_ids != generation_config.pad_token_id
+        elif prompt_ids is not None:
+            prev_tokens = prompt_ids[None].repeat(decoder_input_ids.shape[0], 1)
+            decoder_input_ids = torch.cat([prev_tokens, decoder_input_ids], dim=-1)
+            # make sure `"decoder_attention_mask"` is not passed to forward
+            kwargs.pop("decoder_attention_mask", None)
+        else:
+            # make sure `"decoder_attention_mask"` is not passed to forward
+            kwargs.pop("decoder_attention_mask", None)
+
+        return decoder_input_ids, kwargs
+
+    def _set_max_new_tokens_and_length(self, config, decoder_input_ids, generation_config):
+        max_new_tokens = generation_config.max_new_tokens if generation_config.max_new_tokens is not None else 0
+        if max_new_tokens + decoder_input_ids.shape[-1] > self.config.max_target_positions:
+            raise ValueError(
+                f"The length of `decoder_input_ids`, including special start tokens, prompt tokens, and previous tokens, is {decoder_input_ids.shape[-1]}, "
+                f" and `max_new_tokens` is {max_new_tokens}. Thus, the combined length of "
+                f"`decoder_input_ids` and `max_new_tokens` is: {max_new_tokens + decoder_input_ids.shape[-1]}. This exceeds the "
+                f"`max_target_positions` of the Whisper model: {self.config.max_target_positions}. "
+                "You should either reduce the length of your prompt, or reduce the value of `max_new_tokens`, "
+                f"so that their combined length is less than {self.config.max_target_positions}."
+            )
+
+        num_initial_tokens = min(config.max_target_positions // 2 - 1, decoder_input_ids.shape[-1] - 1)
+
+        # Make sure we don't get larger than `max_length`
+        if generation_config.max_length is not None and generation_config.max_new_tokens is None:
+            max_length = min(generation_config.max_length + num_initial_tokens, config.max_target_positions)
+            logger.info(
+                f"Increase max_length from {generation_config.max_length} to {max_length} since input is conditioned on previous segment."
+            )
+        elif (
+            generation_config.max_new_tokens is not None
+            and generation_config.max_new_tokens + decoder_input_ids.shape[-1] > config.max_target_positions
+        ):
+            max_new_tokens = config.max_target_positions - decoder_input_ids.shape[-1]
+            generation_config.max_new_tokens = max_new_tokens
+
+    @staticmethod
+    def _retrieve_compression_ratio(tokens, vocab_size):
+        """Compute byte length of zlib compressed token bytes vs. byte length of raw token bytes"""
+        length = int(math.log2(vocab_size) / 8) + 1
+        token_bytes = b"".join([t.to_bytes(length, "little") for t in tokens.tolist()])
+        compression_ratio = len(token_bytes) / len(zlib.compress(token_bytes))
+
+        return compression_ratio
+
+    @staticmethod
+    def _retrieve_avg_logprobs(scores, tokens, eos_token_id, temperature):
+        rescale_temperature = temperature if temperature > 0.0 else 1
+        scores = torch.stack(scores).to(tokens.device)
+
+        if scores.shape[0] > tokens.shape[0]:
+            scores = scores[: tokens.shape[0]]
+        else:
+            tokens = tokens[-scores.shape[0] :]
+
+        logprobs = F.log_softmax((scores * rescale_temperature).float(), dim=-1).to(scores.dtype)
+
+        # retrieve logprob of selected tokens and sum
+        sum_logprobs = sum((logprobs[i][tokens[i]] * (tokens[i] != eos_token_id)) for i in range(logprobs.shape[0]))
+        length = (tokens != eos_token_id).sum(-1) if eos_token_id is not None else tokens.shape[0]
+
+        avg_logprobs = sum_logprobs / (length + 1)
+        return avg_logprobs
+
+    @staticmethod
+    def _retrieve_segment(
+        seek_sequence,
+        seek_outputs,
+        time_offset,
+        timestamp_begin,
+        seek_num_frames,
+        time_precision,
+        time_precision_features,
+        input_stride,
+        prev_idx,
+        idx,
+        return_token_timestamps,
+    ):
+        # find the predicted "end of segment" predictions of Whisper
+        # "end of segment" predictions occur whenever Whisper predicts a timestamp token
+        timestamp_tokens: torch.Tensor = seek_sequence.ge(timestamp_begin)
+        single_timestamp_ending = timestamp_tokens[-2:].tolist() == [False, True]
+        timestamp_segment_indices = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0]
+        timestamp_segment_indices.add_(1)
+        token_timestamps = seek_outputs[idx]["token_timestamps"] if return_token_timestamps else []
+        device = seek_sequence.device
+
+        # If whisper predicted a "end of segment" via a timestep token, let's go ever each
+        # "end of segment" prediction and slice the decoding into segments accordingly
+        if len(timestamp_segment_indices) > 0:
+            # if the output contains two consecutive timestamp tokens
+            slices = timestamp_segment_indices.tolist()
+            segments = []
+            if single_timestamp_ending:
+                slices.append(len(seek_sequence))
+            else:
+                # we want to include the last timestamp token in the last segment to know it was no single ending
+                slices[-1] += 1
+
+            last_slice = 0
+            # Add each segment to list of all segments
+            for i, current_slice in enumerate(slices):
+                is_last_slice = i == len(slices) - 1
+                sliced_tokens = seek_sequence[last_slice:current_slice]
+                start_timestamp_pos = sliced_tokens[0] - timestamp_begin
+                idx_sliced_tokens = -1 if not is_last_slice or single_timestamp_ending else -2
+                end_timestamp_pos = sliced_tokens[idx_sliced_tokens] - timestamp_begin
+                segments.append(
+                    {
+                        "start": time_offset[prev_idx]
+                        + start_timestamp_pos.to(torch.float32 if device.type == "mps" else torch.float64)
+                        * time_precision,
+                        "end": time_offset[prev_idx]
+                        + end_timestamp_pos.to(torch.float32 if device.type == "mps" else torch.float64)
+                        * time_precision,
+                        "tokens": sliced_tokens,
+                        "result": seek_outputs[idx],
+                    }
+                )
+                if return_token_timestamps:
+                    segments[-1]["token_timestamps"] = (
+                        token_timestamps[last_slice:current_slice] + time_offset[prev_idx]
+                    )
+                last_slice = current_slice
+
+            if single_timestamp_ending:
+                # single timestamp at the end means no speech after the last timestamp.
+                segment_offset = seek_num_frames[prev_idx]
+            else:
+                # otherwise, ignore the unfinished segment and seek to the last timestamp
+                # here we throw away all predictions after the last predicted "end of segment"
+                # since we are cutting right in the middle of an audio
+                last_timestamp_pos = seek_sequence[last_slice - 2].item() - timestamp_begin
+                segment_offset = last_timestamp_pos * input_stride
+        else:
+            # If whisper does not predict any "end of segment" token, then
+            # the whole decoding is considered a segment and we add it to the list of segments
+            timestamps = seek_sequence[timestamp_tokens.nonzero().flatten()]
+            last_timestamp_pos = int(seek_num_frames[prev_idx] * time_precision_features / time_precision)
+            if timestamps.numel() > 0 and timestamps[-1] != timestamp_begin:
+                # no consecutive timestamps but it has a timestamp; use the last one.
+                last_timestamp_pos = (timestamps[-1] - timestamp_begin).to(
+                    torch.float32 if device.type == "mps" else torch.float64
+                )
+            segments = [
+                {
+                    "start": time_offset[prev_idx],
+                    "end": time_offset[prev_idx] + last_timestamp_pos * time_precision,
+                    "tokens": seek_sequence,
+                    "result": seek_outputs[idx],
+                }
+            ]
+            if return_token_timestamps:
+                segments[-1]["token_timestamps"] = token_timestamps + time_offset[prev_idx]
+            segment_offset = seek_num_frames[prev_idx]
+
+        return segments, segment_offset

modeling_flax_whisper (1).py

@@ -0,0 +1,1696 @@
+# coding=utf-8
+# Copyright 2022 The OpenAI Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Flax whisper model."""
+
+import math
+import random
+from functools import partial
+from typing import Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen import partitioning as nn_partitioning
+from flax.linen.attention import dot_product_attention_weights
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax import lax
+from jax.random import PRNGKey
+
+from ...generation.flax_logits_process import FlaxWhisperTimeStampLogitsProcessor
+from ...modeling_flax_outputs import (
+    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
+    FlaxSeq2SeqLMOutput,
+    FlaxSeq2SeqModelOutput,
+    FlaxSequenceClassifierOutput,
+)
+from ...modeling_flax_utils import (
+    ACT2FN,
+    FlaxPreTrainedModel,
+    append_call_sample_docstring,
+    append_replace_return_docstrings,
+    overwrite_call_docstring,
+)
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_whisper import WhisperConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+_CHECKPOINT_FOR_DOC = "openai/whisper-tiny"
+_CONFIG_FOR_DOC = "WhisperConfig"
+
+remat = nn_partitioning.remat
+
+
+def sinusoidal_embedding_init(key, shape, dtype=jnp.float_) -> jax.Array:
+    """Returns sinusoids for positional embedding"""
+    length, channels = shape
+    if channels % 2 != 0:
+        raise ValueError(
+            f"Number of channels has to be divisible by 2 for sinusoidal positional embeddings, got {channels} channels."
+        )
+    log_timescale_increment = math.log(10000) / (channels // 2 - 1)
+    inv_timescales = jnp.exp(-log_timescale_increment * jnp.arange(channels // 2))
+    scaled_time = jnp.arange(length).reshape(-1, 1) * inv_timescales.reshape(1, -1)
+    return jnp.concatenate([jnp.sin(scaled_time), jnp.cos(scaled_time)], axis=1).astype(dtype)
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its models (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.) This model is also a Flax Linen
+    [flax.nn.Module](https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html) subclass. Use it as a
+    regular Flax Module and refer to the Flax documentation for all matter related to general usage and behavior.
+    Finally, this model supports inherent JAX features such as:
+    - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+    - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+    - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+    - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+    Parameters:
+        config ([`WhisperConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~FlaxPreTrainedModel.from_pretrained`] method to load the model weights.
+        dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+            The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
+            `jax.numpy.bfloat16` (on TPUs). This can be used to enable mixed-precision training or half-precision
+            inference on GPUs or TPUs. If specified all the computation will be performed with the given `dtype`.
+            **Note that this only specifies the dtype of the computation and does not influence the dtype of model
+            parameters.** If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`]
+            and [`~FlaxPreTrainedModel.to_bf16`].
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`numpy.ndarray` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`WhisperFeatureExtractor`] should be used for extracting the features, padding and conversion into a
+            tensor of type `numpy.ndarray`. See [`~WhisperFeatureExtractor.__call__`]
+        attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Whisper does not support masking of the `input_features`, this argument is preserved for compatibility, but
+            is not used. By default the silence in the input log mel spectrogram are ignored.
+        decoder_input_ids (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using
+            [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
+            [What are decoder input IDs?](../glossary#decoder-input-ids) Whisper uses the `decoder_start_token_id` as
+            the starting token for `decoder_input_ids` generation.
+        decoder_attention_mask (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default. If you want to change padding behavior, you should modify to your needs. See diagram 1
+            in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Whisper does not use `position_ids` in the encoder as `input_features` is always the same size and doesn't
+            use masking, but this argument is preserved for compatibility. By default the silence in the input log mel
+            spectrogram are ignored.
+        decoder_position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+WHISPER_ENCODE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`numpy.ndarray` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`WhisperFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `numpy.ndarray`. See [`~WhisperFeatureExtractor.__call__`].
+        attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Whisper does not support masking of the `input_features`, this argument is preserved for compatibility, but
+            is not used. By default the silence in the input log mel spectrogram are ignored.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+WHISPER_DECODE_INPUTS_DOCSTRING = r"""
+    Args:
+        decoder_input_ids (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`):
+            Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using
+            [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+        encoder_outputs (`tuple(tuple(numpy.ndarray)`):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        encoder_attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+           Whisper does not support masking of the `input_features`, this argument is preserved for compatibility,
+            but it is not used. By default the silence in the input log mel spectrogram are ignored.
+        decoder_attention_mask (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default. If you want to change padding behavior, you should modify to your needs. See diagram 1
+            in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        decoder_position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        past_key_values (`Dict[str, numpy.ndarray]`, *optional*, returned by `init_cache` or when passing previous `past_key_values`):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape *[batch_size, max_length]*.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class FlaxWhisperAttention(nn.Module):
+    config: WhisperConfig
+    embed_dim: int
+    num_heads: int
+    dropout: float = 0.0
+    causal: bool = False
+    bias: bool = True
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        dense = partial(
+            nn.Dense,
+            self.embed_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+        self.q_proj = dense(use_bias=self.bias)
+        self.k_proj = dense(use_bias=False)
+        self.v_proj = dense(use_bias=self.bias)
+        self.out_proj = dense(use_bias=self.bias)
+
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_target_positions), dtype="bool"), dtype="bool"
+            )
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        key_value_states: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        query_states = self.q_proj(hidden_states)
+
+        if is_cross_attention:
+            key_states = self.k_proj(key_value_states)
+            value_states = self.v_proj(key_value_states)
+        else:
+            key_states = self.k_proj(hidden_states)
+            value_states = self.v_proj(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask,
+                    (0, 0, mask_shift, 0),
+                    (1, 1, query_length, max_decoder_length),
+                )
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
+            )
+
+        # Convert the boolean attention mask to an attention bias.
+        if attention_mask is not None:
+            # attention mask in the form of attention bias
+            attention_bias = lax.select(
+                attention_mask > 0,
+                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
+            )
+        else:
+            attention_bias = None
+
+        dropout_rng = None
+        if not deterministic and self.dropout > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        attn_weights = dot_product_attention_weights(
+            query_states,
+            key_states,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.dropout,
+            broadcast_dropout=True,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+    def _split_heads(self, hidden_state) -> jnp.ndarray:
+        return hidden_state.reshape(hidden_state.shape[:2] + (self.num_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_state) -> jnp.ndarray:
+        return hidden_state.reshape(hidden_state.shape[:2] + (self.embed_dim,))
+
+    @nn.compact
+    def _concatenate_to_cache(self, key, value, query, attention_mask) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only
+            # attend to those key positions that have already been generated and cached, not the
+            # remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+
+        return key, value, attention_mask
+
+
+# Copied from transformers.models.mbart.modeling_flax_mbart.FlaxMBartEncoderLayer with MBart->Whisper
+class FlaxWhisperEncoderLayer(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxWhisperAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.encoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            dtype=self.dtype,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+        self.activation_fn = ACT2FN[self.config.activation_function]
+        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
+        self.fc1 = nn.Dense(
+            self.config.encoder_ffn_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.fc2 = nn.Dense(
+            self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights = self.self_attn(hidden_states=hidden_states, attention_mask=attention_mask)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class FlaxWhisperEncoderLayerCollection(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
+
+    def setup(self):
+        if self.gradient_checkpointing:
+            FlaxWhisperEncoderCheckpointLayer = remat(FlaxWhisperEncoderLayer, static_argnums=(2, 3))
+            self.layers = [
+                FlaxWhisperEncoderCheckpointLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.encoder_layers)
+            ]
+        else:
+            self.layers = [
+                FlaxWhisperEncoderLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.encoder_layers)
+            ]
+        self.layerdrop = self.config.encoder_layerdrop
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        all_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if not deterministic and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions,
+                    deterministic,
+                )
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = (hidden_states, all_hidden_states, all_attentions)
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+# Copied from transformers.models.mbart.modeling_flax_mbart.FlaxMBartDecoderLayer with MBart->Whisper
+class FlaxWhisperDecoderLayer(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxWhisperAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.decoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            causal=True,
+            dtype=self.dtype,
+        )
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+        self.activation_fn = ACT2FN[self.config.activation_function]
+        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
+
+        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.encoder_attn = FlaxWhisperAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.decoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            dtype=self.dtype,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.fc1 = nn.Dense(
+            self.config.decoder_ffn_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.fc2 = nn.Dense(
+            self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, init_cache=init_cache
+        )
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+            hidden_states, cross_attn_weights = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+            hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        return outputs
+
+
+class FlaxWhisperDecoderLayerCollection(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
+
+    def setup(self):
+        if self.gradient_checkpointing:
+            FlaxWhisperDecoderCheckpointLayer = remat(FlaxWhisperDecoderLayer, static_argnums=(4, 5, 6))
+            self.layers = [
+                FlaxWhisperDecoderCheckpointLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.decoder_layers)
+            ]
+        else:
+            self.layers = [
+                FlaxWhisperDecoderLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.decoder_layers)
+            ]
+        self.layerdrop = self.config.decoder_layerdrop
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+                # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if not deterministic and (dropout_probability < self.layerdrop):
+                layer_outputs = (None, None, None)
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    init_cache,
+                    output_attentions,
+                    deterministic,
+                )
+
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = [hidden_states, all_hidden_states, all_self_attns, all_cross_attentions]
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class FlaxWhisperEncoder(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.conv1 = nn.Conv(
+            self.config.d_model,
+            kernel_size=(3,),
+            padding=1,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            dtype=self.dtype,
+        )
+        self.conv2 = nn.Conv(
+            self.config.d_model,
+            kernel_size=(3,),
+            strides=2,
+            padding=1,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            dtype=self.dtype,
+        )
+
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+
+        self.layers = FlaxWhisperEncoderLayerCollection(
+            self.config,
+            dtype=self.dtype,
+            gradient_checkpointing=self.gradient_checkpointing,
+        )
+
+        self.embed_positions = nn.Embed(
+            self.config.max_source_positions,
+            self.config.d_model,
+            dtype=self.dtype,
+            embedding_init=sinusoidal_embedding_init,
+        )
+
+        self.layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        if input_features.shape[1:] != (self.config.num_mel_bins, self.config.max_source_positions * 2):
+            raise ValueError(
+                "input_features.shape[1:], must be equal to (self.config.num_mel_bins,"
+                f" self.config.max_source_positions * 2) (got {input_features.shape[1:]}, but should be"
+                f" ({self.config.num_mel_bins}, {self.config.max_source_positions * 2}))"
+            )
+
+        input_features = input_features.transpose(0, 2, 1)
+        hidden_states = jax.nn.gelu(self.conv1(input_features), approximate=False)
+        hidden_states = jax.nn.gelu(self.conv2(hidden_states), approximate=False)
+
+        embed_positions = self.embed_positions(jnp.arange(self.config.max_source_positions))
+        # freeze the sinusoidal embeddings by stopping the back-prop
+        embed_positions = jax.lax.stop_gradient(embed_positions)
+        hidden_states = hidden_states + embed_positions
+
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+
+        outputs = self.layers(
+            hidden_states,
+            attention_mask=None,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = self.layer_norm(last_hidden_states)
+
+        # update the last element in `hidden_states` after applying `layernorm` above
+        hidden_states = None
+        if output_hidden_states:
+            hidden_states = outputs[1]
+            hidden_states = hidden_states[:-1] + (last_hidden_states,)
+
+        if not return_dict:
+            outputs = (last_hidden_states, hidden_states) + (outputs[2:] if output_hidden_states else outputs[1:])
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=last_hidden_states,
+            hidden_states=hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class FlaxWhisperDecoder(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.embed_tokens = nn.Embed(self.config.vocab_size, self.config.d_model, dtype=self.dtype)
+        self.embed_positions = nn.Embed(self.config.max_target_positions, self.config.d_model, dtype=self.dtype)
+
+        self.layers = FlaxWhisperDecoderLayerCollection(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+
+        self.layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-5)
+
+    def __call__(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        position_ids: jnp.ndarray,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        input_embeds = self.embed_tokens(input_ids)
+        position_embeds = self.embed_positions(position_ids)
+
+        hidden_states = input_embeds + position_embeds
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+
+        outputs = self.layers(
+            hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = self.layer_norm(last_hidden_states)
+
+        # update the last element in `hidden_states` after applying `layernorm` above
+        hidden_states = None
+        if output_hidden_states:
+            hidden_states = outputs[1]
+            hidden_states = hidden_states[:-1] + (last_hidden_states,)
+
+        if not return_dict:
+            outputs = (last_hidden_states, hidden_states) + (outputs[2:] if output_hidden_states else outputs[1:])
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=last_hidden_states,
+            hidden_states=hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+class FlaxWhisperModule(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.encoder = FlaxWhisperEncoder(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+        self.decoder = FlaxWhisperDecoder(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        decoder_input_ids: jnp.ndarray,
+        decoder_attention_mask: jnp.ndarray,
+        decoder_position_ids: jnp.ndarray,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        encoder_outputs = self.encoder(
+            input_features,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return FlaxSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def _get_encoder_module(self):
+        return self.encoder
+
+    def _get_decoder_module(self):
+        return self.decoder
+
+
+class FlaxWhisperPreTrainedModel(FlaxPreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix: str = "model"
+    main_input_name = "input_features"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: WhisperConfig,
+        input_shape: Tuple[int] = None,
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        gradient_checkpointing: bool = False,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, gradient_checkpointing=gradient_checkpointing, **kwargs)
+        if input_shape is None:
+            input_shape = (1, config.num_mel_bins, 2 * config.max_source_positions)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def enable_gradient_checkpointing(self):
+        self._module = self.module_class(
+            config=self.config,
+            dtype=self.dtype,
+            gradient_checkpointing=True,
+        )
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_features = jnp.zeros(input_shape, dtype="f4")
+        input_features = input_features.at[(..., -1)].set(self.config.eos_token_id)
+
+        decoder_input_ids = jnp.zeros((input_shape[0], 1), dtype="i4")
+        decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        decoder_position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(
+            rngs,
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+        )["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartPreTrainedModel.init_cache with Bart->Whisper
+    def init_cache(self, batch_size, max_length, encoder_outputs):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+            encoder_outputs (`Union[FlaxBaseModelOutput, tuple(tuple(jnp.ndarray)]`):
+                `encoder_outputs` consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*:
+                `attentions`). `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*)
+                is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+                cross-attention of the decoder.
+        """
+        # init input variables to retrieve cache
+        decoder_input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+        decoder_position_ids = jnp.broadcast_to(
+            jnp.arange(jnp.atleast_2d(decoder_input_ids).shape[-1]), decoder_input_ids.shape
+        )
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                decoder_position_ids,
+                **kwargs,
+            )
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0),
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            init_cache=True,
+            method=_decoder_forward,  # we only need to call the decoder to init the cache
+        )
+        return unfreeze(init_variables["cache"])
+
+    @add_start_docstrings(WHISPER_ENCODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxBaseModelOutput, config_class=WhisperConfig)
+    def encode(
+        self,
+        input_features: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        **kwargs,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="np")
+        >>> input_features = inputs.input_features
+        >>> encoder_outputs = model.encode(input_features=input_features)
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _encoder_forward(module, input_features, **kwargs):
+            encode_module = module._get_encoder_module()
+            return encode_module(input_features, **kwargs)
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_features=jnp.array(input_features, dtype="f4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            method=_encoder_forward,
+        )
+
+    @add_start_docstrings(WHISPER_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxBaseModelOutputWithPastAndCrossAttentions, config_class=WhisperConfig)
+    def decode(
+        self,
+        decoder_input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+        >>> import jax.numpy as jnp
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        >>> input_features = processor(ds[0]["audio"]["array"], return_tensors="np").input_features
+
+        >>> encoder_outputs = model.encode(input_features=input_features)
+        >>> decoder_start_token_id = model.config.decoder_start_token_id
+
+        >>> decoder_input_ids = jnp.ones((input_features.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+        >>> outputs = model.decode(decoder_input_ids, encoder_outputs)
+        >>> last_decoder_hidden_states = outputs.last_hidden_state
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
+
+            if decoder_attention_mask is not None:
+                decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
+            else:
+                decoder_position_ids = jnp.broadcast_to(
+                    jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+                )
+
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxWhisperAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                input_ids=decoder_input_ids,
+                attention_mask=decoder_attention_mask,
+                position_ids=decoder_position_ids,
+                **kwargs,
+            )
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past = outputs
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past = outputs
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        decoder_input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # prepare decoder inputs
+        if decoder_position_ids is None:
+            if decoder_attention_mask is not None:
+                decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
+            else:
+                batch_size, sequence_length = decoder_input_ids.shape
+                decoder_position_ids = jnp.broadcast_to(
+                    jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+                )
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_features=jnp.array(input_features, dtype="f4"),
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+
+@add_start_docstrings(
+    "The bare Whisper Model transformer outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class FlaxWhisperModel(FlaxWhisperPreTrainedModel):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    module_class = FlaxWhisperModule
+
+
+append_call_sample_docstring(FlaxWhisperModel, _CHECKPOINT_FOR_DOC, FlaxSeq2SeqModelOutput, _CONFIG_FOR_DOC)
+
+
+class FlaxWhisperForConditionalGenerationModule(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.model = FlaxWhisperModule(
+            config=self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+        self.lm_head = nn.Dense(
+            self.config.vocab_size,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+    def _get_encoder_module(self):
+        return self.model.encoder
+
+    def _get_decoder_module(self):
+        return self.model.decoder
+
+    def __call__(
+        self,
+        input_features,
+        decoder_input_ids,
+        decoder_attention_mask: jnp.ndarray = None,
+        decoder_position_ids: jnp.ndarray = None,
+        position_ids: jnp.ndarray = None,
+        attention_mask: jnp.ndarray = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        outputs = self.model(
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.model.decoder.embed_tokens.variables["params"]["embedding"]
+            lm_logits = self.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+        else:
+            lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return output
+
+        return FlaxSeq2SeqLMOutput(
+            logits=lm_logits,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+
+@add_start_docstrings("The Whisper Model with a language modeling head.", WHISPER_START_DOCSTRING)
+class FlaxWhisperForConditionalGeneration(FlaxWhisperPreTrainedModel):
+    module_class = FlaxWhisperForConditionalGenerationModule
+    dtype: jnp.dtype = jnp.float32
+
+    @add_start_docstrings(WHISPER_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxCausalLMOutputWithCrossAttentions, config_class=WhisperConfig)
+    def decode(
+        self,
+        decoder_input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="np")
+        >>> input_features = inputs.input_features
+        >>> encoder_outputs = model.encode(input_features=input_features)
+        >>> decoder_start_token_id = model.config.decoder_start_token_id
+
+        >>> decoder_input_ids = jnp.ones((inputs.input_ids.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+        >>> outputs = model.decode(decoder_input_ids, encoder_outputs)
+        >>> last_decoder_hidden_states = outputs.last_hidden_state
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
+
+            if decoder_attention_mask is not None:
+                decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
+            else:
+                decoder_position_ids = jnp.broadcast_to(
+                    jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+                )
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length), dtype="i4")
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxWhisperAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            outputs = decoder_module(
+                input_ids=decoder_input_ids,
+                attention_mask=decoder_attention_mask,
+                position_ids=decoder_position_ids,
+                **kwargs,
+            )
+            hidden_states = outputs[0]
+
+            if self.config.tie_word_embeddings:
+                shared_embedding = module.model.decoder.embed_tokens.variables["params"]["embedding"]
+                lm_logits = module.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+            else:
+                lm_logits = module.lm_head(hidden_states)
+
+            return lm_logits, outputs
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        if past_key_values is None:
+            lm_logits, decoder_outputs = outputs
+        else:
+            (lm_logits, decoder_outputs), past = outputs
+
+        if return_dict:
+            outputs = FlaxCausalLMOutputWithCrossAttentions(
+                logits=lm_logits,
+                hidden_states=decoder_outputs.hidden_states,
+                attentions=decoder_outputs.attentions,
+                cross_attentions=decoder_outputs.cross_attentions,
+            )
+        else:
+            outputs = (lm_logits,) + decoder_outputs[1:]
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def generate(
+        self,
+        input_features,
+        generation_config=None,
+        logits_processor=None,
+        return_timestamps=None,
+        task=None,
+        language=None,
+        is_multilingual=None,
+        **kwargs,
+    ):
+        if generation_config is None:
+            generation_config = self.generation_config
+
+        if return_timestamps is not None:
+            generation_config.return_timestamps = return_timestamps
+
+        if task is not None:
+            generation_config.task = task
+
+        if is_multilingual is not None:
+            generation_config.is_multilingual = is_multilingual
+
+        if language is not None:
+            generation_config.language = language
+
+        if kwargs is not None and "decoder_input_ids" in kwargs:
+            decoder_input_length = len(kwargs["decoder_input_ids"])
+        else:
+            decoder_input_length = 1
+
+        forced_decoder_ids = []
+
+        if hasattr(generation_config, "is_multilingual") and generation_config.is_multilingual:
+            if hasattr(generation_config, "language"):
+                forced_decoder_ids.append((1, generation_config.lang_to_id[generation_config.language]))
+            else:
+                forced_decoder_ids.append((1, None))
+
+            if hasattr(generation_config, "task"):
+                forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
+            else:
+                forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))
+
+        if (
+            hasattr(generation_config, "return_timestamps") and generation_config.return_timestamps
+        ) or return_timestamps:
+            logits_processor = [
+                FlaxWhisperTimeStampLogitsProcessor(generation_config, self.config, decoder_input_length)
+            ]
+        else:
+            if forced_decoder_ids and forced_decoder_ids[-1][0] != generation_config.no_timestamps_token_id:
+                idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
+                forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
+
+        if len(forced_decoder_ids) > 0:
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        return super().generate(
+            input_features,
+            generation_config,
+            logits_processor=logits_processor,
+            **kwargs,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        max_length,
+        attention_mask: Optional[jax.Array] = None,
+        decoder_attention_mask: Optional[jax.Array] = None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # initializing the cache
+        batch_size, seq_length = decoder_input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length, encoder_outputs)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyways.
+        # Thus we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if decoder_attention_mask is not None:
+            position_ids = decoder_attention_mask.cumsum(-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, decoder_attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "encoder_outputs": encoder_outputs,
+            "encoder_attention_mask": attention_mask,
+            "decoder_attention_mask": extended_attention_mask,
+            "decoder_position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["decoder_position_ids"] = model_kwargs["decoder_position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+FLAX_WHISPER_CONDITIONAL_GENERATION_DOCSTRING = r"""
+    Returns:
+
+    Transcription example:
+
+    ```python
+    >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+    >>> from datasets import load_dataset
+
+    >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+    >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+    >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+    >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="np")
+    >>> input_features = inputs.input_features
+    >>> generated_ids = model.generate(input_ids=input_features)
+    >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    >>> transcription
+    ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxWhisperForConditionalGeneration, WHISPER_INPUTS_DOCSTRING + FLAX_WHISPER_CONDITIONAL_GENERATION_DOCSTRING
+)
+append_replace_return_docstrings(
+    FlaxWhisperForConditionalGeneration, output_type=FlaxSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
+)
+
+
+class FlaxWhisperForAudioClassificationModule(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.encoder = FlaxWhisperEncoder(
+            config=self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+        self.config.is_encoder_decoder = False
+        num_layers = self.config.num_hidden_layers + 1
+        if self.config.use_weighted_layer_sum:
+            self.layer_weights = jnp.repeat(1 / num_layers, num_layers)
+        self.projector = nn.Dense(self.config.classifier_proj_size, dtype=self.dtype)
+        self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_features,
+        encoder_outputs=None,
+        output_attentions=None,
+        output_hidden_states: bool = True,
+        return_dict: bool = True,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        if self.config.use_weighted_layer_sum:
+            hidden_states = jnp.stack(encoder_outputs, axis=1)
+            norm_weights = jax.nn.softmax(self.layer_weights, axis=-1)
+            hidden_states = jnp.sum(hidden_states * jnp.reshape(norm_weights, [-1, 1, 1]), axis=1)
+        else:
+            hidden_states = encoder_outputs[0]
+
+        hidden_states = self.projector(hidden_states)
+        pooled_output = jnp.mean(hidden_states, axis=1)
+
+        logits = self.classifier(pooled_output)
+
+        if not return_dict:
+            return (logits,) + encoder_outputs[1:]
+
+        return FlaxSequenceClassifierOutput(
+            logits=logits,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings("The Whisper Model with an audio classification head on top.", WHISPER_START_DOCSTRING)
+class FlaxWhisperForAudioClassification(FlaxWhisperPreTrainedModel):
+    module_class = FlaxWhisperForAudioClassificationModule
+    dtype: jnp.dtype = jnp.float32
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_features = jnp.zeros(input_shape, dtype="f4")
+        input_features = input_features.at[(..., -1)].set(self.config.eos_token_id)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(
+            rngs,
+            input_features=input_features,
+        )["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        **kwargs,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_features=jnp.array(input_features, dtype="f4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            rngs=rngs,
+        )
+
+
+FLAX_WHISPER_AUDIO_CLASSIFICATION_DOCSTRING = r"""
+    Returns:
+
+    Transcription example:
+
+    ```python
+    >>> import jax.numpy as jnp
+    >>> from transformers import AutoFeatureExtractor, FlaxWhisperForAudioClassification
+    >>> from datasets import load_dataset
+
+    >>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
+    >>> model = FlaxWhisperForAudioClassification.from_pretrained(
+    ...     "sanchit-gandhi/whisper-medium-fleurs-lang-id", from_pt=True
+    ... )
+    >>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True, trust_remote_code=True)
+
+    >>> sample = next(iter(ds))
+
+    >>> inputs = feature_extractor(
+    ...     sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="np"
+    ... )
+    >>> input_features = inputs.input_features
+
+    >>> logits = model(input_features).logits
+
+    >>> predicted_class_ids = jnp.argmax(logits).item()
+    >>> predicted_label = model.config.id2label[predicted_class_ids]
+    >>> predicted_label
+    'af_za'
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxWhisperForAudioClassification, WHISPER_INPUTS_DOCSTRING + FLAX_WHISPER_AUDIO_CLASSIFICATION_DOCSTRING
+)
+append_replace_return_docstrings(
+    FlaxWhisperForAudioClassification, output_type=FlaxSequenceClassifierOutput, config_class=_CONFIG_FOR_DOC
+)

modeling_flax_whisper.py

@@ -0,0 +1,1696 @@
+# coding=utf-8
+# Copyright 2022 The OpenAI Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Flax whisper model."""
+
+import math
+import random
+from functools import partial
+from typing import Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen import partitioning as nn_partitioning
+from flax.linen.attention import dot_product_attention_weights
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax import lax
+from jax.random import PRNGKey
+
+from ...generation.flax_logits_process import FlaxWhisperTimeStampLogitsProcessor
+from ...modeling_flax_outputs import (
+    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
+    FlaxSeq2SeqLMOutput,
+    FlaxSeq2SeqModelOutput,
+    FlaxSequenceClassifierOutput,
+)
+from ...modeling_flax_utils import (
+    ACT2FN,
+    FlaxPreTrainedModel,
+    append_call_sample_docstring,
+    append_replace_return_docstrings,
+    overwrite_call_docstring,
+)
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_whisper import WhisperConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+_CHECKPOINT_FOR_DOC = "openai/whisper-tiny"
+_CONFIG_FOR_DOC = "WhisperConfig"
+
+remat = nn_partitioning.remat
+
+
+def sinusoidal_embedding_init(key, shape, dtype=jnp.float_) -> jax.Array:
+    """Returns sinusoids for positional embedding"""
+    length, channels = shape
+    if channels % 2 != 0:
+        raise ValueError(
+            f"Number of channels has to be divisible by 2 for sinusoidal positional embeddings, got {channels} channels."
+        )
+    log_timescale_increment = math.log(10000) / (channels // 2 - 1)
+    inv_timescales = jnp.exp(-log_timescale_increment * jnp.arange(channels // 2))
+    scaled_time = jnp.arange(length).reshape(-1, 1) * inv_timescales.reshape(1, -1)
+    return jnp.concatenate([jnp.sin(scaled_time), jnp.cos(scaled_time)], axis=1).astype(dtype)
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its models (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.) This model is also a Flax Linen
+    [flax.nn.Module](https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html) subclass. Use it as a
+    regular Flax Module and refer to the Flax documentation for all matter related to general usage and behavior.
+    Finally, this model supports inherent JAX features such as:
+    - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+    - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+    - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+    - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+    Parameters:
+        config ([`WhisperConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~FlaxPreTrainedModel.from_pretrained`] method to load the model weights.
+        dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+            The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
+            `jax.numpy.bfloat16` (on TPUs). This can be used to enable mixed-precision training or half-precision
+            inference on GPUs or TPUs. If specified all the computation will be performed with the given `dtype`.
+            **Note that this only specifies the dtype of the computation and does not influence the dtype of model
+            parameters.** If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`]
+            and [`~FlaxPreTrainedModel.to_bf16`].
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`numpy.ndarray` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`WhisperFeatureExtractor`] should be used for extracting the features, padding and conversion into a
+            tensor of type `numpy.ndarray`. See [`~WhisperFeatureExtractor.__call__`]
+        attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Whisper does not support masking of the `input_features`, this argument is preserved for compatibility, but
+            is not used. By default the silence in the input log mel spectrogram are ignored.
+        decoder_input_ids (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using
+            [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
+            [What are decoder input IDs?](../glossary#decoder-input-ids) Whisper uses the `decoder_start_token_id` as
+            the starting token for `decoder_input_ids` generation.
+        decoder_attention_mask (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default. If you want to change padding behavior, you should modify to your needs. See diagram 1
+            in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Whisper does not use `position_ids` in the encoder as `input_features` is always the same size and doesn't
+            use masking, but this argument is preserved for compatibility. By default the silence in the input log mel
+            spectrogram are ignored.
+        decoder_position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+WHISPER_ENCODE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`numpy.ndarray` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`WhisperFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `numpy.ndarray`. See [`~WhisperFeatureExtractor.__call__`].
+        attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Whisper does not support masking of the `input_features`, this argument is preserved for compatibility, but
+            is not used. By default the silence in the input log mel spectrogram are ignored.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+WHISPER_DECODE_INPUTS_DOCSTRING = r"""
+    Args:
+        decoder_input_ids (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`):
+            Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using
+            [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+        encoder_outputs (`tuple(tuple(numpy.ndarray)`):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        encoder_attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+           Whisper does not support masking of the `input_features`, this argument is preserved for compatibility,
+            but it is not used. By default the silence in the input log mel spectrogram are ignored.
+        decoder_attention_mask (`numpy.ndarray` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default. If you want to change padding behavior, you should modify to your needs. See diagram 1
+            in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        decoder_position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+            range `[0, config.max_position_embeddings - 1]`.
+        past_key_values (`Dict[str, numpy.ndarray]`, *optional*, returned by `init_cache` or when passing previous `past_key_values`):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape *[batch_size, max_length]*.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class FlaxWhisperAttention(nn.Module):
+    config: WhisperConfig
+    embed_dim: int
+    num_heads: int
+    dropout: float = 0.0
+    causal: bool = False
+    bias: bool = True
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        dense = partial(
+            nn.Dense,
+            self.embed_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+        self.q_proj = dense(use_bias=self.bias)
+        self.k_proj = dense(use_bias=False)
+        self.v_proj = dense(use_bias=self.bias)
+        self.out_proj = dense(use_bias=self.bias)
+
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_target_positions), dtype="bool"), dtype="bool"
+            )
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        key_value_states: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        query_states = self.q_proj(hidden_states)
+
+        if is_cross_attention:
+            key_states = self.k_proj(key_value_states)
+            value_states = self.v_proj(key_value_states)
+        else:
+            key_states = self.k_proj(hidden_states)
+            value_states = self.v_proj(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask,
+                    (0, 0, mask_shift, 0),
+                    (1, 1, query_length, max_decoder_length),
+                )
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
+            )
+
+        # Convert the boolean attention mask to an attention bias.
+        if attention_mask is not None:
+            # attention mask in the form of attention bias
+            attention_bias = lax.select(
+                attention_mask > 0,
+                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
+            )
+        else:
+            attention_bias = None
+
+        dropout_rng = None
+        if not deterministic and self.dropout > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        attn_weights = dot_product_attention_weights(
+            query_states,
+            key_states,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.dropout,
+            broadcast_dropout=True,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+    def _split_heads(self, hidden_state) -> jnp.ndarray:
+        return hidden_state.reshape(hidden_state.shape[:2] + (self.num_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_state) -> jnp.ndarray:
+        return hidden_state.reshape(hidden_state.shape[:2] + (self.embed_dim,))
+
+    @nn.compact
+    def _concatenate_to_cache(self, key, value, query, attention_mask) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only
+            # attend to those key positions that have already been generated and cached, not the
+            # remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+
+        return key, value, attention_mask
+
+
+# Copied from transformers.models.mbart.modeling_flax_mbart.FlaxMBartEncoderLayer with MBart->Whisper
+class FlaxWhisperEncoderLayer(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxWhisperAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.encoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            dtype=self.dtype,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+        self.activation_fn = ACT2FN[self.config.activation_function]
+        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
+        self.fc1 = nn.Dense(
+            self.config.encoder_ffn_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.fc2 = nn.Dense(
+            self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights = self.self_attn(hidden_states=hidden_states, attention_mask=attention_mask)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class FlaxWhisperEncoderLayerCollection(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
+
+    def setup(self):
+        if self.gradient_checkpointing:
+            FlaxWhisperEncoderCheckpointLayer = remat(FlaxWhisperEncoderLayer, static_argnums=(2, 3))
+            self.layers = [
+                FlaxWhisperEncoderCheckpointLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.encoder_layers)
+            ]
+        else:
+            self.layers = [
+                FlaxWhisperEncoderLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.encoder_layers)
+            ]
+        self.layerdrop = self.config.encoder_layerdrop
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        all_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if not deterministic and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions,
+                    deterministic,
+                )
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = (hidden_states, all_hidden_states, all_attentions)
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+# Copied from transformers.models.mbart.modeling_flax_mbart.FlaxMBartDecoderLayer with MBart->Whisper
+class FlaxWhisperDecoderLayer(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self) -> None:
+        self.embed_dim = self.config.d_model
+        self.self_attn = FlaxWhisperAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.decoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            causal=True,
+            dtype=self.dtype,
+        )
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+        self.activation_fn = ACT2FN[self.config.activation_function]
+        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
+
+        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.encoder_attn = FlaxWhisperAttention(
+            config=self.config,
+            embed_dim=self.embed_dim,
+            num_heads=self.config.decoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            dtype=self.dtype,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.fc1 = nn.Dense(
+            self.config.decoder_ffn_dim,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.fc2 = nn.Dense(
+            self.embed_dim, dtype=self.dtype, kernel_init=jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states, attention_mask=attention_mask, init_cache=init_cache
+        )
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+            hidden_states, cross_attn_weights = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+            hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        return outputs
+
+
+class FlaxWhisperDecoderLayerCollection(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
+
+    def setup(self):
+        if self.gradient_checkpointing:
+            FlaxWhisperDecoderCheckpointLayer = remat(FlaxWhisperDecoderLayer, static_argnums=(4, 5, 6))
+            self.layers = [
+                FlaxWhisperDecoderCheckpointLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.decoder_layers)
+            ]
+        else:
+            self.layers = [
+                FlaxWhisperDecoderLayer(self.config, name=str(i), dtype=self.dtype)
+                for i in range(self.config.decoder_layers)
+            ]
+        self.layerdrop = self.config.decoder_layerdrop
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+                # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if not deterministic and (dropout_probability < self.layerdrop):
+                layer_outputs = (None, None, None)
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    init_cache,
+                    output_attentions,
+                    deterministic,
+                )
+
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = [hidden_states, all_hidden_states, all_self_attns, all_cross_attentions]
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class FlaxWhisperEncoder(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.conv1 = nn.Conv(
+            self.config.d_model,
+            kernel_size=(3,),
+            padding=1,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            dtype=self.dtype,
+        )
+        self.conv2 = nn.Conv(
+            self.config.d_model,
+            kernel_size=(3,),
+            strides=2,
+            padding=1,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            dtype=self.dtype,
+        )
+
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+
+        self.layers = FlaxWhisperEncoderLayerCollection(
+            self.config,
+            dtype=self.dtype,
+            gradient_checkpointing=self.gradient_checkpointing,
+        )
+
+        self.embed_positions = nn.Embed(
+            self.config.max_source_positions,
+            self.config.d_model,
+            dtype=self.dtype,
+            embedding_init=sinusoidal_embedding_init,
+        )
+
+        self.layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        if input_features.shape[1:] != (self.config.num_mel_bins, self.config.max_source_positions * 2):
+            raise ValueError(
+                "input_features.shape[1:], must be equal to (self.config.num_mel_bins,"
+                f" self.config.max_source_positions * 2) (got {input_features.shape[1:]}, but should be"
+                f" ({self.config.num_mel_bins}, {self.config.max_source_positions * 2}))"
+            )
+
+        input_features = input_features.transpose(0, 2, 1)
+        hidden_states = jax.nn.gelu(self.conv1(input_features), approximate=False)
+        hidden_states = jax.nn.gelu(self.conv2(hidden_states), approximate=False)
+
+        embed_positions = self.embed_positions(jnp.arange(self.config.max_source_positions))
+        # freeze the sinusoidal embeddings by stopping the back-prop
+        embed_positions = jax.lax.stop_gradient(embed_positions)
+        hidden_states = hidden_states + embed_positions
+
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+
+        outputs = self.layers(
+            hidden_states,
+            attention_mask=None,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = self.layer_norm(last_hidden_states)
+
+        # update the last element in `hidden_states` after applying `layernorm` above
+        hidden_states = None
+        if output_hidden_states:
+            hidden_states = outputs[1]
+            hidden_states = hidden_states[:-1] + (last_hidden_states,)
+
+        if not return_dict:
+            outputs = (last_hidden_states, hidden_states) + (outputs[2:] if output_hidden_states else outputs[1:])
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=last_hidden_states,
+            hidden_states=hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class FlaxWhisperDecoder(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.embed_tokens = nn.Embed(self.config.vocab_size, self.config.d_model, dtype=self.dtype)
+        self.embed_positions = nn.Embed(self.config.max_target_positions, self.config.d_model, dtype=self.dtype)
+
+        self.layers = FlaxWhisperDecoderLayerCollection(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+
+        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
+
+        self.layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-5)
+
+    def __call__(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        position_ids: jnp.ndarray,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        input_embeds = self.embed_tokens(input_ids)
+        position_embeds = self.embed_positions(position_ids)
+
+        hidden_states = input_embeds + position_embeds
+        hidden_states = self.dropout_layer(hidden_states, deterministic=deterministic)
+
+        outputs = self.layers(
+            hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_states = outputs[0]
+        last_hidden_states = self.layer_norm(last_hidden_states)
+
+        # update the last element in `hidden_states` after applying `layernorm` above
+        hidden_states = None
+        if output_hidden_states:
+            hidden_states = outputs[1]
+            hidden_states = hidden_states[:-1] + (last_hidden_states,)
+
+        if not return_dict:
+            outputs = (last_hidden_states, hidden_states) + (outputs[2:] if output_hidden_states else outputs[1:])
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=last_hidden_states,
+            hidden_states=hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+class FlaxWhisperModule(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.encoder = FlaxWhisperEncoder(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+        self.decoder = FlaxWhisperDecoder(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        decoder_input_ids: jnp.ndarray,
+        decoder_attention_mask: jnp.ndarray,
+        decoder_position_ids: jnp.ndarray,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        encoder_outputs = self.encoder(
+            input_features,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return FlaxSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def _get_encoder_module(self):
+        return self.encoder
+
+    def _get_decoder_module(self):
+        return self.decoder
+
+
+class FlaxWhisperPreTrainedModel(FlaxPreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix: str = "model"
+    main_input_name = "input_features"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: WhisperConfig,
+        input_shape: Tuple[int] = None,
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        gradient_checkpointing: bool = False,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, gradient_checkpointing=gradient_checkpointing, **kwargs)
+        if input_shape is None:
+            input_shape = (1, config.num_mel_bins, 2 * config.max_source_positions)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def enable_gradient_checkpointing(self):
+        self._module = self.module_class(
+            config=self.config,
+            dtype=self.dtype,
+            gradient_checkpointing=True,
+        )
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_features = jnp.zeros(input_shape, dtype="f4")
+        input_features = input_features.at[(..., -1)].set(self.config.eos_token_id)
+
+        decoder_input_ids = jnp.zeros((input_shape[0], 1), dtype="i4")
+        decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        decoder_position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(
+            rngs,
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+        )["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartPreTrainedModel.init_cache with Bart->Whisper
+    def init_cache(self, batch_size, max_length, encoder_outputs):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+            encoder_outputs (`Union[FlaxBaseModelOutput, tuple(tuple(jnp.ndarray)]`):
+                `encoder_outputs` consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*:
+                `attentions`). `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*)
+                is a sequence of hidden-states at the output of the last layer of the encoder. Used in the
+                cross-attention of the decoder.
+        """
+        # init input variables to retrieve cache
+        decoder_input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+        decoder_position_ids = jnp.broadcast_to(
+            jnp.arange(jnp.atleast_2d(decoder_input_ids).shape[-1]), decoder_input_ids.shape
+        )
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                decoder_input_ids,
+                decoder_attention_mask,
+                decoder_position_ids,
+                **kwargs,
+            )
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0),
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            init_cache=True,
+            method=_decoder_forward,  # we only need to call the decoder to init the cache
+        )
+        return unfreeze(init_variables["cache"])
+
+    @add_start_docstrings(WHISPER_ENCODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxBaseModelOutput, config_class=WhisperConfig)
+    def encode(
+        self,
+        input_features: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        **kwargs,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="np")
+        >>> input_features = inputs.input_features
+        >>> encoder_outputs = model.encode(input_features=input_features)
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _encoder_forward(module, input_features, **kwargs):
+            encode_module = module._get_encoder_module()
+            return encode_module(input_features, **kwargs)
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_features=jnp.array(input_features, dtype="f4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            method=_encoder_forward,
+        )
+
+    @add_start_docstrings(WHISPER_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxBaseModelOutputWithPastAndCrossAttentions, config_class=WhisperConfig)
+    def decode(
+        self,
+        decoder_input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+        >>> import jax.numpy as jnp
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        >>> input_features = processor(ds[0]["audio"]["array"], return_tensors="np").input_features
+
+        >>> encoder_outputs = model.encode(input_features=input_features)
+        >>> decoder_start_token_id = model.config.decoder_start_token_id
+
+        >>> decoder_input_ids = jnp.ones((input_features.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+        >>> outputs = model.decode(decoder_input_ids, encoder_outputs)
+        >>> last_decoder_hidden_states = outputs.last_hidden_state
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
+
+            if decoder_attention_mask is not None:
+                decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
+            else:
+                decoder_position_ids = jnp.broadcast_to(
+                    jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+                )
+
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxWhisperAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                input_ids=decoder_input_ids,
+                attention_mask=decoder_attention_mask,
+                position_ids=decoder_position_ids,
+                **kwargs,
+            )
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past = outputs
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past = outputs
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        decoder_input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # prepare decoder inputs
+        if decoder_position_ids is None:
+            if decoder_attention_mask is not None:
+                decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
+            else:
+                batch_size, sequence_length = decoder_input_ids.shape
+                decoder_position_ids = jnp.broadcast_to(
+                    jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+                )
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones_like(decoder_input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_features=jnp.array(input_features, dtype="f4"),
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+
+@add_start_docstrings(
+    "The bare Whisper Model transformer outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class FlaxWhisperModel(FlaxWhisperPreTrainedModel):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    module_class = FlaxWhisperModule
+
+
+append_call_sample_docstring(FlaxWhisperModel, _CHECKPOINT_FOR_DOC, FlaxSeq2SeqModelOutput, _CONFIG_FOR_DOC)
+
+
+class FlaxWhisperForConditionalGenerationModule(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.model = FlaxWhisperModule(
+            config=self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+        self.lm_head = nn.Dense(
+            self.config.vocab_size,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+    def _get_encoder_module(self):
+        return self.model.encoder
+
+    def _get_decoder_module(self):
+        return self.model.decoder
+
+    def __call__(
+        self,
+        input_features,
+        decoder_input_ids,
+        decoder_attention_mask: jnp.ndarray = None,
+        decoder_position_ids: jnp.ndarray = None,
+        position_ids: jnp.ndarray = None,
+        attention_mask: jnp.ndarray = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        outputs = self.model(
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.model.decoder.embed_tokens.variables["params"]["embedding"]
+            lm_logits = self.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+        else:
+            lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return output
+
+        return FlaxSeq2SeqLMOutput(
+            logits=lm_logits,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+
+@add_start_docstrings("The Whisper Model with a language modeling head.", WHISPER_START_DOCSTRING)
+class FlaxWhisperForConditionalGeneration(FlaxWhisperPreTrainedModel):
+    module_class = FlaxWhisperForConditionalGenerationModule
+    dtype: jnp.dtype = jnp.float32
+
+    @add_start_docstrings(WHISPER_DECODE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=FlaxCausalLMOutputWithCrossAttentions, config_class=WhisperConfig)
+    def decode(
+        self,
+        decoder_input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_attention_mask: Optional[jnp.ndarray] = None,
+        decoder_position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="np")
+        >>> input_features = inputs.input_features
+        >>> encoder_outputs = model.encode(input_features=input_features)
+        >>> decoder_start_token_id = model.config.decoder_start_token_id
+
+        >>> decoder_input_ids = jnp.ones((inputs.input_ids.shape[0], 1), dtype="i4") * decoder_start_token_id
+
+        >>> outputs = model.decode(decoder_input_ids, encoder_outputs)
+        >>> last_decoder_hidden_states = outputs.last_hidden_state
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        batch_size, sequence_length = decoder_input_ids.shape
+        if decoder_position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
+
+            if decoder_attention_mask is not None:
+                decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
+            else:
+                decoder_position_ids = jnp.broadcast_to(
+                    jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+                )
+        if decoder_attention_mask is None:
+            decoder_attention_mask = jnp.ones((batch_size, sequence_length), dtype="i4")
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxWhisperAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, decoder_position_ids, **kwargs):
+            decoder_module = module._get_decoder_module()
+            outputs = decoder_module(
+                input_ids=decoder_input_ids,
+                attention_mask=decoder_attention_mask,
+                position_ids=decoder_position_ids,
+                **kwargs,
+            )
+            hidden_states = outputs[0]
+
+            if self.config.tie_word_embeddings:
+                shared_embedding = module.model.decoder.embed_tokens.variables["params"]["embedding"]
+                lm_logits = module.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+            else:
+                lm_logits = module.lm_head(hidden_states)
+
+            return lm_logits, outputs
+
+        outputs = self.module.apply(
+            inputs,
+            decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
+            decoder_position_ids=jnp.array(decoder_position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        if past_key_values is None:
+            lm_logits, decoder_outputs = outputs
+        else:
+            (lm_logits, decoder_outputs), past = outputs
+
+        if return_dict:
+            outputs = FlaxCausalLMOutputWithCrossAttentions(
+                logits=lm_logits,
+                hidden_states=decoder_outputs.hidden_states,
+                attentions=decoder_outputs.attentions,
+                cross_attentions=decoder_outputs.cross_attentions,
+            )
+        else:
+            outputs = (lm_logits,) + decoder_outputs[1:]
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def generate(
+        self,
+        input_features,
+        generation_config=None,
+        logits_processor=None,
+        return_timestamps=None,
+        task=None,
+        language=None,
+        is_multilingual=None,
+        **kwargs,
+    ):
+        if generation_config is None:
+            generation_config = self.generation_config
+
+        if return_timestamps is not None:
+            generation_config.return_timestamps = return_timestamps
+
+        if task is not None:
+            generation_config.task = task
+
+        if is_multilingual is not None:
+            generation_config.is_multilingual = is_multilingual
+
+        if language is not None:
+            generation_config.language = language
+
+        if kwargs is not None and "decoder_input_ids" in kwargs:
+            decoder_input_length = len(kwargs["decoder_input_ids"])
+        else:
+            decoder_input_length = 1
+
+        forced_decoder_ids = []
+
+        if hasattr(generation_config, "is_multilingual") and generation_config.is_multilingual:
+            if hasattr(generation_config, "language"):
+                forced_decoder_ids.append((1, generation_config.lang_to_id[generation_config.language]))
+            else:
+                forced_decoder_ids.append((1, None))
+
+            if hasattr(generation_config, "task"):
+                forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
+            else:
+                forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))
+
+        if (
+            hasattr(generation_config, "return_timestamps") and generation_config.return_timestamps
+        ) or return_timestamps:
+            logits_processor = [
+                FlaxWhisperTimeStampLogitsProcessor(generation_config, self.config, decoder_input_length)
+            ]
+        else:
+            if forced_decoder_ids and forced_decoder_ids[-1][0] != generation_config.no_timestamps_token_id:
+                idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
+                forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
+
+        if len(forced_decoder_ids) > 0:
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        return super().generate(
+            input_features,
+            generation_config,
+            logits_processor=logits_processor,
+            **kwargs,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        max_length,
+        attention_mask: Optional[jax.Array] = None,
+        decoder_attention_mask: Optional[jax.Array] = None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # initializing the cache
+        batch_size, seq_length = decoder_input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length, encoder_outputs)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyways.
+        # Thus we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if decoder_attention_mask is not None:
+            position_ids = decoder_attention_mask.cumsum(-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, decoder_attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "encoder_outputs": encoder_outputs,
+            "encoder_attention_mask": attention_mask,
+            "decoder_attention_mask": extended_attention_mask,
+            "decoder_position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["decoder_position_ids"] = model_kwargs["decoder_position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+FLAX_WHISPER_CONDITIONAL_GENERATION_DOCSTRING = r"""
+    Returns:
+
+    Transcription example:
+
+    ```python
+    >>> from transformers import WhisperProcessor, FlaxWhisperForConditionalGeneration
+    >>> from datasets import load_dataset
+
+    >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+    >>> model = FlaxWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", from_pt=True)
+    >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+    >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="np")
+    >>> input_features = inputs.input_features
+    >>> generated_ids = model.generate(input_ids=input_features)
+    >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    >>> transcription
+    ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxWhisperForConditionalGeneration, WHISPER_INPUTS_DOCSTRING + FLAX_WHISPER_CONDITIONAL_GENERATION_DOCSTRING
+)
+append_replace_return_docstrings(
+    FlaxWhisperForConditionalGeneration, output_type=FlaxSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC
+)
+
+
+class FlaxWhisperForAudioClassificationModule(nn.Module):
+    config: WhisperConfig
+    dtype: jnp.dtype = jnp.float32
+    gradient_checkpointing: bool = False
+
+    def setup(self) -> None:
+        self.encoder = FlaxWhisperEncoder(
+            config=self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
+        self.config.is_encoder_decoder = False
+        num_layers = self.config.num_hidden_layers + 1
+        if self.config.use_weighted_layer_sum:
+            self.layer_weights = jnp.repeat(1 / num_layers, num_layers)
+        self.projector = nn.Dense(self.config.classifier_proj_size, dtype=self.dtype)
+        self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_features,
+        encoder_outputs=None,
+        output_attentions=None,
+        output_hidden_states: bool = True,
+        return_dict: bool = True,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        if self.config.use_weighted_layer_sum:
+            hidden_states = jnp.stack(encoder_outputs, axis=1)
+            norm_weights = jax.nn.softmax(self.layer_weights, axis=-1)
+            hidden_states = jnp.sum(hidden_states * jnp.reshape(norm_weights, [-1, 1, 1]), axis=1)
+        else:
+            hidden_states = encoder_outputs[0]
+
+        hidden_states = self.projector(hidden_states)
+        pooled_output = jnp.mean(hidden_states, axis=1)
+
+        logits = self.classifier(pooled_output)
+
+        if not return_dict:
+            return (logits,) + encoder_outputs[1:]
+
+        return FlaxSequenceClassifierOutput(
+            logits=logits,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings("The Whisper Model with an audio classification head on top.", WHISPER_START_DOCSTRING)
+class FlaxWhisperForAudioClassification(FlaxWhisperPreTrainedModel):
+    module_class = FlaxWhisperForAudioClassificationModule
+    dtype: jnp.dtype = jnp.float32
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_features = jnp.zeros(input_shape, dtype="f4")
+        input_features = input_features.at[(..., -1)].set(self.config.eos_token_id)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(
+            rngs,
+            input_features=input_features,
+        )["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_features: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+        **kwargs,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_features=jnp.array(input_features, dtype="f4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            rngs=rngs,
+        )
+
+
+FLAX_WHISPER_AUDIO_CLASSIFICATION_DOCSTRING = r"""
+    Returns:
+
+    Transcription example:
+
+    ```python
+    >>> import jax.numpy as jnp
+    >>> from transformers import AutoFeatureExtractor, FlaxWhisperForAudioClassification
+    >>> from datasets import load_dataset
+
+    >>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
+    >>> model = FlaxWhisperForAudioClassification.from_pretrained(
+    ...     "sanchit-gandhi/whisper-medium-fleurs-lang-id", from_pt=True
+    ... )
+    >>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True, trust_remote_code=True)
+
+    >>> sample = next(iter(ds))
+
+    >>> inputs = feature_extractor(
+    ...     sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="np"
+    ... )
+    >>> input_features = inputs.input_features
+
+    >>> logits = model(input_features).logits
+
+    >>> predicted_class_ids = jnp.argmax(logits).item()
+    >>> predicted_label = model.config.id2label[predicted_class_ids]
+    >>> predicted_label
+    'af_za'
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxWhisperForAudioClassification, WHISPER_INPUTS_DOCSTRING + FLAX_WHISPER_AUDIO_CLASSIFICATION_DOCSTRING
+)
+append_replace_return_docstrings(
+    FlaxWhisperForAudioClassification, output_type=FlaxSequenceClassifierOutput, config_class=_CONFIG_FOR_DOC
+)

modeling_tf_whisper (1).py

@@ -0,0 +1,1758 @@
+# coding=utf-8
+# Copyright 2022 The OpenAI Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""TensorFlow Whisper model."""
+
+from __future__ import annotations
+
+import math
+import random
+from typing import Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...generation.configuration_utils import GenerationConfig
+from ...generation.tf_logits_process import TFLogitsProcessorList
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPastAndCrossAttentions,
+    TFSeq2SeqLMOutput,
+    TFSeq2SeqModelOutput,
+)
+from ...modeling_tf_utils import (
+    TFCausalLanguageModelingLoss,
+    TFModelInputType,
+    TFPreTrainedModel,
+    keras,
+    keras_serializable,
+    unpack_inputs,
+)
+from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_whisper import WhisperConfig
+from .tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "WhisperConfig"
+
+
+LARGE_NEGATIVE = -1e8
+
+
+def sinusoidal_embedding_init(shape, dtype=tf.float32) -> tf.Tensor:
+    """Returns sinusoids for positional embedding"""
+    length, channels = shape
+    if channels % 2 != 0:
+        raise ValueError(
+            f"Number of channels has to be divisible by 2 for sinusoidal positional embeddings, got {channels} channels."
+        )
+    log_timescale_increment = math.log(10000) / (channels // 2 - 1)
+    inv_timescales = tf.exp(-log_timescale_increment * tf.range(channels // 2, dtype=tf.float32))
+    scaled_time = tf.reshape(tf.range(length, dtype=tf.float32), (-1, 1)) * tf.reshape(inv_timescales, (1, -1))
+    return tf.cast(tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=1), dtype)
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.shift_tokens_right
+def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    pad_token_id = tf.cast(pad_token_id, input_ids.dtype)
+    decoder_start_token_id = tf.cast(decoder_start_token_id, input_ids.dtype)
+    start_tokens = tf.fill(
+        (shape_list(input_ids)[0], 1), tf.convert_to_tensor(decoder_start_token_id, input_ids.dtype)
+    )
+    shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids = tf.where(
+        shifted_input_ids == -100,
+        tf.fill(shape_list(shifted_input_ids), tf.convert_to_tensor(pad_token_id, input_ids.dtype)),
+        shifted_input_ids,
+    )
+
+    # "Verify that `labels` has only positive values and -100"
+    assert_gte0 = tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=input_ids.dtype))
+
+    # Make sure the assertion op is called by wrapping the result in an identity no-op
+    with tf.control_dependencies([assert_gte0]):
+        shifted_input_ids = tf.identity(shifted_input_ids)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz = input_ids_shape[0]
+    tgt_len = input_ids_shape[1]
+    mask = tf.ones((tgt_len, tgt_len)) * LARGE_NEGATIVE
+    mask_cond = tf.range(shape_list(mask)[-1])
+
+    mask = tf.where(mask_cond < tf.reshape(mask_cond + 1, (shape_list(mask)[-1], 1)), 0.0, mask)
+
+    if past_key_values_length > 0:
+        mask = tf.concat([tf.zeros((tgt_len, past_key_values_length)), mask], axis=-1)
+
+    return tf.tile(mask[None, None, :, :], (bsz, 1, 1, 1))
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._expand_mask
+def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    src_len = shape_list(mask)[1]
+    tgt_len = tgt_len if tgt_len is not None else src_len
+    one_cst = tf.constant(1.0)
+    mask = tf.cast(mask, dtype=one_cst.dtype)
+    expanded_mask = tf.tile(mask[:, None, None, :], (1, 1, tgt_len, 1))
+
+    return (one_cst - expanded_mask) * LARGE_NEGATIVE
+
+
+class TFWhisperPositionalEmbedding(keras.layers.Layer):
+    def __init__(
+        self,
+        num_positions: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        embedding_initializer=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_positions = num_positions
+        self.embedding_dim = embedding_dim
+        self.padding_idx = padding_idx
+        self.embedding_initializer = keras.initializers.get(embedding_initializer)
+
+    def build(self, input_shape):
+        self.weight = self.add_weight(
+            name="weight",
+            shape=[self.num_positions, self.embedding_dim],
+            initializer=self.embedding_initializer,
+            trainable=True,
+        )
+        super().build(input_shape)
+
+    def call(self, input_ids, past_key_values_length=0):
+        past_key_values_length = tf.cast(past_key_values_length, tf.int32)
+        gather_indices = tf.range(tf.shape(input_ids)[1], delta=1) + past_key_values_length
+        return tf.gather(self.weight, gather_indices)
+
+
+class TFWhisperAttention(keras.layers.Layer):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = keras.layers.Dropout(dropout)
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = keras.layers.Dense(embed_dim, use_bias=False, name="k_proj")
+        self.v_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
+        self.q_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
+        self.out_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention._shape with BART->whisper
+    def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int):
+        return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), (0, 2, 1, 3))
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention.call with BART->whisper
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        key_value_states: tf.Tensor | None = None,
+        past_key_value: Tuple[Tuple[tf.Tensor]] | None = None,
+        attention_mask: tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Tuple[tf.Tensor, tf.Tensor | None]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = shape_list(hidden_states)
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = tf.concat([past_key_value[0], key_states], axis=2)
+            value_states = tf.concat([past_key_value[1], value_states], axis=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(tf.Tensor, tf.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(tf.Tensor, tf.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = tf.reshape(self._shape(query_states, tgt_len, bsz), proj_shape)
+        key_states = tf.reshape(key_states, proj_shape)
+        value_states = tf.reshape(value_states, proj_shape)
+
+        src_len = shape_list(key_states)[1]
+        attn_weights = tf.matmul(query_states, key_states, transpose_b=True)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_weights),
+            [bsz * self.num_heads, tgt_len, src_len],
+            message=(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {shape_list(attn_weights)}"
+            ),
+        )
+
+        if attention_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(attention_mask),
+                [bsz, 1, tgt_len, src_len],
+                message=(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {shape_list(attention_mask)}"
+                ),
+            )
+
+            attention_mask = tf.cast(attention_mask, dtype=attn_weights.dtype)
+            attn_weights = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len)) + attention_mask
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_weights = stable_softmax(attn_weights, axis=-1)
+
+        if layer_head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(layer_head_mask),
+                [self.num_heads],
+                message=(
+                    f"Head mask for a single layer should be of size {(self.num_heads)}, but is"
+                    f" {shape_list(layer_head_mask)}"
+                ),
+            )
+
+            attn_weights = tf.reshape(layer_head_mask, (1, -1, 1, 1)) * tf.reshape(
+                attn_weights, (bsz, self.num_heads, tgt_len, src_len)
+            )
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_probs = self.dropout(attn_weights, training=training)
+        attn_output = tf.matmul(attn_probs, value_states)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_output),
+            [bsz * self.num_heads, tgt_len, self.head_dim],
+            message=(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {shape_list(attn_output)}"
+            ),
+        )
+
+        attn_output = tf.transpose(
+            tf.reshape(attn_output, (bsz, self.num_heads, tgt_len, self.head_dim)), (0, 2, 1, 3)
+        )
+        attn_output = tf.reshape(attn_output, (bsz, tgt_len, embed_dim))
+
+        attn_output = self.out_proj(attn_output)
+        attn_weights: tf.Tensor = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len))
+
+        return attn_output, attn_weights, past_key_value
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "k_proj", None) is not None:
+            with tf.name_scope(self.k_proj.name):
+                self.k_proj.build([None, None, self.embed_dim])
+        if getattr(self, "v_proj", None) is not None:
+            with tf.name_scope(self.v_proj.name):
+                self.v_proj.build([None, None, self.embed_dim])
+        if getattr(self, "q_proj", None) is not None:
+            with tf.name_scope(self.q_proj.name):
+                self.q_proj.build([None, None, self.embed_dim])
+        if getattr(self, "out_proj", None) is not None:
+            with tf.name_scope(self.out_proj.name):
+                self.out_proj.build([None, None, self.embed_dim])
+
+
+# Copied from transformers.models.speech_to_text.modeling_tf_speech_to_text.TFSpeech2TextEncoderLayer with Speech2Text->Whisper
+class TFWhisperEncoderLayer(keras.layers.Layer):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFWhisperAttention(
+            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
+        )
+        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
+        self.fc1 = keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
+        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+        self.config = config
+
+    def call(
+        self, hidden_states: tf.Tensor, attention_mask: tf.Tensor, layer_head_mask: tf.Tensor, training: bool = False
+    ):
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, self_attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            training=training,
+        )
+
+        tf.debugging.assert_equal(
+            shape_list(hidden_states),
+            shape_list(residual),
+            message=f"Self attn modified the shape of query {shape_list(residual)} to {shape_list(hidden_states)}",
+        )
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        return hidden_states, self_attn_weights
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "self_attn", None) is not None:
+            with tf.name_scope(self.self_attn.name):
+                self.self_attn.build(None)
+        if getattr(self, "self_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.self_attn_layer_norm.name):
+                self.self_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.embed_dim])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.encoder_ffn_dim])
+        if getattr(self, "final_layer_norm", None) is not None:
+            with tf.name_scope(self.final_layer_norm.name):
+                self.final_layer_norm.build([None, None, self.embed_dim])
+
+
+# Copied from transformers.models.speech_to_text.modeling_tf_speech_to_text.TFSpeech2TextDecoderLayer with Speech2Text->Whisper
+class TFWhisperDecoderLayer(keras.layers.Layer):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+
+        self.self_attn = TFWhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="self_attn",
+            is_decoder=True,
+        )
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
+
+        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.encoder_attn = TFWhisperAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="encoder_attn",
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
+        self.fc1 = keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
+        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+        self.config = config
+
+    def call(
+        self,
+        hidden_states,
+        attention_mask: tf.Tensor | None = None,
+        encoder_hidden_states: tf.Tensor | None = None,
+        encoder_attention_mask: tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        cross_attn_layer_head_mask: tf.Tensor | None = None,
+        past_key_value: Tuple[tf.Tensor] | None = None,
+        training=False,
+    ) -> Tuple[tf.Tensor, tf.Tensor, Tuple[Tuple[tf.Tensor]]]:
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (`tf.Tensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`tf.Tensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(decoder_attention_heads,)`
+            cross_attn_layer_head_mask (`tf.Tensor`): mask for heads of the cross-attention module.
+                `(decoder_attention_heads,)`
+            past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            training=training,
+        )
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                training=training,
+            )
+            hidden_states = self.dropout(hidden_states, training=training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        return (
+            hidden_states,
+            self_attn_weights,
+            cross_attn_weights,
+            present_key_value,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "self_attn", None) is not None:
+            with tf.name_scope(self.self_attn.name):
+                self.self_attn.build(None)
+        if getattr(self, "self_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.self_attn_layer_norm.name):
+                self.self_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "encoder_attn", None) is not None:
+            with tf.name_scope(self.encoder_attn.name):
+                self.encoder_attn.build(None)
+        if getattr(self, "encoder_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.encoder_attn_layer_norm.name):
+                self.encoder_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.embed_dim])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.decoder_ffn_dim])
+        if getattr(self, "final_layer_norm", None) is not None:
+            with tf.name_scope(self.final_layer_norm.name):
+                self.final_layer_norm.build([None, None, self.embed_dim])
+
+
+class TFWhisperPreTrainedModel(TFPreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix = "model"
+    main_input_name = "input_features"
+
+    def _get_feat_extract_output_lengths(self, input_lengths: tf.Tensor) -> int:
+        """
+        Computes the output length of the convolutional layers
+        """
+        input_lengths = (input_lengths - 1) // 2 + 1
+
+        return input_lengths
+
+    @property
+    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
+        """
+        Dummy inputs to build the network.
+
+        Returns:
+            `Dict[str, tf.Tensor]`: The dummy inputs.
+        """
+        return {
+            self.main_input_name: tf.random.uniform(
+                [1, self.config.num_mel_bins, self.config.max_source_positions * 2 - 1], dtype=tf.float32
+            ),
+            "decoder_input_ids": tf.constant([[1, 3]], dtype=tf.int32),
+        }
+
+    @property
+    def input_signature(self):
+        return {
+            "input_features": tf.TensorSpec((None, self.config.num_mel_bins, None), tf.float32, name="input_features"),
+            "decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
+            "decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
+        }
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
+    as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`WhisperConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~TFPreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`tf.Tensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be obtained
+            by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.*
+            via the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the fbank features, padding and conversion into a
+            tensor of type `tf.Tensor`. See [`~WhisperFeatureExtractor.__call__`]
+        decoder_input_ids (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`SpeechToTextTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            SpeechToText uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should read
+            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(tf.Tensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        decoder_inputs_embeds (`tf.Tensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
+            input (see `past_key_values`). This is useful if you want more control over how to convert
+            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@keras_serializable
+class TFWhisperEncoder(keras.layers.Layer):
+    config_class = WhisperConfig
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`TFWhisperEncoderLayer`].
+
+    Args:
+        config: WhisperConfig
+        embed_tokens (TFWhisperEmbedding): output embedding
+    """
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.layerdrop = config.encoder_layerdrop
+
+        self.embed_dim = config.d_model
+        self.num_mel_bins = config.num_mel_bins
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(self.embed_dim) if config.scale_embedding else 1.0
+
+        # Padding is added in call() to match the PyTorch implementation
+        self.conv1 = keras.layers.Conv1D(self.embed_dim, kernel_size=3, strides=1, padding="valid", name="conv1")
+        self.conv2 = keras.layers.Conv1D(self.embed_dim, kernel_size=3, strides=2, padding="valid", name="conv2")
+
+        self.embed_positions = TFWhisperPositionalEmbedding(
+            num_positions=self.max_source_positions,
+            embedding_dim=self.embed_dim,
+            embedding_initializer=sinusoidal_embedding_init,
+            name="embed_positions",
+        )
+        self.embed_positions.trainable = False
+
+        self.encoder_layers = [TFWhisperEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
+        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+        self.dropout = keras.layers.Dropout(config.dropout)
+
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_features (`tf.Tensor` of shape `(batch_size, feature_size, sequence_length)`):
+                Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be
+                obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
+                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the fbank features,
+                padding and conversion into a tensor of type `tf.Tensor`. See [`~WhisperFeatureExtractor.__call__`]
+            head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # TF 2.0 layers can't use channels first format when running on CPU.
+        input_features = tf.transpose(input_features, perm=(0, 2, 1))
+        input_features = tf.pad(input_features, [[0, 0], [1, 1], [0, 0]])
+        inputs_embeds = keras.activations.gelu(self.conv1(input_features))
+        inputs_embeds = tf.pad(inputs_embeds, [[0, 0], [1, 1], [0, 0]])
+        inputs_embeds = keras.activations.gelu(self.conv2(inputs_embeds))
+        inputs_embeds = tf.transpose(inputs_embeds, perm=(0, 1, 2))
+
+        embed_pos = self.embed_positions(input_ids=tf.zeros((1, self.max_source_positions), dtype=tf.int32))
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(head_mask)[0],
+                len(self.encoder_layers),
+                message=(
+                    f"The head_mask should be specified for {len(self.encoder_layers)} layers, but it is for"
+                    f" {shape_list(head_mask)[0]}."
+                ),
+            )
+
+        for idx, encoder_layer in enumerate(self.encoder_layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):  # skip the layer
+                continue
+
+            hidden_states, attn = encoder_layer(
+                hidden_states,
+                None,
+                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                training=training,
+            )
+
+            if output_attentions:
+                all_attentions += (attn,)
+
+        hidden_states = self.layer_norm(hidden_states)
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "conv1", None) is not None:
+            with tf.name_scope(self.conv1.name):
+                self.conv1.build([None, None, self.num_mel_bins])
+        if getattr(self, "conv2", None) is not None:
+            with tf.name_scope(self.conv2.name):
+                self.conv2.build([None, None, self.embed_dim])
+        if getattr(self, "embed_positions", None) is not None:
+            with tf.name_scope(self.embed_positions.name):
+                self.embed_positions.build(None)
+        if getattr(self, "layer_norm", None) is not None:
+            with tf.name_scope(self.layer_norm.name):
+                self.layer_norm.build([None, None, self.config.d_model])
+        if getattr(self, "encoder_layers", None) is not None:
+            for layer in self.encoder_layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+@keras_serializable
+class TFWhisperDecoder(keras.layers.Layer):
+    config_class = WhisperConfig
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFWhisperDecoderLayer`]
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_target_positions
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = keras.layers.Embedding(
+            input_dim=config.vocab_size,
+            output_dim=config.d_model,
+            embeddings_initializer=keras.initializers.TruncatedNormal(stddev=self.config.init_std),
+            name="embed_tokens",
+        )
+        self.embed_positions = TFWhisperPositionalEmbedding(
+            self.max_target_positions, config.d_model, name="embed_positions"
+        )
+
+        self.decoder_layers = [TFWhisperDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
+
+        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        batch_size, seq_len = input_shape[0], input_shape[1]
+
+        combined_attention_mask = tf.cond(
+            tf.math.greater(seq_len, 1),
+            lambda: _make_causal_mask(input_shape, past_key_values_length=past_key_values_length),
+            lambda: _expand_mask(tf.ones((batch_size, seq_len + past_key_values_length)), tgt_len=seq_len),
+        )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, tgt_len=input_shape[-1])
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+        return combined_attention_mask
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        encoder_hidden_states=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+                range `[0, config.max_position_embeddings - 1]`.
+            encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
+                on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+                `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+                `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = tf.shape(input_ids)
+            input_ids = tf.reshape(input_ids, (-1, input_shape[-1]))
+        elif inputs_embeds is not None:
+            input_shape = tf.shape(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = tf.shape(past_key_values[0][0])[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        attention_mask = self._prepare_decoder_attention_mask(attention_mask, input_shape, past_key_values_length)
+
+        # embed positions
+        filled_past_positions = past_key_values_length if position_ids is None else position_ids[0, -1]
+        positions = self.embed_positions(input_ids, past_key_values_length=filled_past_positions)
+
+        hidden_states = inputs_embeds + positions
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask_name, attn_mask in [("head_mask", head_mask), ("cross_attn_head_mask", cross_attn_head_mask)]:
+            if attn_mask is not None:
+                tf.debugging.assert_equal(
+                    shape_list(attn_mask)[0],
+                    len(self.decoder_layers),
+                    message=(
+                        f"The {attn_mask_name} should be specified for {len(self.decoder_layers)} layers, but it is"
+                        f" for {shape_list(attn_mask)[0]}."
+                    ),
+                )
+
+        for idx, decoder_layer in enumerate(self.decoder_layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                cross_attn_layer_head_mask=(cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None),
+                past_key_value=past_key_value,
+                training=training,
+            )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return TFBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embed_tokens", None) is not None:
+            with tf.name_scope(self.embed_tokens.name):
+                self.embed_tokens.build(None)
+        if getattr(self, "embed_positions", None) is not None:
+            with tf.name_scope(self.embed_positions.name):
+                self.embed_positions.build(None)
+        if getattr(self, "layer_norm", None) is not None:
+            with tf.name_scope(self.layer_norm.name):
+                self.layer_norm.build([None, None, self.config.d_model])
+        if getattr(self, "decoder_layers", None) is not None:
+            for layer in self.decoder_layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+@keras_serializable
+class TFWhisperMainLayer(keras.layers.Layer):
+    config_class = WhisperConfig
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.encoder = TFWhisperEncoder(config, name="encoder")
+        self.decoder = TFWhisperDecoder(config, name="decoder")
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        decoder_position_ids=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+         ```python
+         >>> import tensorflow as tf
+         >>> from transformers import TFWhisperModel, AutoFeatureExtractor
+         >>> from datasets import load_dataset
+
+         >>> model = TFWhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="tf")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = tf.convert_to_tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                training=training,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a TFBaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, TFBaseModelOutput):
+            encoder_outputs = TFBaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "encoder", None) is not None:
+            with tf.name_scope(self.encoder.name):
+                self.encoder.build(None)
+        if getattr(self, "decoder", None) is not None:
+            with tf.name_scope(self.decoder.name):
+                self.decoder.build(None)
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class TFWhisperModel(TFWhisperPreTrainedModel):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(config, **kwargs)
+
+        self.model = TFWhisperMainLayer(config, name="model")
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.model.encoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    def decoder(self):
+        return self.model.decoder
+
+    def encoder(self):
+        return self.model.encoder
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features: TFModelInputType | None = None,
+        decoder_input_ids: np.ndarray | tf.Tensor | None = None,
+        decoder_attention_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_head_mask: np.ndarray | tf.Tensor | None = None,
+        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
+        encoder_outputs: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[Union[np.ndarray, tf.Tensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+
+         ```python
+         >>> import tensorflow as tf
+         >>> from transformers import TFWhisperModel, AutoFeatureExtractor
+         >>> from datasets import load_dataset
+
+         >>> model = TFWhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="tf")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = tf.convert_to_tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        outputs = self.model(
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            encoder_outputs=encoder_outputs,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        return outputs
+
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=output.last_hidden_state,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "model", None) is not None:
+            with tf.name_scope(self.model.name):
+                self.model.build(None)
+
+
+@add_start_docstrings(
+    "The Whisper Model with a language modeling head. Can be used for automatic speech recognition.",
+    WHISPER_START_DOCSTRING,
+)
+class TFWhisperForConditionalGeneration(TFWhisperPreTrainedModel, TFCausalLanguageModelingLoss):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.version",
+        r"decoder.version",
+        r"proj_out.weight",
+    ]
+    _keys_to_ignore_on_save = [
+        r"proj_out.weight",
+    ]
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.model = TFWhisperMainLayer(config, name="model")
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def get_output_embeddings(self):
+        return self.get_input_embeddings()
+
+    def set_output_embeddings(self, value):
+        self.set_input_embeddings(value)
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> keras.layers.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        return new_embeddings
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features: TFModelInputType | None = None,
+        decoder_input_ids: np.ndarray | tf.Tensor | None = None,
+        decoder_attention_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_head_mask: np.ndarray | tf.Tensor | None = None,
+        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
+        encoder_outputs: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[Union[np.ndarray, tf.Tensor]]] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqLMOutput]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
+            or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import AutoProcessor, TFWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="tf")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(input_features=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_features,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        decoder_last_hidden_state = outputs[0]
+        # Decoder and encoder embeddings are tied
+        lm_logits = tf.matmul(decoder_last_hidden_state, self.get_output_embeddings().weights, transpose_b=True)
+
+        loss = None if labels is None else self.hf_compute_loss(labels, lm_logits)
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSeq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def generate(
+        self,
+        inputs: Optional[tf.Tensor] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        logits_processor: Optional[TFLogitsProcessorList] = None,
+        seed: Optional[List[int]] = None,
+        return_timestamps: Optional[bool] = None,
+        task: Optional[str] = None,
+        language: Optional[str] = None,
+        is_multilingual: Optional[bool] = None,
+        prompt_ids: Optional[tf.Tensor] = None,
+        return_token_timestamps=None,
+        **kwargs,
+    ):
+        r"""
+        Generates sequences of token ids for models with a language modeling head.
+
+        <Tip warning={true}>
+
+        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
+        model's default generation configuration. You can override any `generation_config` by passing the corresponding
+        parameters to generate, e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
+
+        For an overview of generation strategies and code examples, check out the [following
+        guide](../generation_strategies).
+
+        </Tip>
+
+        Parameters:
+            inputs (`tf.Tensor` of varying shape depending on the modality, *optional*):
+                The sequence used as a prompt for the generation or as model inputs to the encoder. If unset the method
+                initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs` should of in
+                the format of `input_ids`. For encoder-decoder models *inputs* can represent any of `input_ids`,
+                `input_values`, `input_features`, or `pixel_values`.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            logits_processor (`LogitsProcessorList`, *optional*):
+                Custom logits processors that complement the default logits processors built from arguments and
+                generation config. If a logit processor is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            seed (`List[int]`, *optional*):
+                Random seed to control sampling, containing two integers, used when `do_sample` is `True`. See the
+                `seed` argument from stateless functions in `tf.random`.
+            return_timestamps (`bool`, *optional*):
+                Whether to return the timestamps with the text. This enables the `TFWhisperTimestampsLogitsProcessor`.
+            task (`str`, *optional*):
+                Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
+                will be updated accordingly.
+            language (`str`, *optional*):
+                Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. You can
+                find all the possible language tokens in the `model.generation_config.lang_to_id` dictionary.
+            is_multilingual (`bool`, *optional*):
+                Whether or not the model is multilingual.
+            prompt_ids (`tf.Tensor`, *optional*):
+                Rank-1 tensor of token IDs created by passing text to [`~WhisperProcessor.get_prompt_ids`] that is
+                provided as a prompt to each chunk. This can be used to provide or "prompt-engineer" a context for
+                transcription, e.g. custom vocabularies or proper nouns to make it more likely to predict those words
+                correctly. It cannot be used in conjunction with `decoder_start_token_id` as it overwrites this value.
+            return_token_timestamps (`bool`, *optional*):
+                Whether to return token-level timestamps with the text. This can be used with or without the
+                `return_timestamps` option. To get word-level timestamps, use the tokenizer to group the tokens into
+                words.
+            kwargs (`Dict[str, Any]`, *optional*):
+                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
+                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
+                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
+
+        Return:
+            [`~utils.ModelOutput`] or `tf.Tensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True` or when
+            `config.return_dict_in_generate=True`) or a `tf.Tensor`.
+
+                If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.TFGreedySearchDecoderOnlyOutput`],
+                    - [`~generation.TFSampleDecoderOnlyOutput`],
+                    - [`~generation.TFBeamSearchDecoderOnlyOutput`],
+                    - [`~generation.TFBeamSampleDecoderOnlyOutput`]
+
+                If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.TFGreedySearchEncoderDecoderOutput`],
+                    - [`~generation.TFSampleEncoderDecoderOutput`],
+                    - [`~generation.TFBeamSearchEncoderDecoderOutput`],
+                    - [`~generation.TFBeamSampleEncoderDecoderOutput`]
+
+        """
+        if generation_config is None:
+            generation_config = self.generation_config
+
+        if return_timestamps is not None:
+            if not hasattr(generation_config, "no_timestamps_token_id"):
+                raise ValueError(
+                    "You are trying to return timestamps, but the generation config is not properly set. "
+                    "Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`. "
+                    "For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
+                )
+
+            generation_config.return_timestamps = return_timestamps
+        else:
+            generation_config.return_timestamps = False
+
+        if language is not None:
+            language = language.lower()
+            generation_config.language = language
+        if task is not None:
+            generation_config.task = task
+
+        forced_decoder_ids = None
+
+        # Legacy code for backward compatibility
+        if hasattr(self.config, "forced_decoder_ids") and self.config.forced_decoder_ids is not None:
+            forced_decoder_ids = self.config.forced_decoder_ids
+        elif (
+            hasattr(self.generation_config, "forced_decoder_ids")
+            and self.generation_config.forced_decoder_ids is not None
+        ):
+            forced_decoder_ids = self.generation_config.forced_decoder_ids
+        else:
+            forced_decoder_ids = kwargs.get("forced_decoder_ids", None)
+
+        if task is not None or language is not None or (forced_decoder_ids is None and prompt_ids is not None):
+            forced_decoder_ids = []
+            if hasattr(generation_config, "language"):
+                if generation_config.language in generation_config.lang_to_id.keys():
+                    language_token = generation_config.language
+                elif generation_config.language in TO_LANGUAGE_CODE.keys():
+                    language_token = f"<|{TO_LANGUAGE_CODE[generation_config.language]}|>"
+                elif generation_config.language in TO_LANGUAGE_CODE.values():
+                    language_token = f"<|{generation_config.language}|>"
+                else:
+                    is_language_code = len(generation_config.language) == 2
+                    raise ValueError(
+                        f"Unsupported language: {generation_config.language}. Language should be one of:"
+                        f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
+                    )
+                if language_token not in generation_config.lang_to_id:
+                    raise ValueError(
+                        f"{language_token} is not supported by this specific model as it is not in the `generation_config.lang_to_id`."
+                        "(You should just add it to the generation config)"
+                    )
+                forced_decoder_ids.append((1, generation_config.lang_to_id[language_token]))
+            else:
+                forced_decoder_ids.append((1, None))  # automatically detect the language
+
+            if hasattr(generation_config, "task"):
+                if generation_config.task in TASK_IDS:
+                    forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
+                else:
+                    raise ValueError(
+                        f"The `{generation_config.task}`task is not supported. The task should be one of `{TASK_IDS}`"
+                    )
+            elif hasattr(generation_config, "task_to_id"):
+                forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))  # defaults to transcribe
+            if hasattr(generation_config, "no_timestamps_token_id") and not generation_config.return_timestamps:
+                idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
+                forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
+
+        if forced_decoder_ids is not None:
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        if prompt_ids is not None:
+            if kwargs.get("decoder_start_token_id") is not None:
+                raise ValueError(
+                    "When specifying `prompt_ids`, you cannot also specify `decoder_start_token_id` as it gets overwritten."
+                )
+            prompt_ids = prompt_ids.tolist()
+            decoder_start_token_id, *text_prompt_ids = prompt_ids
+            # Slicing the text prompt ids in a manner consistent with the OpenAI implementation
+            # to accommodate context space for the prefix (see https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/decoding.py#L599)
+            text_prompt_ids = text_prompt_ids[-self.config.max_length // 2 - 1 :]
+            # Set the decoder_start_token_id to <|startofprev|>
+            kwargs.update({"decoder_start_token_id": decoder_start_token_id})
+
+            # Update the max generation length to include the prompt
+            specified_max_length = kwargs.pop("max_new_tokens", None) or kwargs.pop("max_length", None)
+            default_max_length = generation_config.max_new_tokens or generation_config.max_length
+            non_prompt_max_length = specified_max_length or default_max_length
+            kwargs["max_new_tokens"] = non_prompt_max_length + len(text_prompt_ids)
+
+            # Reformat the forced_decoder_ids to incorporate the prompt
+            non_prompt_forced_decoder_ids = (
+                kwargs.pop("forced_decoder_ids", None) or generation_config.forced_decoder_ids
+            )
+            forced_decoder_ids = [
+                *text_prompt_ids,
+                generation_config.decoder_start_token_id,
+                *[token for _rank, token in non_prompt_forced_decoder_ids],
+            ]
+            forced_decoder_ids = [(rank + 1, token) for rank, token in enumerate(forced_decoder_ids)]
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        # TODO: Implement `WhisperTimeStampLogitsProcessor`.
+        if generation_config.return_timestamps:
+            # logits_processor = [TFWhisperTimeStampLogitsProcessor(generation_config)]
+            raise ValueError("`TFWhisperForConditionalGeneration` doesn't support returning the timestamps yet.")
+
+        if return_token_timestamps:
+            kwargs["output_attentions"] = True
+            kwargs["return_dict_in_generate"] = True
+
+            if getattr(generation_config, "task", None) == "translate":
+                logger.warning("Token-level timestamps may not be reliable for task 'translate'.")
+            if not hasattr(generation_config, "alignment_heads"):
+                raise ValueError(
+                    "Model generation config has no `alignment_heads`, token-level timestamps not available. "
+                    "See https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a on how to add this property to the generation config."
+                )
+
+        outputs = super().generate(
+            inputs,
+            generation_config,
+            logits_processor,
+            **kwargs,
+        )
+
+        if return_token_timestamps and hasattr(generation_config, "alignment_heads"):
+            outputs["token_timestamps"] = self._extract_token_timestamps(outputs, generation_config.alignment_heads)
+
+        return outputs
+
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqLMOutput(
+            logits=output.logits,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        past_key_values=None,
+        use_cache=None,
+        encoder_outputs=None,
+        attention_mask=None,
+        decoder_attention_mask=None,
+        **kwargs,
+    ):
+        # cut decoder_input_ids if past is used
+        if past_key_values is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        if decoder_attention_mask is not None:  # xla
+            decoder_position_ids = tf.math.cumsum(decoder_attention_mask, axis=-1, exclusive=True)[:, -1:]
+        elif past_key_values is not None:  # no xla + past
+            decoder_position_ids = past_key_values[0][0].shape[2]
+        else:  # no xla + no past
+            decoder_position_ids = tf.range(decoder_input_ids.shape[1])
+        decoder_position_ids = tf.broadcast_to(decoder_position_ids, decoder_input_ids.shape)
+
+        return {
+            "input_features": None,  # Needs to be passed to make Keras.layer.__call__ happy
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past_key_values,
+            "decoder_input_ids": decoder_input_ids,
+            "use_cache": use_cache,
+            "decoder_attention_mask": decoder_attention_mask,
+            "decoder_position_ids": decoder_position_ids,
+        }
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "model", None) is not None:
+            with tf.name_scope(self.model.name):
+                self.model.build(None)

modeling_tf_whisper.py

@@ -0,0 +1,1758 @@
+# coding=utf-8
+# Copyright 2022 The OpenAI Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""TensorFlow Whisper model."""
+
+from __future__ import annotations
+
+import math
+import random
+from typing import Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...generation.configuration_utils import GenerationConfig
+from ...generation.tf_logits_process import TFLogitsProcessorList
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPastAndCrossAttentions,
+    TFSeq2SeqLMOutput,
+    TFSeq2SeqModelOutput,
+)
+from ...modeling_tf_utils import (
+    TFCausalLanguageModelingLoss,
+    TFModelInputType,
+    TFPreTrainedModel,
+    keras,
+    keras_serializable,
+    unpack_inputs,
+)
+from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_whisper import WhisperConfig
+from .tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "WhisperConfig"
+
+
+LARGE_NEGATIVE = -1e8
+
+
+def sinusoidal_embedding_init(shape, dtype=tf.float32) -> tf.Tensor:
+    """Returns sinusoids for positional embedding"""
+    length, channels = shape
+    if channels % 2 != 0:
+        raise ValueError(
+            f"Number of channels has to be divisible by 2 for sinusoidal positional embeddings, got {channels} channels."
+        )
+    log_timescale_increment = math.log(10000) / (channels // 2 - 1)
+    inv_timescales = tf.exp(-log_timescale_increment * tf.range(channels // 2, dtype=tf.float32))
+    scaled_time = tf.reshape(tf.range(length, dtype=tf.float32), (-1, 1)) * tf.reshape(inv_timescales, (1, -1))
+    return tf.cast(tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=1), dtype)
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.shift_tokens_right
+def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    pad_token_id = tf.cast(pad_token_id, input_ids.dtype)
+    decoder_start_token_id = tf.cast(decoder_start_token_id, input_ids.dtype)
+    start_tokens = tf.fill(
+        (shape_list(input_ids)[0], 1), tf.convert_to_tensor(decoder_start_token_id, input_ids.dtype)
+    )
+    shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids = tf.where(
+        shifted_input_ids == -100,
+        tf.fill(shape_list(shifted_input_ids), tf.convert_to_tensor(pad_token_id, input_ids.dtype)),
+        shifted_input_ids,
+    )
+
+    # "Verify that `labels` has only positive values and -100"
+    assert_gte0 = tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=input_ids.dtype))
+
+    # Make sure the assertion op is called by wrapping the result in an identity no-op
+    with tf.control_dependencies([assert_gte0]):
+        shifted_input_ids = tf.identity(shifted_input_ids)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz = input_ids_shape[0]
+    tgt_len = input_ids_shape[1]
+    mask = tf.ones((tgt_len, tgt_len)) * LARGE_NEGATIVE
+    mask_cond = tf.range(shape_list(mask)[-1])
+
+    mask = tf.where(mask_cond < tf.reshape(mask_cond + 1, (shape_list(mask)[-1], 1)), 0.0, mask)
+
+    if past_key_values_length > 0:
+        mask = tf.concat([tf.zeros((tgt_len, past_key_values_length)), mask], axis=-1)
+
+    return tf.tile(mask[None, None, :, :], (bsz, 1, 1, 1))
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._expand_mask
+def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    src_len = shape_list(mask)[1]
+    tgt_len = tgt_len if tgt_len is not None else src_len
+    one_cst = tf.constant(1.0)
+    mask = tf.cast(mask, dtype=one_cst.dtype)
+    expanded_mask = tf.tile(mask[:, None, None, :], (1, 1, tgt_len, 1))
+
+    return (one_cst - expanded_mask) * LARGE_NEGATIVE
+
+
+class TFWhisperPositionalEmbedding(keras.layers.Layer):
+    def __init__(
+        self,
+        num_positions: int,
+        embedding_dim: int,
+        padding_idx: Optional[int] = None,
+        embedding_initializer=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_positions = num_positions
+        self.embedding_dim = embedding_dim
+        self.padding_idx = padding_idx
+        self.embedding_initializer = keras.initializers.get(embedding_initializer)
+
+    def build(self, input_shape):
+        self.weight = self.add_weight(
+            name="weight",
+            shape=[self.num_positions, self.embedding_dim],
+            initializer=self.embedding_initializer,
+            trainable=True,
+        )
+        super().build(input_shape)
+
+    def call(self, input_ids, past_key_values_length=0):
+        past_key_values_length = tf.cast(past_key_values_length, tf.int32)
+        gather_indices = tf.range(tf.shape(input_ids)[1], delta=1) + past_key_values_length
+        return tf.gather(self.weight, gather_indices)
+
+
+class TFWhisperAttention(keras.layers.Layer):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = keras.layers.Dropout(dropout)
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = keras.layers.Dense(embed_dim, use_bias=False, name="k_proj")
+        self.v_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
+        self.q_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
+        self.out_proj = keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention._shape with BART->whisper
+    def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int):
+        return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), (0, 2, 1, 3))
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention.call with BART->whisper
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        key_value_states: tf.Tensor | None = None,
+        past_key_value: Tuple[Tuple[tf.Tensor]] | None = None,
+        attention_mask: tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        training: Optional[bool] = False,
+    ) -> Tuple[tf.Tensor, tf.Tensor | None]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = shape_list(hidden_states)
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = tf.concat([past_key_value[0], key_states], axis=2)
+            value_states = tf.concat([past_key_value[1], value_states], axis=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(tf.Tensor, tf.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(tf.Tensor, tf.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = tf.reshape(self._shape(query_states, tgt_len, bsz), proj_shape)
+        key_states = tf.reshape(key_states, proj_shape)
+        value_states = tf.reshape(value_states, proj_shape)
+
+        src_len = shape_list(key_states)[1]
+        attn_weights = tf.matmul(query_states, key_states, transpose_b=True)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_weights),
+            [bsz * self.num_heads, tgt_len, src_len],
+            message=(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {shape_list(attn_weights)}"
+            ),
+        )
+
+        if attention_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(attention_mask),
+                [bsz, 1, tgt_len, src_len],
+                message=(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {shape_list(attention_mask)}"
+                ),
+            )
+
+            attention_mask = tf.cast(attention_mask, dtype=attn_weights.dtype)
+            attn_weights = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len)) + attention_mask
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_weights = stable_softmax(attn_weights, axis=-1)
+
+        if layer_head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(layer_head_mask),
+                [self.num_heads],
+                message=(
+                    f"Head mask for a single layer should be of size {(self.num_heads)}, but is"
+                    f" {shape_list(layer_head_mask)}"
+                ),
+            )
+
+            attn_weights = tf.reshape(layer_head_mask, (1, -1, 1, 1)) * tf.reshape(
+                attn_weights, (bsz, self.num_heads, tgt_len, src_len)
+            )
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_probs = self.dropout(attn_weights, training=training)
+        attn_output = tf.matmul(attn_probs, value_states)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_output),
+            [bsz * self.num_heads, tgt_len, self.head_dim],
+            message=(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {shape_list(attn_output)}"
+            ),
+        )
+
+        attn_output = tf.transpose(
+            tf.reshape(attn_output, (bsz, self.num_heads, tgt_len, self.head_dim)), (0, 2, 1, 3)
+        )
+        attn_output = tf.reshape(attn_output, (bsz, tgt_len, embed_dim))
+
+        attn_output = self.out_proj(attn_output)
+        attn_weights: tf.Tensor = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len))
+
+        return attn_output, attn_weights, past_key_value
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "k_proj", None) is not None:
+            with tf.name_scope(self.k_proj.name):
+                self.k_proj.build([None, None, self.embed_dim])
+        if getattr(self, "v_proj", None) is not None:
+            with tf.name_scope(self.v_proj.name):
+                self.v_proj.build([None, None, self.embed_dim])
+        if getattr(self, "q_proj", None) is not None:
+            with tf.name_scope(self.q_proj.name):
+                self.q_proj.build([None, None, self.embed_dim])
+        if getattr(self, "out_proj", None) is not None:
+            with tf.name_scope(self.out_proj.name):
+                self.out_proj.build([None, None, self.embed_dim])
+
+
+# Copied from transformers.models.speech_to_text.modeling_tf_speech_to_text.TFSpeech2TextEncoderLayer with Speech2Text->Whisper
+class TFWhisperEncoderLayer(keras.layers.Layer):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFWhisperAttention(
+            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
+        )
+        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
+        self.fc1 = keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
+        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+        self.config = config
+
+    def call(
+        self, hidden_states: tf.Tensor, attention_mask: tf.Tensor, layer_head_mask: tf.Tensor, training: bool = False
+    ):
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, self_attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            training=training,
+        )
+
+        tf.debugging.assert_equal(
+            shape_list(hidden_states),
+            shape_list(residual),
+            message=f"Self attn modified the shape of query {shape_list(residual)} to {shape_list(hidden_states)}",
+        )
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        return hidden_states, self_attn_weights
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "self_attn", None) is not None:
+            with tf.name_scope(self.self_attn.name):
+                self.self_attn.build(None)
+        if getattr(self, "self_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.self_attn_layer_norm.name):
+                self.self_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.embed_dim])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.encoder_ffn_dim])
+        if getattr(self, "final_layer_norm", None) is not None:
+            with tf.name_scope(self.final_layer_norm.name):
+                self.final_layer_norm.build([None, None, self.embed_dim])
+
+
+# Copied from transformers.models.speech_to_text.modeling_tf_speech_to_text.TFSpeech2TextDecoderLayer with Speech2Text->Whisper
+class TFWhisperDecoderLayer(keras.layers.Layer):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+
+        self.self_attn = TFWhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="self_attn",
+            is_decoder=True,
+        )
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = keras.layers.Dropout(config.activation_dropout)
+
+        self.self_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.encoder_attn = TFWhisperAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="encoder_attn",
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
+        self.fc1 = keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
+        self.fc2 = keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+        self.config = config
+
+    def call(
+        self,
+        hidden_states,
+        attention_mask: tf.Tensor | None = None,
+        encoder_hidden_states: tf.Tensor | None = None,
+        encoder_attention_mask: tf.Tensor | None = None,
+        layer_head_mask: tf.Tensor | None = None,
+        cross_attn_layer_head_mask: tf.Tensor | None = None,
+        past_key_value: Tuple[tf.Tensor] | None = None,
+        training=False,
+    ) -> Tuple[tf.Tensor, tf.Tensor, Tuple[Tuple[tf.Tensor]]]:
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (`tf.Tensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`tf.Tensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(decoder_attention_heads,)`
+            cross_attn_layer_head_mask (`tf.Tensor`): mask for heads of the cross-attention module.
+                `(decoder_attention_heads,)`
+            past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            training=training,
+        )
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                training=training,
+            )
+            hidden_states = self.dropout(hidden_states, training=training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        return (
+            hidden_states,
+            self_attn_weights,
+            cross_attn_weights,
+            present_key_value,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "self_attn", None) is not None:
+            with tf.name_scope(self.self_attn.name):
+                self.self_attn.build(None)
+        if getattr(self, "self_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.self_attn_layer_norm.name):
+                self.self_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "encoder_attn", None) is not None:
+            with tf.name_scope(self.encoder_attn.name):
+                self.encoder_attn.build(None)
+        if getattr(self, "encoder_attn_layer_norm", None) is not None:
+            with tf.name_scope(self.encoder_attn_layer_norm.name):
+                self.encoder_attn_layer_norm.build([None, None, self.embed_dim])
+        if getattr(self, "fc1", None) is not None:
+            with tf.name_scope(self.fc1.name):
+                self.fc1.build([None, None, self.embed_dim])
+        if getattr(self, "fc2", None) is not None:
+            with tf.name_scope(self.fc2.name):
+                self.fc2.build([None, None, self.config.decoder_ffn_dim])
+        if getattr(self, "final_layer_norm", None) is not None:
+            with tf.name_scope(self.final_layer_norm.name):
+                self.final_layer_norm.build([None, None, self.embed_dim])
+
+
+class TFWhisperPreTrainedModel(TFPreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix = "model"
+    main_input_name = "input_features"
+
+    def _get_feat_extract_output_lengths(self, input_lengths: tf.Tensor) -> int:
+        """
+        Computes the output length of the convolutional layers
+        """
+        input_lengths = (input_lengths - 1) // 2 + 1
+
+        return input_lengths
+
+    @property
+    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
+        """
+        Dummy inputs to build the network.
+
+        Returns:
+            `Dict[str, tf.Tensor]`: The dummy inputs.
+        """
+        return {
+            self.main_input_name: tf.random.uniform(
+                [1, self.config.num_mel_bins, self.config.max_source_positions * 2 - 1], dtype=tf.float32
+            ),
+            "decoder_input_ids": tf.constant([[1, 3]], dtype=tf.int32),
+        }
+
+    @property
+    def input_signature(self):
+        return {
+            "input_features": tf.TensorSpec((None, self.config.num_mel_bins, None), tf.float32, name="input_features"),
+            "decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
+            "decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
+        }
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
+    as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`WhisperConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~TFPreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`tf.Tensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be obtained
+            by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.*
+            via the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the fbank features, padding and conversion into a
+            tensor of type `tf.Tensor`. See [`~WhisperFeatureExtractor.__call__`]
+        decoder_input_ids (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`SpeechToTextTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            SpeechToText uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should read
+            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(tf.Tensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        decoder_inputs_embeds (`tf.Tensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
+            input (see `past_key_values`). This is useful if you want more control over how to convert
+            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@keras_serializable
+class TFWhisperEncoder(keras.layers.Layer):
+    config_class = WhisperConfig
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`TFWhisperEncoderLayer`].
+
+    Args:
+        config: WhisperConfig
+        embed_tokens (TFWhisperEmbedding): output embedding
+    """
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.layerdrop = config.encoder_layerdrop
+
+        self.embed_dim = config.d_model
+        self.num_mel_bins = config.num_mel_bins
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(self.embed_dim) if config.scale_embedding else 1.0
+
+        # Padding is added in call() to match the PyTorch implementation
+        self.conv1 = keras.layers.Conv1D(self.embed_dim, kernel_size=3, strides=1, padding="valid", name="conv1")
+        self.conv2 = keras.layers.Conv1D(self.embed_dim, kernel_size=3, strides=2, padding="valid", name="conv2")
+
+        self.embed_positions = TFWhisperPositionalEmbedding(
+            num_positions=self.max_source_positions,
+            embedding_dim=self.embed_dim,
+            embedding_initializer=sinusoidal_embedding_init,
+            name="embed_positions",
+        )
+        self.embed_positions.trainable = False
+
+        self.encoder_layers = [TFWhisperEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
+        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+        self.dropout = keras.layers.Dropout(config.dropout)
+
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_features (`tf.Tensor` of shape `(batch_size, feature_size, sequence_length)`):
+                Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be
+                obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
+                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the fbank features,
+                padding and conversion into a tensor of type `tf.Tensor`. See [`~WhisperFeatureExtractor.__call__`]
+            head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # TF 2.0 layers can't use channels first format when running on CPU.
+        input_features = tf.transpose(input_features, perm=(0, 2, 1))
+        input_features = tf.pad(input_features, [[0, 0], [1, 1], [0, 0]])
+        inputs_embeds = keras.activations.gelu(self.conv1(input_features))
+        inputs_embeds = tf.pad(inputs_embeds, [[0, 0], [1, 1], [0, 0]])
+        inputs_embeds = keras.activations.gelu(self.conv2(inputs_embeds))
+        inputs_embeds = tf.transpose(inputs_embeds, perm=(0, 1, 2))
+
+        embed_pos = self.embed_positions(input_ids=tf.zeros((1, self.max_source_positions), dtype=tf.int32))
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(head_mask)[0],
+                len(self.encoder_layers),
+                message=(
+                    f"The head_mask should be specified for {len(self.encoder_layers)} layers, but it is for"
+                    f" {shape_list(head_mask)[0]}."
+                ),
+            )
+
+        for idx, encoder_layer in enumerate(self.encoder_layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):  # skip the layer
+                continue
+
+            hidden_states, attn = encoder_layer(
+                hidden_states,
+                None,
+                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                training=training,
+            )
+
+            if output_attentions:
+                all_attentions += (attn,)
+
+        hidden_states = self.layer_norm(hidden_states)
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "conv1", None) is not None:
+            with tf.name_scope(self.conv1.name):
+                self.conv1.build([None, None, self.num_mel_bins])
+        if getattr(self, "conv2", None) is not None:
+            with tf.name_scope(self.conv2.name):
+                self.conv2.build([None, None, self.embed_dim])
+        if getattr(self, "embed_positions", None) is not None:
+            with tf.name_scope(self.embed_positions.name):
+                self.embed_positions.build(None)
+        if getattr(self, "layer_norm", None) is not None:
+            with tf.name_scope(self.layer_norm.name):
+                self.layer_norm.build([None, None, self.config.d_model])
+        if getattr(self, "encoder_layers", None) is not None:
+            for layer in self.encoder_layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+@keras_serializable
+class TFWhisperDecoder(keras.layers.Layer):
+    config_class = WhisperConfig
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFWhisperDecoderLayer`]
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.dropout = keras.layers.Dropout(config.dropout)
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_target_positions
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = keras.layers.Embedding(
+            input_dim=config.vocab_size,
+            output_dim=config.d_model,
+            embeddings_initializer=keras.initializers.TruncatedNormal(stddev=self.config.init_std),
+            name="embed_tokens",
+        )
+        self.embed_positions = TFWhisperPositionalEmbedding(
+            self.max_target_positions, config.d_model, name="embed_positions"
+        )
+
+        self.decoder_layers = [TFWhisperDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
+
+        self.layer_norm = keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        batch_size, seq_len = input_shape[0], input_shape[1]
+
+        combined_attention_mask = tf.cond(
+            tf.math.greater(seq_len, 1),
+            lambda: _make_causal_mask(input_shape, past_key_values_length=past_key_values_length),
+            lambda: _expand_mask(tf.ones((batch_size, seq_len + past_key_values_length)), tgt_len=seq_len),
+        )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, tgt_len=input_shape[-1])
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+        return combined_attention_mask
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        encoder_hidden_states=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+                range `[0, config.max_position_embeddings - 1]`.
+            encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
+                on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+                `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+                `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = tf.shape(input_ids)
+            input_ids = tf.reshape(input_ids, (-1, input_shape[-1]))
+        elif inputs_embeds is not None:
+            input_shape = tf.shape(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = tf.shape(past_key_values[0][0])[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            check_embeddings_within_bounds(input_ids, self.embed_tokens.input_dim)
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        attention_mask = self._prepare_decoder_attention_mask(attention_mask, input_shape, past_key_values_length)
+
+        # embed positions
+        filled_past_positions = past_key_values_length if position_ids is None else position_ids[0, -1]
+        positions = self.embed_positions(input_ids, past_key_values_length=filled_past_positions)
+
+        hidden_states = inputs_embeds + positions
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask_name, attn_mask in [("head_mask", head_mask), ("cross_attn_head_mask", cross_attn_head_mask)]:
+            if attn_mask is not None:
+                tf.debugging.assert_equal(
+                    shape_list(attn_mask)[0],
+                    len(self.decoder_layers),
+                    message=(
+                        f"The {attn_mask_name} should be specified for {len(self.decoder_layers)} layers, but it is"
+                        f" for {shape_list(attn_mask)[0]}."
+                    ),
+                )
+
+        for idx, decoder_layer in enumerate(self.decoder_layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                cross_attn_layer_head_mask=(cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None),
+                past_key_value=past_key_value,
+                training=training,
+            )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return TFBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "embed_tokens", None) is not None:
+            with tf.name_scope(self.embed_tokens.name):
+                self.embed_tokens.build(None)
+        if getattr(self, "embed_positions", None) is not None:
+            with tf.name_scope(self.embed_positions.name):
+                self.embed_positions.build(None)
+        if getattr(self, "layer_norm", None) is not None:
+            with tf.name_scope(self.layer_norm.name):
+                self.layer_norm.build([None, None, self.config.d_model])
+        if getattr(self, "decoder_layers", None) is not None:
+            for layer in self.decoder_layers:
+                with tf.name_scope(layer.name):
+                    layer.build(None)
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+@keras_serializable
+class TFWhisperMainLayer(keras.layers.Layer):
+    config_class = WhisperConfig
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.encoder = TFWhisperEncoder(config, name="encoder")
+        self.decoder = TFWhisperDecoder(config, name="decoder")
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        decoder_position_ids=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+         ```python
+         >>> import tensorflow as tf
+         >>> from transformers import TFWhisperModel, AutoFeatureExtractor
+         >>> from datasets import load_dataset
+
+         >>> model = TFWhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="tf")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = tf.convert_to_tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                training=training,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a TFBaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, TFBaseModelOutput):
+            encoder_outputs = TFBaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "encoder", None) is not None:
+            with tf.name_scope(self.encoder.name):
+                self.encoder.build(None)
+        if getattr(self, "decoder", None) is not None:
+            with tf.name_scope(self.decoder.name):
+                self.decoder.build(None)
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class TFWhisperModel(TFWhisperPreTrainedModel):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(config, **kwargs)
+
+        self.model = TFWhisperMainLayer(config, name="model")
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.model.encoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    def decoder(self):
+        return self.model.decoder
+
+    def encoder(self):
+        return self.model.encoder
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features: TFModelInputType | None = None,
+        decoder_input_ids: np.ndarray | tf.Tensor | None = None,
+        decoder_attention_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_head_mask: np.ndarray | tf.Tensor | None = None,
+        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
+        encoder_outputs: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[Union[np.ndarray, tf.Tensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+
+         ```python
+         >>> import tensorflow as tf
+         >>> from transformers import TFWhisperModel, AutoFeatureExtractor
+         >>> from datasets import load_dataset
+
+         >>> model = TFWhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="tf")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = tf.convert_to_tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        outputs = self.model(
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            encoder_outputs=encoder_outputs,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        return outputs
+
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=output.last_hidden_state,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "model", None) is not None:
+            with tf.name_scope(self.model.name):
+                self.model.build(None)
+
+
+@add_start_docstrings(
+    "The Whisper Model with a language modeling head. Can be used for automatic speech recognition.",
+    WHISPER_START_DOCSTRING,
+)
+class TFWhisperForConditionalGeneration(TFWhisperPreTrainedModel, TFCausalLanguageModelingLoss):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.version",
+        r"decoder.version",
+        r"proj_out.weight",
+    ]
+    _keys_to_ignore_on_save = [
+        r"proj_out.weight",
+    ]
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.model = TFWhisperMainLayer(config, name="model")
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def get_output_embeddings(self):
+        return self.get_input_embeddings()
+
+    def set_output_embeddings(self, value):
+        self.set_input_embeddings(value)
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> keras.layers.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        return new_embeddings
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features: TFModelInputType | None = None,
+        decoder_input_ids: np.ndarray | tf.Tensor | None = None,
+        decoder_attention_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_position_ids: np.ndarray | tf.Tensor | None = None,
+        head_mask: np.ndarray | tf.Tensor | None = None,
+        decoder_head_mask: np.ndarray | tf.Tensor | None = None,
+        cross_attn_head_mask: np.ndarray | tf.Tensor | None = None,
+        encoder_outputs: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[Union[np.ndarray, tf.Tensor]]] = None,
+        labels: np.ndarray | tf.Tensor | None = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[Tuple[tf.Tensor], TFSeq2SeqLMOutput]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
+            or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import AutoProcessor, TFWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="tf")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(input_features=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_features,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        decoder_last_hidden_state = outputs[0]
+        # Decoder and encoder embeddings are tied
+        lm_logits = tf.matmul(decoder_last_hidden_state, self.get_output_embeddings().weights, transpose_b=True)
+
+        loss = None if labels is None else self.hf_compute_loss(labels, lm_logits)
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSeq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def generate(
+        self,
+        inputs: Optional[tf.Tensor] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        logits_processor: Optional[TFLogitsProcessorList] = None,
+        seed: Optional[List[int]] = None,
+        return_timestamps: Optional[bool] = None,
+        task: Optional[str] = None,
+        language: Optional[str] = None,
+        is_multilingual: Optional[bool] = None,
+        prompt_ids: Optional[tf.Tensor] = None,
+        return_token_timestamps=None,
+        **kwargs,
+    ):
+        r"""
+        Generates sequences of token ids for models with a language modeling head.
+
+        <Tip warning={true}>
+
+        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
+        model's default generation configuration. You can override any `generation_config` by passing the corresponding
+        parameters to generate, e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
+
+        For an overview of generation strategies and code examples, check out the [following
+        guide](../generation_strategies).
+
+        </Tip>
+
+        Parameters:
+            inputs (`tf.Tensor` of varying shape depending on the modality, *optional*):
+                The sequence used as a prompt for the generation or as model inputs to the encoder. If unset the method
+                initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs` should of in
+                the format of `input_ids`. For encoder-decoder models *inputs* can represent any of `input_ids`,
+                `input_values`, `input_features`, or `pixel_values`.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            logits_processor (`LogitsProcessorList`, *optional*):
+                Custom logits processors that complement the default logits processors built from arguments and
+                generation config. If a logit processor is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            seed (`List[int]`, *optional*):
+                Random seed to control sampling, containing two integers, used when `do_sample` is `True`. See the
+                `seed` argument from stateless functions in `tf.random`.
+            return_timestamps (`bool`, *optional*):
+                Whether to return the timestamps with the text. This enables the `TFWhisperTimestampsLogitsProcessor`.
+            task (`str`, *optional*):
+                Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
+                will be updated accordingly.
+            language (`str`, *optional*):
+                Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. You can
+                find all the possible language tokens in the `model.generation_config.lang_to_id` dictionary.
+            is_multilingual (`bool`, *optional*):
+                Whether or not the model is multilingual.
+            prompt_ids (`tf.Tensor`, *optional*):
+                Rank-1 tensor of token IDs created by passing text to [`~WhisperProcessor.get_prompt_ids`] that is
+                provided as a prompt to each chunk. This can be used to provide or "prompt-engineer" a context for
+                transcription, e.g. custom vocabularies or proper nouns to make it more likely to predict those words
+                correctly. It cannot be used in conjunction with `decoder_start_token_id` as it overwrites this value.
+            return_token_timestamps (`bool`, *optional*):
+                Whether to return token-level timestamps with the text. This can be used with or without the
+                `return_timestamps` option. To get word-level timestamps, use the tokenizer to group the tokens into
+                words.
+            kwargs (`Dict[str, Any]`, *optional*):
+                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
+                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
+                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
+
+        Return:
+            [`~utils.ModelOutput`] or `tf.Tensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True` or when
+            `config.return_dict_in_generate=True`) or a `tf.Tensor`.
+
+                If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.TFGreedySearchDecoderOnlyOutput`],
+                    - [`~generation.TFSampleDecoderOnlyOutput`],
+                    - [`~generation.TFBeamSearchDecoderOnlyOutput`],
+                    - [`~generation.TFBeamSampleDecoderOnlyOutput`]
+
+                If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.TFGreedySearchEncoderDecoderOutput`],
+                    - [`~generation.TFSampleEncoderDecoderOutput`],
+                    - [`~generation.TFBeamSearchEncoderDecoderOutput`],
+                    - [`~generation.TFBeamSampleEncoderDecoderOutput`]
+
+        """
+        if generation_config is None:
+            generation_config = self.generation_config
+
+        if return_timestamps is not None:
+            if not hasattr(generation_config, "no_timestamps_token_id"):
+                raise ValueError(
+                    "You are trying to return timestamps, but the generation config is not properly set. "
+                    "Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`. "
+                    "For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
+                )
+
+            generation_config.return_timestamps = return_timestamps
+        else:
+            generation_config.return_timestamps = False
+
+        if language is not None:
+            language = language.lower()
+            generation_config.language = language
+        if task is not None:
+            generation_config.task = task
+
+        forced_decoder_ids = None
+
+        # Legacy code for backward compatibility
+        if hasattr(self.config, "forced_decoder_ids") and self.config.forced_decoder_ids is not None:
+            forced_decoder_ids = self.config.forced_decoder_ids
+        elif (
+            hasattr(self.generation_config, "forced_decoder_ids")
+            and self.generation_config.forced_decoder_ids is not None
+        ):
+            forced_decoder_ids = self.generation_config.forced_decoder_ids
+        else:
+            forced_decoder_ids = kwargs.get("forced_decoder_ids", None)
+
+        if task is not None or language is not None or (forced_decoder_ids is None and prompt_ids is not None):
+            forced_decoder_ids = []
+            if hasattr(generation_config, "language"):
+                if generation_config.language in generation_config.lang_to_id.keys():
+                    language_token = generation_config.language
+                elif generation_config.language in TO_LANGUAGE_CODE.keys():
+                    language_token = f"<|{TO_LANGUAGE_CODE[generation_config.language]}|>"
+                elif generation_config.language in TO_LANGUAGE_CODE.values():
+                    language_token = f"<|{generation_config.language}|>"
+                else:
+                    is_language_code = len(generation_config.language) == 2
+                    raise ValueError(
+                        f"Unsupported language: {generation_config.language}. Language should be one of:"
+                        f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
+                    )
+                if language_token not in generation_config.lang_to_id:
+                    raise ValueError(
+                        f"{language_token} is not supported by this specific model as it is not in the `generation_config.lang_to_id`."
+                        "(You should just add it to the generation config)"
+                    )
+                forced_decoder_ids.append((1, generation_config.lang_to_id[language_token]))
+            else:
+                forced_decoder_ids.append((1, None))  # automatically detect the language
+
+            if hasattr(generation_config, "task"):
+                if generation_config.task in TASK_IDS:
+                    forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
+                else:
+                    raise ValueError(
+                        f"The `{generation_config.task}`task is not supported. The task should be one of `{TASK_IDS}`"
+                    )
+            elif hasattr(generation_config, "task_to_id"):
+                forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))  # defaults to transcribe
+            if hasattr(generation_config, "no_timestamps_token_id") and not generation_config.return_timestamps:
+                idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
+                forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
+
+        if forced_decoder_ids is not None:
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        if prompt_ids is not None:
+            if kwargs.get("decoder_start_token_id") is not None:
+                raise ValueError(
+                    "When specifying `prompt_ids`, you cannot also specify `decoder_start_token_id` as it gets overwritten."
+                )
+            prompt_ids = prompt_ids.tolist()
+            decoder_start_token_id, *text_prompt_ids = prompt_ids
+            # Slicing the text prompt ids in a manner consistent with the OpenAI implementation
+            # to accommodate context space for the prefix (see https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/decoding.py#L599)
+            text_prompt_ids = text_prompt_ids[-self.config.max_length // 2 - 1 :]
+            # Set the decoder_start_token_id to <|startofprev|>
+            kwargs.update({"decoder_start_token_id": decoder_start_token_id})
+
+            # Update the max generation length to include the prompt
+            specified_max_length = kwargs.pop("max_new_tokens", None) or kwargs.pop("max_length", None)
+            default_max_length = generation_config.max_new_tokens or generation_config.max_length
+            non_prompt_max_length = specified_max_length or default_max_length
+            kwargs["max_new_tokens"] = non_prompt_max_length + len(text_prompt_ids)
+
+            # Reformat the forced_decoder_ids to incorporate the prompt
+            non_prompt_forced_decoder_ids = (
+                kwargs.pop("forced_decoder_ids", None) or generation_config.forced_decoder_ids
+            )
+            forced_decoder_ids = [
+                *text_prompt_ids,
+                generation_config.decoder_start_token_id,
+                *[token for _rank, token in non_prompt_forced_decoder_ids],
+            ]
+            forced_decoder_ids = [(rank + 1, token) for rank, token in enumerate(forced_decoder_ids)]
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        # TODO: Implement `WhisperTimeStampLogitsProcessor`.
+        if generation_config.return_timestamps:
+            # logits_processor = [TFWhisperTimeStampLogitsProcessor(generation_config)]
+            raise ValueError("`TFWhisperForConditionalGeneration` doesn't support returning the timestamps yet.")
+
+        if return_token_timestamps:
+            kwargs["output_attentions"] = True
+            kwargs["return_dict_in_generate"] = True
+
+            if getattr(generation_config, "task", None) == "translate":
+                logger.warning("Token-level timestamps may not be reliable for task 'translate'.")
+            if not hasattr(generation_config, "alignment_heads"):
+                raise ValueError(
+                    "Model generation config has no `alignment_heads`, token-level timestamps not available. "
+                    "See https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a on how to add this property to the generation config."
+                )
+
+        outputs = super().generate(
+            inputs,
+            generation_config,
+            logits_processor,
+            **kwargs,
+        )
+
+        if return_token_timestamps and hasattr(generation_config, "alignment_heads"):
+            outputs["token_timestamps"] = self._extract_token_timestamps(outputs, generation_config.alignment_heads)
+
+        return outputs
+
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqLMOutput(
+            logits=output.logits,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        past_key_values=None,
+        use_cache=None,
+        encoder_outputs=None,
+        attention_mask=None,
+        decoder_attention_mask=None,
+        **kwargs,
+    ):
+        # cut decoder_input_ids if past is used
+        if past_key_values is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        if decoder_attention_mask is not None:  # xla
+            decoder_position_ids = tf.math.cumsum(decoder_attention_mask, axis=-1, exclusive=True)[:, -1:]
+        elif past_key_values is not None:  # no xla + past
+            decoder_position_ids = past_key_values[0][0].shape[2]
+        else:  # no xla + no past
+            decoder_position_ids = tf.range(decoder_input_ids.shape[1])
+        decoder_position_ids = tf.broadcast_to(decoder_position_ids, decoder_input_ids.shape)
+
+        return {
+            "input_features": None,  # Needs to be passed to make Keras.layer.__call__ happy
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past_key_values,
+            "decoder_input_ids": decoder_input_ids,
+            "use_cache": use_cache,
+            "decoder_attention_mask": decoder_attention_mask,
+            "decoder_position_ids": decoder_position_ids,
+        }
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "model", None) is not None:
+            with tf.name_scope(self.model.name):
+                self.model.build(None)

modeling_whisper.cpython-312 (1).pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4526e23b9c4686aa6e18e6c4e49c76bb1c06cc8c70ac6c84f5368cf281a5615
+size 105050

modeling_whisper.cpython-312.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4526e23b9c4686aa6e18e6c4e49c76bb1c06cc8c70ac6c84f5368cf281a5615
+size 105050

processing_whisper (1).py

@@ -0,0 +1,97 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Speech processor class for Whisper
+"""
+
+from ...processing_utils import ProcessorMixin
+
+
+class WhisperProcessor(ProcessorMixin):
+    r"""
+    Constructs a Whisper processor which wraps a Whisper feature extractor and a Whisper tokenizer into a single
+    processor.
+
+    [`WhisperProcessor`] offers all the functionalities of [`WhisperFeatureExtractor`] and [`WhisperTokenizer`]. See
+    the [`~WhisperProcessor.__call__`] and [`~WhisperProcessor.decode`] for more information.
+
+    Args:
+        feature_extractor (`WhisperFeatureExtractor`):
+            An instance of [`WhisperFeatureExtractor`]. The feature extractor is a required input.
+        tokenizer (`WhisperTokenizer`):
+            An instance of [`WhisperTokenizer`]. The tokenizer is a required input.
+    """
+
+    feature_extractor_class = "WhisperFeatureExtractor"
+    tokenizer_class = "WhisperTokenizer"
+
+    def __init__(self, feature_extractor, tokenizer):
+        super().__init__(feature_extractor, tokenizer)
+        self.current_processor = self.feature_extractor
+        self._in_target_context_manager = False
+
+    def get_decoder_prompt_ids(self, task=None, language=None, no_timestamps=True):
+        return self.tokenizer.get_decoder_prompt_ids(task=task, language=language, no_timestamps=no_timestamps)
+
+    def __call__(self, *args, **kwargs):
+        """
+        Forwards the `audio` argument to WhisperFeatureExtractor's [`~WhisperFeatureExtractor.__call__`] and the `text`
+        argument to [`~WhisperTokenizer.__call__`]. Please refer to the doctsring of the above two methods for more
+        information.
+        """
+        # For backward compatibility
+        if self._in_target_context_manager:
+            return self.current_processor(*args, **kwargs)
+
+        audio = kwargs.pop("audio", None)
+        sampling_rate = kwargs.pop("sampling_rate", None)
+        text = kwargs.pop("text", None)
+        if len(args) > 0:
+            audio = args[0]
+            args = args[1:]
+
+        if audio is None and text is None:
+            raise ValueError("You need to specify either an `audio` or `text` input to process.")
+
+        if audio is not None:
+            inputs = self.feature_extractor(audio, *args, sampling_rate=sampling_rate, **kwargs)
+        if text is not None:
+            encodings = self.tokenizer(text, **kwargs)
+
+        if text is None:
+            return inputs
+
+        elif audio is None:
+            return encodings
+        else:
+            inputs["labels"] = encodings["input_ids"]
+            return inputs
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to WhisperTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to WhisperTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def get_prompt_ids(self, text: str, return_tensors="np"):
+        return self.tokenizer.get_prompt_ids(text, return_tensors=return_tensors)

processing_whisper.py

@@ -0,0 +1,97 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Speech processor class for Whisper
+"""
+
+from ...processing_utils import ProcessorMixin
+
+
+class WhisperProcessor(ProcessorMixin):
+    r"""
+    Constructs a Whisper processor which wraps a Whisper feature extractor and a Whisper tokenizer into a single
+    processor.
+
+    [`WhisperProcessor`] offers all the functionalities of [`WhisperFeatureExtractor`] and [`WhisperTokenizer`]. See
+    the [`~WhisperProcessor.__call__`] and [`~WhisperProcessor.decode`] for more information.
+
+    Args:
+        feature_extractor (`WhisperFeatureExtractor`):
+            An instance of [`WhisperFeatureExtractor`]. The feature extractor is a required input.
+        tokenizer (`WhisperTokenizer`):
+            An instance of [`WhisperTokenizer`]. The tokenizer is a required input.
+    """
+
+    feature_extractor_class = "WhisperFeatureExtractor"
+    tokenizer_class = "WhisperTokenizer"
+
+    def __init__(self, feature_extractor, tokenizer):
+        super().__init__(feature_extractor, tokenizer)
+        self.current_processor = self.feature_extractor
+        self._in_target_context_manager = False
+
+    def get_decoder_prompt_ids(self, task=None, language=None, no_timestamps=True):
+        return self.tokenizer.get_decoder_prompt_ids(task=task, language=language, no_timestamps=no_timestamps)
+
+    def __call__(self, *args, **kwargs):
+        """
+        Forwards the `audio` argument to WhisperFeatureExtractor's [`~WhisperFeatureExtractor.__call__`] and the `text`
+        argument to [`~WhisperTokenizer.__call__`]. Please refer to the doctsring of the above two methods for more
+        information.
+        """
+        # For backward compatibility
+        if self._in_target_context_manager:
+            return self.current_processor(*args, **kwargs)
+
+        audio = kwargs.pop("audio", None)
+        sampling_rate = kwargs.pop("sampling_rate", None)
+        text = kwargs.pop("text", None)
+        if len(args) > 0:
+            audio = args[0]
+            args = args[1:]
+
+        if audio is None and text is None:
+            raise ValueError("You need to specify either an `audio` or `text` input to process.")
+
+        if audio is not None:
+            inputs = self.feature_extractor(audio, *args, sampling_rate=sampling_rate, **kwargs)
+        if text is not None:
+            encodings = self.tokenizer(text, **kwargs)
+
+        if text is None:
+            return inputs
+
+        elif audio is None:
+            return encodings
+        else:
+            inputs["labels"] = encodings["input_ids"]
+            return inputs
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to WhisperTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to WhisperTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def get_prompt_ids(self, text: str, return_tensors="np"):
+        return self.tokenizer.get_prompt_ids(text, return_tensors=return_tensors)

realtime-whisper-webgpu/.eslintrc.cjs

@@ -0,0 +1,21 @@
+module.exports = {
+  root: true,
+  env: { browser: true, es2020: true },
+  extends: [
+    "eslint:recommended",
+    "plugin:react/recommended",
+    "plugin:react/jsx-runtime",
+    "plugin:react-hooks/recommended",
+  ],
+  ignorePatterns: ["dist", ".eslintrc.cjs"],
+  parserOptions: { ecmaVersion: "latest", sourceType: "module" },
+  settings: { react: { version: "18.2" } },
+  plugins: ["react-refresh"],
+  rules: {
+    "react-refresh/only-export-components": [
+      "warn",
+      { allowConstantExport: true },
+    ],
+    "react/prop-types": "off",
+  },
+};

realtime-whisper-webgpu/.gitignore

@@ -0,0 +1,24 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+pnpm-debug.log*
+lerna-debug.log*
+
+node_modules
+dist
+dist-ssr
+*.local
+
+# Editor directories and files
+.vscode/*
+!.vscode/extensions.json
+.idea
+.DS_Store
+*.suo
+*.ntvs*
+*.njsproj
+*.sln
+*.sw?

realtime-whisper-webgpu/README.md

@@ -0,0 +1,8 @@
+# React + Vite
+
+This template provides a minimal setup to get React working in Vite with HMR and some ESLint rules.
+
+Currently, two official plugins are available:
+
+- [@vitejs/plugin-react](https://github.com/vitejs/vite-plugin-react/blob/main/packages/plugin-react/README.md) uses [Babel](https://babeljs.io/) for Fast Refresh
+- [@vitejs/plugin-react-swc](https://github.com/vitejs/vite-plugin-react-swc) uses [SWC](https://swc.rs/) for Fast Refresh

realtime-whisper-webgpu/index.html

@@ -0,0 +1,13 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <link rel="icon" type="image/png" href="/logo.png" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>Whisper WebGPU</title>
+  </head>
+  <body>
+    <div id="root"></div>
+    <script type="module" src="/src/main.jsx"></script>
+  </body>
+</html>

realtime-whisper-webgpu/package.json

@@ -0,0 +1,30 @@
+{
+  "name": "realtime-whisper-webgpu",
+  "private": true,
+  "version": "0.0.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "vite build",
+    "lint": "eslint . --ext js,jsx --report-unused-disable-directives --max-warnings 0",
+    "preview": "vite preview"
+  },
+  "dependencies": {
+    "@huggingface/transformers": "3.0.0",
+    "react": "^18.2.0",
+    "react-dom": "^18.2.0"
+  },
+  "devDependencies": {
+    "@types/react": "^18.2.43",
+    "@types/react-dom": "^18.2.17",
+    "@vitejs/plugin-react": "^4.2.1",
+    "autoprefixer": "^10.4.19",
+    "eslint": "^8.55.0",
+    "eslint-plugin-react": "^7.33.2",
+    "eslint-plugin-react-hooks": "^4.6.0",
+    "eslint-plugin-react-refresh": "^0.4.5",
+    "postcss": "^8.4.38",
+    "tailwindcss": "^3.4.3",
+    "vite": "^5.2.11"
+  }
+}

realtime-whisper-webgpu/postcss.config.js

@@ -0,0 +1,6 @@
+export default {
+  plugins: {
+    tailwindcss: {},
+    autoprefixer: {},
+  },
+};

realtime-whisper-webgpu/public/realtime-whisper-webgpu.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:523ddcfcab7baaca8ee13187a6074e7cbe4422359b1208fa8d8228095c0ee9aa
+size 15302693

realtime-whisper-webgpu/src/App.jsx

@@ -0,0 +1,321 @@
+import { useEffect, useState, useRef } from "react";
+
+import { AudioVisualizer } from "./components/AudioVisualizer";
+import Progress from "./components/Progress";
+import { LanguageSelector } from "./components/LanguageSelector";
+
+const IS_WEBGPU_AVAILABLE = !!navigator.gpu;
+
+const WHISPER_SAMPLING_RATE = 16_000;
+const MAX_AUDIO_LENGTH = 30; // seconds
+const MAX_SAMPLES = WHISPER_SAMPLING_RATE * MAX_AUDIO_LENGTH;
+
+function App() {
+  // Create a reference to the worker object.
+  const worker = useRef(null);
+
+  const recorderRef = useRef(null);
+
+  // Model loading and progress
+  const [status, setStatus] = useState(null);
+  const [loadingMessage, setLoadingMessage] = useState("");
+  const [progressItems, setProgressItems] = useState([]);
+
+  // Inputs and outputs
+  const [text, setText] = useState("");
+  const [tps, setTps] = useState(null);
+  const [language, setLanguage] = useState("en");
+
+  // Processing
+  const [recording, setRecording] = useState(false);
+  const [isProcessing, setIsProcessing] = useState(false);
+  const [chunks, setChunks] = useState([]);
+  const [stream, setStream] = useState(null);
+  const audioContextRef = useRef(null);
+
+  // We use the `useEffect` hook to setup the worker as soon as the `App` component is mounted.
+  useEffect(() => {
+    if (!worker.current) {
+      // Create the worker if it does not yet exist.
+      worker.current = new Worker(new URL("./worker.js", import.meta.url), {
+        type: "module",
+      });
+    }
+
+    // Create a callback function for messages from the worker thread.
+    const onMessageReceived = (e) => {
+      switch (e.data.status) {
+        case "loading":
+          // Model file start load: add a new progress item to the list.
+          setStatus("loading");
+          setLoadingMessage(e.data.data);
+          break;
+
+        case "initiate":
+          setProgressItems((prev) => [...prev, e.data]);
+          break;
+
+        case "progress":
+          // Model file progress: update one of the progress items.
+          setProgressItems((prev) =>
+            prev.map((item) => {
+              if (item.file === e.data.file) {
+                return { ...item, ...e.data };
+              }
+              return item;
+            }),
+          );
+          break;
+
+        case "done":
+          // Model file loaded: remove the progress item from the list.
+          setProgressItems((prev) =>
+            prev.filter((item) => item.file !== e.data.file),
+          );
+          break;
+
+        case "ready":
+          // Pipeline ready: the worker is ready to accept messages.
+          setStatus("ready");
+          recorderRef.current?.start();
+          break;
+
+        case "start":
+          {
+            // Start generation
+            setIsProcessing(true);
+
+            // Request new data from the recorder
+            recorderRef.current?.requestData();
+          }
+          break;
+
+        case "update":
+          {
+            // Generation update: update the output text.
+            const { tps } = e.data;
+            setTps(tps);
+          }
+          break;
+
+        case "complete":
+          // Generation complete: re-enable the "Generate" button
+          setIsProcessing(false);
+          setText(e.data.output);
+          break;
+      }
+    };
+
+    // Attach the callback function as an event listener.
+    worker.current.addEventListener("message", onMessageReceived);
+
+    // Define a cleanup function for when the component is unmounted.
+    return () => {
+      worker.current.removeEventListener("message", onMessageReceived);
+    };
+  }, []);
+
+  useEffect(() => {
+    if (recorderRef.current) return; // Already set
+
+    if (navigator.mediaDevices.getUserMedia) {
+      navigator.mediaDevices
+        .getUserMedia({ audio: true })
+        .then((stream) => {
+          setStream(stream);
+
+          recorderRef.current = new MediaRecorder(stream);
+          audioContextRef.current = new AudioContext({
+            sampleRate: WHISPER_SAMPLING_RATE,
+          });
+
+          recorderRef.current.onstart = () => {
+            setRecording(true);
+            setChunks([]);
+          };
+          recorderRef.current.ondataavailable = (e) => {
+            if (e.data.size > 0) {
+              setChunks((prev) => [...prev, e.data]);
+            } else {
+              // Empty chunk received, so we request new data after a short timeout
+              setTimeout(() => {
+                recorderRef.current.requestData();
+              }, 25);
+            }
+          };
+
+          recorderRef.current.onstop = () => {
+            setRecording(false);
+          };
+        })
+        .catch((err) => console.error("The following error occurred: ", err));
+    } else {
+      console.error("getUserMedia not supported on your browser!");
+    }
+
+    return () => {
+      recorderRef.current?.stop();
+      recorderRef.current = null;
+    };
+  }, []);
+
+  useEffect(() => {
+    if (!recorderRef.current) return;
+    if (!recording) return;
+    if (isProcessing) return;
+    if (status !== "ready") return;
+
+    if (chunks.length > 0) {
+      // Generate from data
+      const blob = new Blob(chunks, { type: recorderRef.current.mimeType });
+
+      const fileReader = new FileReader();
+
+      fileReader.onloadend = async () => {
+        const arrayBuffer = fileReader.result;
+        const decoded =
+          await audioContextRef.current.decodeAudioData(arrayBuffer);
+        let audio = decoded.getChannelData(0);
+        if (audio.length > MAX_SAMPLES) {
+          // Get last MAX_SAMPLES
+          audio = audio.slice(-MAX_SAMPLES);
+        }
+
+        worker.current.postMessage({
+          type: "generate",
+          data: { audio, language },
+        });
+      };
+      fileReader.readAsArrayBuffer(blob);
+    } else {
+      recorderRef.current?.requestData();
+    }
+  }, [status, recording, isProcessing, chunks, language]);
+
+  return IS_WEBGPU_AVAILABLE ? (
+    <div className="flex flex-col h-screen mx-auto justify-end text-gray-800 dark:text-gray-200 bg-white dark:bg-gray-900">
+      {
+        <div className="h-full overflow-auto scrollbar-thin flex justify-center items-center flex-col relative">
+          <div className="flex flex-col items-center mb-1 max-w-[400px] text-center">
+            <img
+              src="logo.png"
+              width="50%"
+              height="auto"
+              className="block"
+            ></img>
+            <h1 className="text-4xl font-bold mb-1">Whisper WebGPU</h1>
+            <h2 className="text-xl font-semibold">
+              Real-time in-browser speech recognition
+            </h2>
+          </div>
+
+          <div className="flex flex-col items-center px-4">
+            {status === null && (
+              <>
+                <p className="max-w-[480px] mb-4">
+                  <br />
+                  You are about to load{" "}
+                  <a
+                    href="https://huggingface.co/onnx-community/whisper-base"
+                    target="_blank"
+                    rel="noreferrer"
+                    className="font-medium underline"
+                  >
+                    whisper-base
+                  </a>
+                  , a 73 million parameter speech recognition model that is
+                  optimized for inference on the web. Once downloaded, the model
+                  (~200&nbsp;MB) will be cached and reused when you revisit the
+                  page.
+                  <br />
+                  <br />
+                  Everything runs directly in your browser using{" "}
+                  <a
+                    href="https://huggingface.co/docs/transformers.js"
+                    target="_blank"
+                    rel="noreferrer"
+                    className="underline"
+                  >
+                    🤗&nbsp;Transformers.js
+                  </a>{" "}
+                  and ONNX Runtime Web, meaning no data is sent to a server. You
+                  can even disconnect from the internet after the model has
+                  loaded!
+                </p>
+
+                <button
+                  className="border px-4 py-2 rounded-lg bg-blue-400 text-white hover:bg-blue-500 disabled:bg-blue-100 disabled:cursor-not-allowed select-none"
+                  onClick={() => {
+                    worker.current.postMessage({ type: "load" });
+                    setStatus("loading");
+                  }}
+                  disabled={status !== null}
+                >
+                  Load model
+                </button>
+              </>
+            )}
+
+            <div className="w-[500px] p-2">
+              <AudioVisualizer className="w-full rounded-lg" stream={stream} />
+              {status === "ready" && (
+                <div className="relative">
+                  <p className="w-full h-[80px] overflow-y-auto overflow-wrap-anywhere border rounded-lg p-2">
+                    {text}
+                  </p>
+                  {tps && (
+                    <span className="absolute bottom-0 right-0 px-1">
+                      {tps.toFixed(2)} tok/s
+                    </span>
+                  )}
+                </div>
+              )}
+            </div>
+            {status === "ready" && (
+              <div className="relative w-full flex justify-center">
+                <LanguageSelector
+                  language={language}
+                  setLanguage={(e) => {
+                    recorderRef.current?.stop();
+                    setLanguage(e);
+                    recorderRef.current?.start();
+                  }}
+                />
+                <button
+                  className="border rounded-lg px-2 absolute right-2"
+                  onClick={() => {
+                    recorderRef.current?.stop();
+                    recorderRef.current?.start();
+                  }}
+                >
+                  Reset
+                </button>
+              </div>
+            )}
+            {status === "loading" && (
+              <div className="w-full max-w-[500px] text-left mx-auto p-4">
+                <p className="text-center">{loadingMessage}</p>
+                {progressItems.map(({ file, progress, total }, i) => (
+                  <Progress
+                    key={i}
+                    text={file}
+                    percentage={progress}
+                    total={total}
+                  />
+                ))}
+              </div>
+            )}
+          </div>
+        </div>
+      }
+    </div>
+  ) : (
+    <div className="fixed w-screen h-screen bg-black z-10 bg-opacity-[92%] text-white text-2xl font-semibold flex justify-center items-center text-center">
+      WebGPU is not supported
+      <br />
+      by this browser :&#40;
+    </div>
+  );
+}
+
+export default App;

realtime-whisper-webgpu/src/components/AudioVisualizer.jsx

@@ -0,0 +1,57 @@
+import { useRef, useCallback, useEffect } from "react";
+
+export function AudioVisualizer({ stream, ...props }) {
+  const canvasRef = useRef(null);
+
+  const visualize = useCallback((stream) => {
+    const audioContext = new (window.AudioContext ||
+      window.webkitAudioContext)();
+    const source = audioContext.createMediaStreamSource(stream);
+    const analyser = audioContext.createAnalyser();
+    analyser.fftSize = 2048;
+    source.connect(analyser);
+
+    const canvas = canvasRef.current;
+    const canvasCtx = canvas.getContext("2d");
+    const bufferLength = analyser.frequencyBinCount;
+    const dataArray = new Uint8Array(bufferLength);
+
+    const drawVisual = () => {
+      requestAnimationFrame(drawVisual);
+      analyser.getByteTimeDomainData(dataArray);
+
+      canvasCtx.fillStyle = "rgb(255, 255, 255)";
+      canvasCtx.fillRect(0, 0, canvas.width, canvas.height);
+
+      canvasCtx.lineWidth = 2;
+      canvasCtx.strokeStyle = "rgb(0, 0, 0)";
+      canvasCtx.beginPath();
+
+      const sliceWidth = (canvas.width * 1.0) / bufferLength;
+
+      let x = 0;
+      for (let i = 0; i < bufferLength; ++i) {
+        const v = dataArray[i] / 128.0;
+        const y = (v * canvas.height) / 2;
+
+        if (i === 0) {
+          canvasCtx.moveTo(x, y);
+        } else {
+          canvasCtx.lineTo(x, y);
+        }
+
+        x += sliceWidth;
+      }
+
+      canvasCtx.lineTo(canvas.width, canvas.height / 2);
+      canvasCtx.stroke();
+    };
+
+    drawVisual();
+  }, []);
+
+  useEffect(() => {
+    stream && visualize(stream);
+  }, [visualize, stream]);
+  return <canvas {...props} width={720} height={240} ref={canvasRef}></canvas>;
+}

realtime-whisper-webgpu/src/components/LanguageSelector.jsx

@@ -0,0 +1,134 @@
+function titleCase(str) {
+  str = str.toLowerCase();
+  return (str.match(/\w+.?/g) || [])
+    .map((word) => {
+      return word.charAt(0).toUpperCase() + word.slice(1);
+    })
+    .join("");
+}
+
+// List of supported languages:
+// https://help.openai.com/en/articles/7031512-whisper-api-faq
+// https://github.com/openai/whisper/blob/248b6cb124225dd263bb9bd32d060b6517e067f8/whisper/tokenizer.py#L79
+const LANGUAGES = {
+  en: "english",
+  zh: "chinese",
+  de: "german",
+  es: "spanish/castilian",
+  ru: "russian",
+  ko: "korean",
+  fr: "french",
+  ja: "japanese",
+  pt: "portuguese",
+  tr: "turkish",
+  pl: "polish",
+  ca: "catalan/valencian",
+  nl: "dutch/flemish",
+  ar: "arabic",
+  sv: "swedish",
+  it: "italian",
+  id: "indonesian",
+  hi: "hindi",
+  fi: "finnish",
+  vi: "vietnamese",
+  he: "hebrew",
+  uk: "ukrainian",
+  el: "greek",
+  ms: "malay",
+  cs: "czech",
+  ro: "romanian/moldavian/moldovan",
+  da: "danish",
+  hu: "hungarian",
+  ta: "tamil",
+  no: "norwegian",
+  th: "thai",
+  ur: "urdu",
+  hr: "croatian",
+  bg: "bulgarian",
+  lt: "lithuanian",
+  la: "latin",
+  mi: "maori",
+  ml: "malayalam",
+  cy: "welsh",
+  sk: "slovak",
+  te: "telugu",
+  fa: "persian",
+  lv: "latvian",
+  bn: "bengali",
+  sr: "serbian",
+  az: "azerbaijani",
+  sl: "slovenian",
+  kn: "kannada",
+  et: "estonian",
+  mk: "macedonian",
+  br: "breton",
+  eu: "basque",
+  is: "icelandic",
+  hy: "armenian",
+  ne: "nepali",
+  mn: "mongolian",
+  bs: "bosnian",
+  kk: "kazakh",
+  sq: "albanian",
+  sw: "swahili",
+  gl: "galician",
+  mr: "marathi",
+  pa: "punjabi/panjabi",
+  si: "sinhala/sinhalese",
+  km: "khmer",
+  sn: "shona",
+  yo: "yoruba",
+  so: "somali",
+  af: "afrikaans",
+  oc: "occitan",
+  ka: "georgian",
+  be: "belarusian",
+  tg: "tajik",
+  sd: "sindhi",
+  gu: "gujarati",
+  am: "amharic",
+  yi: "yiddish",
+  lo: "lao",
+  uz: "uzbek",
+  fo: "faroese",
+  ht: "haitian creole/haitian",
+  ps: "pashto/pushto",
+  tk: "turkmen",
+  nn: "nynorsk",
+  mt: "maltese",
+  sa: "sanskrit",
+  lb: "luxembourgish/letzeburgesch",
+  my: "myanmar/burmese",
+  bo: "tibetan",
+  tl: "tagalog",
+  mg: "malagasy",
+  as: "assamese",
+  tt: "tatar",
+  haw: "hawaiian",
+  ln: "lingala",
+  ha: "hausa",
+  ba: "bashkir",
+  jw: "javanese",
+  su: "sundanese",
+};
+export function LanguageSelector({ language, setLanguage }) {
+  const handleLanguageChange = (event) => {
+    setLanguage(event.target.value);
+  };
+
+  const names = Object.values(LANGUAGES).map(titleCase);
+
+  return (
+    <select
+      className="border rounded-lg p-2 max-w-[100px]"
+      value={language}
+      onChange={handleLanguageChange}
+    >
+      {Object.keys(LANGUAGES).map((key, i) => (
+        <option key={key} value={key}>
+          {names[i]}
+        </option>
+      ))}
+    </select>
+  );
+}

realtime-whisper-webgpu/src/components/Progress.jsx

@@ -0,0 +1,22 @@
+function formatBytes(size) {
+  const i = size == 0 ? 0 : Math.floor(Math.log(size) / Math.log(1024));
+  return (
+    +(size / Math.pow(1024, i)).toFixed(2) * 1 +
+    ["B", "kB", "MB", "GB", "TB"][i]
+  );
+}
+
+export default function Progress({ text, percentage, total }) {
+  percentage ??= 0;
+  return (
+    <div className="w-full bg-gray-100 dark:bg-gray-700 text-left rounded-lg overflow-hidden mb-0.5">
+      <div
+        className="bg-blue-400 whitespace-nowrap px-1 text-sm"
+        style={{ width: `${percentage}%` }}
+      >
+        {text} ({percentage.toFixed(2)}%
+        {isNaN(total) ? "" : ` of ${formatBytes(total)}`})
+      </div>
+    </div>
+  );
+}

realtime-whisper-webgpu/src/index.css

@@ -0,0 +1,32 @@
+@tailwind base;
+@tailwind components;
+@tailwind utilities;
+
+@layer utilities {
+  .scrollbar-thin::-webkit-scrollbar {
+    @apply w-2;
+  }
+
+  .scrollbar-thin::-webkit-scrollbar-track {
+    @apply rounded-full bg-gray-100 dark:bg-gray-700;
+  }
+
+  .scrollbar-thin::-webkit-scrollbar-thumb {
+    @apply rounded-full bg-gray-300 dark:bg-gray-600;
+  }
+
+  .scrollbar-thin::-webkit-scrollbar-thumb:hover {
+    @apply bg-gray-500;
+  }
+
+  .animation-delay-200 {
+    animation-delay: 200ms;
+  }
+  .animation-delay-400 {
+    animation-delay: 400ms;
+  }
+
+  .overflow-wrap-anywhere {
+    overflow-wrap: anywhere;
+  }
+}

realtime-whisper-webgpu/src/main.jsx

@@ -0,0 +1,10 @@
+import React from "react";
+import ReactDOM from "react-dom/client";
+import App from "./App.jsx";
+import "./index.css";
+
+ReactDOM.createRoot(document.getElementById("root")).render(
+  <React.StrictMode>
+    <App />
+  </React.StrictMode>,
+);

realtime-whisper-webgpu/src/worker.js

@@ -0,0 +1,143 @@
+import {
+  AutoTokenizer,
+  AutoProcessor,
+  WhisperForConditionalGeneration,
+  TextStreamer,
+  full,
+} from "@huggingface/transformers";
+
+const MAX_NEW_TOKENS = 64;
+
+/**
+ * This class uses the Singleton pattern to ensure that only one instance of the model is loaded.
+ */
+class AutomaticSpeechRecognitionPipeline {
+  static model_id = "onnx-community/whisper-base";
+  static tokenizer = null;
+  static processor = null;
+  static model = null;
+
+  static async getInstance(progress_callback = null) {
+    this.tokenizer ??= AutoTokenizer.from_pretrained(this.model_id, {
+      progress_callback,
+    });
+    this.processor ??= AutoProcessor.from_pretrained(this.model_id, {
+      progress_callback,
+    });
+
+    this.model ??= WhisperForConditionalGeneration.from_pretrained(
+      this.model_id,
+      {
+        dtype: {
+          encoder_model: "fp32", // 'fp16' works too
+          decoder_model_merged: "q4", // or 'fp32' ('fp16' is broken)
+        },
+        device: "webgpu",
+        progress_callback,
+      },
+    );
+
+    return Promise.all([this.tokenizer, this.processor, this.model]);
+  }
+}
+
+let processing = false;
+async function generate({ audio, language }) {
+  if (processing) return;
+  processing = true;
+
+  // Tell the main thread we are starting
+  self.postMessage({ status: "start" });
+
+  // Retrieve the text-generation pipeline.
+  const [tokenizer, processor, model] =
+    await AutomaticSpeechRecognitionPipeline.getInstance();
+
+  let startTime;
+  let numTokens = 0;
+  let tps;
+  const token_callback_function = () => {
+    startTime ??= performance.now();
+
+    if (numTokens++ > 0) {
+      tps = (numTokens / (performance.now() - startTime)) * 1000;
+    }
+  };
+  const callback_function = (output) => {
+    self.postMessage({
+      status: "update",
+      output,
+      tps,
+      numTokens,
+    });
+  };
+
+  const streamer = new TextStreamer(tokenizer, {
+    skip_prompt: true,
+    skip_special_tokens: true,
+    callback_function,
+    token_callback_function,
+  });
+
+  const inputs = await processor(audio);
+
+  const outputs = await model.generate({
+    ...inputs,
+    max_new_tokens: MAX_NEW_TOKENS,
+    language,
+    streamer,
+  });
+
+  const decoded = tokenizer.batch_decode(outputs, {
+    skip_special_tokens: true,
+  });
+
+  // Send the output back to the main thread
+  self.postMessage({
+    status: "complete",
+    output: decoded,
+  });
+  processing = false;
+}
+
+async function load() {
+  self.postMessage({
+    status: "loading",
+    data: "Loading model...",
+  });
+
+  // Load the pipeline and save it for future use.
+  const [tokenizer, processor, model] =
+    await AutomaticSpeechRecognitionPipeline.getInstance((x) => {
+      // We also add a progress callback to the pipeline so that we can
+      // track model loading.
+      self.postMessage(x);
+    });
+
+  self.postMessage({
+    status: "loading",
+    data: "Compiling shaders and warming up model...",
+  });
+
+  // Run model with dummy input to compile shaders
+  await model.generate({
+    input_features: full([1, 80, 3000], 0.0),
+    max_new_tokens: 1,
+  });
+  self.postMessage({ status: "ready" });
+}
+
+// Listen for messages from the main thread
+self.addEventListener("message", async (e) => {
+  const { type, data } = e.data;
+
+  switch (type) {
+    case "load":
+      load();
+      break;
+
+    case "generate":
+      generate(data);
+      break;
+  }
+});

realtime-whisper-webgpu/tailwind.config.js

@@ -0,0 +1,8 @@
+/** @type {import('tailwindcss').Config} */
+export default {
+  content: ["./index.html", "./src/**/*.{js,ts,jsx,tsx}"],
+  theme: {
+    extend: {},
+  },
+  plugins: [],
+};