Upload model

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,35 @@
+{
+ "architectures": [
+ "T5MIMOModel"
+ ],
+ "auto_map": {
+ "AutoConfig": "configuration_t5mimo.T5MIMOConfig",
+ "AutoModel": "modeling_t5mimo.T5MIMOModel"
+ },
+ "classifier_dropout": 0.0,
+ "d_ff": 1024,
+ "d_kv": 64,
+ "d_model": 256,
+ "decoder_start_token_id": 0,
+ "dense_act_fn": "relu",
+ "dropout_rate": 0.1,
+ "eos_token_id": 1,
+ "feed_forward_proj": "relu",
+ "initializer_factor": 0.05,
+ "is_encoder_decoder": true,
+ "is_gated_act": false,
+ "layer_norm_epsilon": 1e-06,
+ "model_type": "t5mimo",
+ "num_decoder_layers": 4,
+ "num_filters": 64,
+ "num_heads": 4,
+ "num_layers": 4,
+ "num_seqs": 3,
+ "pad_token_id": 0,
+ "relative_attention_max_distance": 128,
+ "relative_attention_num_buckets": 32,
+ "torch_dtype": "float32",
+ "transformers_version": "4.41.1",
+ "use_cache": true,
+ "vocab_size": 4096
+}

configuration_t5mimo.py

@@ -0,0 +1,152 @@
+from typing import Mapping
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxSeq2SeqConfigWithPast
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class T5MIMOConfig(PretrainedConfig):
+ r"""
+ This is the configuration class to store the configuration of a [`T5Model`] or a [`TFT5Model`]. It is used to
+ instantiate a T5 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 T5
+ [google-t5/t5-small](https://huggingface.co/google-t5/t5-small) architecture.
+
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+ documentation from [`PretrainedConfig`] for more information.
+
+ Arguments:
+ vocab_size (`int`, *optional*, defaults to 32128):
+ Vocabulary size of the T5 model. Defines the number of different tokens that can be represented by the
+ `inputs_ids` passed when calling [`T5Model`] or [`TFT5Model`].
+ d_model (`int`, *optional*, defaults to 512):
+ Size of the encoder layers and the pooler layer.
+ d_kv (`int`, *optional*, defaults to 64):
+ Size of the key, query, value projections per attention head. The `inner_dim` of the projection layer will
+ be defined as `num_heads * d_kv`.
+ d_ff (`int`, *optional*, defaults to 2048):
+ Size of the intermediate feed forward layer in each `T5Block`.
+ num_layers (`int`, *optional*, defaults to 6):
+ Number of hidden layers in the Transformer encoder.
+ num_decoder_layers (`int`, *optional*):
+ Number of hidden layers in the Transformer decoder. Will use the same value as `num_layers` if not set.
+ num_heads (`int`, *optional*, defaults to 8):
+ Number of attention heads for each attention layer in the Transformer encoder.
+ relative_attention_num_buckets (`int`, *optional*, defaults to 32):
+ The number of buckets to use for each attention layer.
+ relative_attention_max_distance (`int`, *optional*, defaults to 128):
+ The maximum distance of the longer sequences for the bucket separation.
+ dropout_rate (`float`, *optional*, defaults to 0.1):
+ The ratio for all dropout layers.
+ classifier_dropout (`float`, *optional*, defaults to 0.0):
+ The dropout ratio for classifier.
+ layer_norm_eps (`float`, *optional*, defaults to 1e-6):
+ The epsilon used by the layer normalization layers.
+ initializer_factor (`float`, *optional*, defaults to 1):
+ A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+ testing).
+ feed_forward_proj (`string`, *optional*, defaults to `"relu"`):
+ Type of feed forward layer to be used. Should be one of `"relu"` or `"gated-gelu"`. T5v1.1 uses the
+ `"gated-gelu"` feed forward projection. Original T5 uses `"relu"`.
+ use_cache (`bool`, *optional*, defaults to `True`):
+ Whether or not the model should return the last key/values attentions (not used by all models).
+ """
+
+ model_type = "t5mimo"
+ keys_to_ignore_at_inference = ["past_key_values"]
+ attribute_map = {"hidden_size": "d_model", "num_attention_heads": "num_heads", "num_hidden_layers": "num_layers"}
+
+ def __init__(
+ self,
+ vocab_size=32128,
+ d_model=512,
+ d_kv=64,
+ d_ff=2048,
+ num_layers=6,
+ num_decoder_layers=None,
+ num_heads=8,
+ relative_attention_num_buckets=32,
+ relative_attention_max_distance=128,
+ dropout_rate=0.1,
+ layer_norm_epsilon=1e-6,
+ initializer_factor=1.0,
+ feed_forward_proj="relu",
+ is_encoder_decoder=True,
+ use_cache=True,
+ pad_token_id=0,
+ eos_token_id=1,
+ decoder_start_token_id = 0,
+ classifier_dropout=0.0,
+ num_seqs=3,
+ num_filters=64,
+ **kwargs,
+ ):
+ self.vocab_size = vocab_size
+ self.d_model = d_model
+ self.d_kv = d_kv
+ self.d_ff = d_ff
+ self.num_layers = num_layers
+ self.num_decoder_layers = (
+ num_decoder_layers if num_decoder_layers is not None else self.num_layers
+ ) # default = symmetry
+ self.num_heads = num_heads
+ self.relative_attention_num_buckets = relative_attention_num_buckets
+ self.relative_attention_max_distance = relative_attention_max_distance
+ self.dropout_rate = dropout_rate
+ self.classifier_dropout = classifier_dropout
+ self.layer_norm_epsilon = layer_norm_epsilon
+ self.initializer_factor = initializer_factor
+ self.feed_forward_proj = feed_forward_proj
+ self.use_cache = use_cache
+ self.num_seqs = num_seqs
+ self.num_filters = num_filters
+
+ act_info = self.feed_forward_proj.split("-")
+ self.dense_act_fn = act_info[-1]
+ self.is_gated_act = act_info[0] == "gated"
+
+ if len(act_info) > 1 and act_info[0] != "gated" or len(act_info) > 2:
+ raise ValueError(
+ f"`feed_forward_proj`: {feed_forward_proj} is not a valid activation function of the dense layer. "
+ "Please make sure `feed_forward_proj` is of the format `gated-{ACT_FN}` or `{ACT_FN}`, e.g. "
+ "'gated-gelu' or 'relu'"
+ )
+
+ # for backwards compatibility
+ if feed_forward_proj == "gated-gelu":
+ self.dense_act_fn = "gelu_new"
+
+ super().__init__(
+ pad_token_id=pad_token_id,
+ eos_token_id=eos_token_id,
+ decoder_start_token_id=decoder_start_token_id,
+ is_encoder_decoder=is_encoder_decoder,
+ **kwargs,
+ )
+
+
+class T5MIMOOnnxConfig(OnnxSeq2SeqConfigWithPast):
+ @property
+ def inputs(self) -> Mapping[str, Mapping[int, str]]:
+ common_inputs = {
+ "input_ids": {0: "batch", 1: "encoder_sequence"},
+ "attention_mask": {0: "batch", 1: "encoder_sequence"},
+ }
+ if self.use_past:
+ common_inputs["attention_mask"][1] = "past_encoder_sequence + sequence"
+ common_inputs["decoder_input_ids"] = {0: "batch"}
+ common_inputs["decoder_attention_mask"] = {0: "batch", 1: "past_decoder_sequence + sequence"}
+ else:
+ common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
+ common_inputs["decoder_attention_mask"] = {0: "batch", 1: "decoder_sequence"}
+
+ if self.use_past:
+ self.fill_with_past_key_values_(common_inputs, direction="inputs")
+
+ return common_inputs
+
+ @property
+ def default_onnx_opset(self) -> int:
+ return 13

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9d97da94a794f0b0aad1566b9e13205267ea4b3b70ae8c6cd147e6fe6e651cb
+size 33588312

modeling_t5mimo.py

@@ -0,0 +1,1745 @@
+import copy
+import math
+import warnings
+from typing import Optional, Tuple, Union
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+ BaseModelOutput,
+ BaseModelOutputWithPastAndCrossAttentions,
+ Seq2SeqLMOutput,
+ Seq2SeqModelOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS, find_pruneable_heads_and_indices, prune_linear_layer
+from transformers.utils import (
+ DUMMY_INPUTS,
+ DUMMY_MASK,
+ is_torch_fx_proxy,
+ logging,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from .configuration_t5mimo import T5MIMOConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+
+class T5LayerNorm(nn.Module):
+ def __init__(self, hidden_size, eps=1e-6):
+ """
+ Construct a layernorm module in the T5 style. No bias and no subtraction of mean.
+ """
+ super().__init__()
+ self.weight = nn.Parameter(torch.ones(hidden_size))
+ self.variance_epsilon = eps
+
+ def forward(self, hidden_states):
+ # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
+ # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated
+ # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
+ # half-precision inputs is done in fp32
+
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+ # convert into half-precision if necessary
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
+ hidden_states = hidden_states.to(self.weight.dtype)
+
+ return self.weight * hidden_states
+
+
+ALL_LAYERNORM_LAYERS.append(T5LayerNorm)
+
+
+class T5DenseActDense(nn.Module):
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__()
+ self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
+ self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+ self.dropout = nn.Dropout(config.dropout_rate)
+ self.act = ACT2FN[config.dense_act_fn]
+
+ def forward(self, hidden_states):
+ hidden_states = self.wi(hidden_states)
+ hidden_states = self.act(hidden_states)
+ hidden_states = self.dropout(hidden_states)
+ if (
+ isinstance(self.wo.weight, torch.Tensor)
+ and hidden_states.dtype != self.wo.weight.dtype
+ and self.wo.weight.dtype != torch.int8
+ ):
+ hidden_states = hidden_states.to(self.wo.weight.dtype)
+ hidden_states = self.wo(hidden_states)
+ return hidden_states
+
+
+class T5DenseGatedActDense(nn.Module):
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__()
+ self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
+ self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
+ self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+ self.dropout = nn.Dropout(config.dropout_rate)
+ self.act = ACT2FN[config.dense_act_fn]
+
+ def forward(self, hidden_states):
+ hidden_gelu = self.act(self.wi_0(hidden_states))
+ hidden_linear = self.wi_1(hidden_states)
+ hidden_states = hidden_gelu * hidden_linear
+ hidden_states = self.dropout(hidden_states)
+
+ # To make 8bit quantization work for google/flan-t5-xxl, self.wo is kept in float32.
+ # See https://github.com/huggingface/transformers/issues/20287
+ # we also make sure the weights are not in `int8` in case users will force `_keep_in_fp32_modules` to be `None``
+ if (
+ isinstance(self.wo.weight, torch.Tensor)
+ and hidden_states.dtype != self.wo.weight.dtype
+ and self.wo.weight.dtype != torch.int8
+ ):
+ hidden_states = hidden_states.to(self.wo.weight.dtype)
+
+ hidden_states = self.wo(hidden_states)
+ return hidden_states
+
+
+class T5LayerFF(nn.Module):
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__()
+ if config.is_gated_act:
+ self.DenseReluDense = T5DenseGatedActDense(config)
+ else:
+ self.DenseReluDense = T5DenseActDense(config)
+
+ self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+ self.dropout = nn.Dropout(config.dropout_rate)
+
+ def forward(self, hidden_states):
+ forwarded_states = self.layer_norm(hidden_states)
+ forwarded_states = self.DenseReluDense(forwarded_states)
+ hidden_states = hidden_states + self.dropout(forwarded_states)
+ return hidden_states
+
+
+
+class MultivariateConvBlock(nn.Module):
+ def __init__(self, config: T5MIMOConfig, kernel_size=3, stride=1, padding=1):
+ super().__init__()
+ # 2D Convolution across sequences and time
+ self.conv1 = nn.Conv2d(
+ in_channels=config.num_seqs,
+ out_channels=config.num_filters,
+ kernel_size=kernel_size, # Kernel spans across time and all features
+ stride=1, # Stride across time, no stride across features
+ padding=1 # Padding to preserve sequence length, no padding across features
+ )
+ 
+ # Batch normalization for stabilization and faster convergence
+ self.bn1 = nn.BatchNorm2d(config.num_filters)
+ 
+ # Second convolution layer to further model interactions and temporal patterns
+ self.conv2 = nn.Conv2d(
+ in_channels=config.num_filters,
+ out_channels=config.num_filters,
+ kernel_size=(kernel_size, 1), # Focus only on temporal patterns
+ stride=(stride, 1),
+ padding=(padding, 0)
+ )
+ 
+ # Batch normalization after second convolution
+ self.bn2 = nn.BatchNorm2d(config.num_filters)
+ 
+ # 1x1 Convolution to reduce the channel dimension back to num_seqs
+ self.conv3 = nn.Conv2d(
+ in_channels=config.num_filters,
+ out_channels=config.num_seqs, # Back to the original number of sequences (channels)
+ kernel_size=(1, 1)
+ )
+ 
+ def forward(self, x):
+ """
+ Forward pass of the multivariate convolutional block.
+ 
+ Args:
+ x (torch.Tensor): Input tensor of shape [batch_size, num_seqs, seq_len, model_dim].
+ 
+ Returns:
+ torch.Tensor: Output tensor of shape [batch_size, num_seqs, seq_len, model_dim].
+ """
+ # Permute to [batch_size, num_seqs, seq_len, model_dim] -> [batch_size, num_seqs, model_dim, seq_len]
+ x = x.permute(0, 1, 3, 2)
+ 
+ # Apply first convolution and activation
+ x = nn.functional.relu(self.bn1(self.conv1(x)))
+ # Apply second convolution and activation
+ x = nn.functional.relu(self.bn2(self.conv2(x)))
+ 
+ # Reduce channel dimension back to num_seqs
+ x = self.conv3(x)
+ 
+ # Permute back to original shape [batch_size, num_seqs, seq_len, model_dim]
+ x = x.permute(0, 1, 3, 2)
+ 
+ return x
+
+
+
+class T5Attention(nn.Module):
+ def __init__(self, config: T5MIMOConfig, has_relative_attention_bias=False):
+ super().__init__()
+ self.is_decoder = config.is_decoder
+ self.has_relative_attention_bias = has_relative_attention_bias
+ self.relative_attention_num_buckets = config.relative_attention_num_buckets
+ self.relative_attention_max_distance = config.relative_attention_max_distance
+ self.d_model = config.d_model
+ self.key_value_proj_dim = config.d_kv
+ self.n_heads = config.num_heads
+ self.dropout = config.dropout_rate
+ self.inner_dim = self.n_heads * self.key_value_proj_dim
+
+ # Mesh TensorFlow initialization to avoid scaling before softmax
+ self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
+ self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
+ self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
+ self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
+
+ if self.has_relative_attention_bias:
+ self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
+ self.pruned_heads = set()
+ self.gradient_checkpointing = False
+
+ def prune_heads(self, heads):
+ if len(heads) == 0:
+ return
+ heads, index = find_pruneable_heads_and_indices(
+ heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads
+ )
+ # Prune linear layers
+ self.q = prune_linear_layer(self.q, index)
+ self.k = prune_linear_layer(self.k, index)
+ self.v = prune_linear_layer(self.v, index)
+ self.o = prune_linear_layer(self.o, index, dim=1)
+ # Update hyper params
+ self.n_heads = self.n_heads - len(heads)
+ self.inner_dim = self.key_value_proj_dim * self.n_heads
+ self.pruned_heads = self.pruned_heads.union(heads)
+
+ @staticmethod
+ def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
+ """
+ Adapted from Mesh Tensorflow:
+ https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+ Translate relative position to a bucket number for relative attention. The relative position is defined as
+ memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
+ position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
+ small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
+ positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
+ This should allow for more graceful generalization to longer sequences than the model has been trained on
+
+ Args:
+ relative_position: an int32 Tensor
+ bidirectional: a boolean - whether the attention is bidirectional
+ num_buckets: an integer
+ max_distance: an integer
+
+ Returns:
+ a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
+ """
+ relative_buckets = 0
+ if bidirectional:
+ num_buckets //= 2
+ relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
+ relative_position = torch.abs(relative_position)
+ else:
+ relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
+ # now relative_position is in the range [0, inf)
+
+ # half of the buckets are for exact increments in positions
+ max_exact = num_buckets // 2
+ is_small = relative_position < max_exact
+
+ # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
+ relative_position_if_large = max_exact + (
+ torch.log(relative_position.float() / max_exact)
+ / math.log(max_distance / max_exact)
+ * (num_buckets - max_exact)
+ ).to(torch.long)
+ relative_position_if_large = torch.min(
+ relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
+ )
+
+ relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
+ return relative_buckets
+
+ def compute_bias(self, query_length, key_length,multivar_dim=-1, device=None):
+ """Compute binned relative position bias"""
+ if device is None:
+ device = self.relative_attention_bias.weight.device
+ context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
+ memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
+ relative_position = memory_position - context_position # shape (query_length, key_length)
+ relative_position_bucket = self._relative_position_bucket(
+ relative_position, # shape (query_length, key_length)
+ bidirectional=(not self.is_decoder),
+ num_buckets=self.relative_attention_num_buckets,
+ max_distance=self.relative_attention_max_distance,
+ )
+ values = self.relative_attention_bias(relative_position_bucket) # shape (query_length, key_length, num_heads)
+ values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, query_length, key_length)
+ if multivar_dim !=-1: # shape (1, multivar_dim, num_heads, query_length, key_length) (copy across)
+ values = values.expand(1, multivar_dim, -1, -1, -1)
+
+ return values
+
+ def forward(
+ self,
+ hidden_states,
+ mask=None,
+ key_value_states=None,
+ position_bias=None,
+ past_key_value=None,
+ layer_head_mask=None,
+ query_length=None,
+ use_cache=False,
+ output_attentions=False,
+ ):
+ """
+ Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).
+ """
+ # Input is (batch_size, seq_length, dim)
+ # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)
+ # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)
+ if len(hidden_states.shape) == 3:
+ batch_size, seq_length = hidden_states.shape[:2]
+ else:
+ batch_size, seq_length = hidden_states.shape[0],hidden_states.shape[2]
+ multivar_dim = hidden_states.shape[1]
+ real_seq_length = seq_length
+
+ if past_key_value is not None:
+ if len(past_key_value) != 2:
+ raise ValueError(
+ f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states"
+ )
+ real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length
+
+ if len(hidden_states.shape) == 3:
+ key_length = real_seq_length if key_value_states is None else key_value_states.shape[1]
+ else:
+ key_length = real_seq_length if key_value_states is None else key_value_states.shape[2]
+
+
+ def shape(states):
+ """projection"""
+ # states: torch.Size([3, 16, 512]) -> query_states: torch.Size([3, 8, 16, 64])
+ # states: torch.Size([3, 6, 16, 512]) -> query_states: torch.Size([3, 6, 8 , 16, 64])
+ if len(states.shape) == 3:
+ return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
+ else:
+ return states.view(batch_size, multivar_dim, -1, self.n_heads, self.key_value_proj_dim).transpose(2, 3)
+ 
+
+ def unshape(states):
+ """reshape"""
+ if len(states.shape) == 4:
+ return states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)
+ else:
+ return states.transpose(2, 3).contiguous().view(batch_size, multivar_dim, -1, self.inner_dim)
+
+ def project(hidden_states, proj_layer, key_value_states, past_key_value):
+ """projects hidden states correctly to key/query states"""
+ if key_value_states is None:
+ # self-attn
+ # (batch_size, n_heads, seq_length, dim_per_head)
+ hidden_states = shape(proj_layer(hidden_states))
+ elif past_key_value is None:
+ # cross-attn
+ # (batch_size, n_heads, seq_length, dim_per_head)
+ hidden_states = shape(proj_layer(key_value_states))
+
+ if past_key_value is not None:
+ if key_value_states is None:
+ # self-attn
+ # (batch_size, n_heads, key_length, dim_per_head)
+ hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
+ elif past_key_value.shape[2] != key_value_states.shape[1]:
+ # checking that the `sequence_length` of the `past_key_value` is the same as
+ # the provided `key_value_states` to support prefix tuning
+ # cross-attn
+ # (batch_size, n_heads, seq_length, dim_per_head)
+ hidden_states = shape(proj_layer(key_value_states))
+ else:
+ # cross-attn
+ hidden_states = past_key_value
+ return hidden_states
+
+ # get query states
+ query_states = shape(self.q(hidden_states)) # (batch_size, n_heads, seq_length, dim_per_head)
+
+
+ # get key/value states
+ key_states = project(
+ hidden_states, self.k, key_value_states, past_key_value[0] if past_key_value is not None else None
+ )
+ value_states = project(
+ hidden_states, self.v, key_value_states, past_key_value[1] if past_key_value is not None else None
+ )
+
+
+
+ # compute scores
+ if len(hidden_states.shape) == 3:
+ scores = torch.matmul(
+ query_states, key_states.transpose(3, 2)
+ ) # equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9
+ else:
+ scores = torch.matmul(
+ query_states, key_states.transpose(4, 3)
+ )
+
+
+
+
+
+ if position_bias is None:
+ if not self.has_relative_attention_bias:
+
+ if len(hidden_states.shape) == 3:
+ position_bias = torch.zeros(
+ (1, self.n_heads, real_seq_length, key_length), device=scores.device, dtype=scores.dtype
+ )
+ else:
+ position_bias = torch.zeros(
+ (1,multivar_dim, self.n_heads, real_seq_length, key_length), device=scores.device, dtype=scores.dtype
+ )
+ if self.gradient_checkpointing and self.training:
+ position_bias.requires_grad = True
+ else:
+
+ if len(hidden_states.shape) == 3:
+ position_bias = self.compute_bias(real_seq_length, key_length, device=scores.device)
+ else:
+ position_bias = self.compute_bias(real_seq_length, key_length,multivar_dim=multivar_dim, device=scores.device)
+
+ # if key and values are already calculated
+ # we want only the last query position bias
+ if past_key_value is not None:
+ position_bias = position_bias[:, :, -hidden_states.size(1) :, :]
+
+ if mask is not None:
+ position_bias = position_bias + mask # (batch_size, n_heads, seq_length, key_length)
+
+
+
+ if self.pruned_heads:
+ mask = torch.ones(position_bias.shape[1])
+ mask[list(self.pruned_heads)] = 0
+ position_bias_masked = position_bias[:, mask.bool()]
+ else:
+ position_bias_masked = position_bias
+
+
+ scores += position_bias_masked
+ attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(
+ scores
+ ) # (batch_size, n_heads, seq_length, key_length)
+ attn_weights = nn.functional.dropout(
+ attn_weights, p=self.dropout, training=self.training
+ ) # (batch_size, n_heads, seq_length, key_length)
+
+ # Mask heads if we want to
+ if layer_head_mask is not None:
+ attn_weights = attn_weights * layer_head_mask
+
+
+ if len(hidden_states.shape) == 3:
+ attn_output = unshape(torch.matmul(attn_weights, value_states)) # (batch_size, seq_length, dim)
+ else:
+ attn_output = unshape(torch.matmul(attn_weights, value_states)) # (batch_size, multivar_dim, seq_length, dim)
+ attn_output = self.o(attn_output)
+
+
+ present_key_value_state = (key_states, value_states) if (self.is_decoder and use_cache) else None
+ outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)
+
+
+ if output_attentions:
+ outputs = outputs + (attn_weights,)
+
+ return outputs
+
+
+class T5LayerSelfAttention(nn.Module):
+ def __init__(self, config, has_relative_attention_bias=False):
+ super().__init__()
+ self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+ self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+ self.dropout = nn.Dropout(config.dropout_rate)
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ position_bias=None,
+ layer_head_mask=None,
+ past_key_value=None,
+ use_cache=False,
+ output_attentions=False,
+ ):
+ normed_hidden_states = self.layer_norm(hidden_states)
+ attention_output = self.SelfAttention(
+ normed_hidden_states,
+ mask=attention_mask,
+ position_bias=position_bias,
+ layer_head_mask=layer_head_mask,
+ past_key_value=past_key_value,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ )
+ 
+ hidden_states = hidden_states + self.dropout(attention_output[0])
+ outputs = (hidden_states,) + attention_output[1:] # add attentions if we output them
+ return outputs
+
+
+class T5LayerCrossAttention(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.EncDecAttention = T5Attention(config, has_relative_attention_bias=False)
+ self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+ self.dropout = nn.Dropout(config.dropout_rate)
+
+ def forward(
+ self,
+ hidden_states,
+ key_value_states,
+ attention_mask=None,
+ position_bias=None,
+ layer_head_mask=None,
+ past_key_value=None,
+ use_cache=False,
+ query_length=None,
+ output_attentions=False,
+ ):
+ 
+ normed_hidden_states = self.layer_norm(hidden_states)
+ attention_output = self.EncDecAttention(
+ normed_hidden_states,
+ mask=attention_mask,
+ key_value_states=key_value_states,
+ position_bias=position_bias,
+ layer_head_mask=layer_head_mask,
+ past_key_value=past_key_value,
+ use_cache=use_cache,
+ query_length=query_length,
+ output_attentions=output_attentions,
+ )
+ layer_output = hidden_states + self.dropout(attention_output[0])
+ outputs = (layer_output,) + attention_output[1:] # add attentions if we output them
+ return outputs
+
+
+class T5Block(nn.Module):
+ def __init__(self, config, has_relative_attention_bias=False):
+ super().__init__()
+ self.is_decoder = config.is_decoder
+ self.layer = nn.ModuleList()
+ self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+ if self.is_decoder:
+ self.layer.append(T5LayerCrossAttention(config))
+
+ self.layer.append(T5LayerFF(config))
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ position_bias=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ encoder_decoder_position_bias=None,
+ layer_head_mask=None,
+ cross_attn_layer_head_mask=None,
+ past_key_value=None,
+ use_cache=False,
+ output_attentions=False,
+ return_dict=True,
+ ):
+ if past_key_value is not None:
+ if not self.is_decoder:
+ logger.warning("`past_key_values` is passed to the encoder. Please make sure this is intended.")
+ expected_num_past_key_values = 2 if encoder_hidden_states is None else 4
+
+ if len(past_key_value) != expected_num_past_key_values:
+ raise ValueError(
+ f"There should be {expected_num_past_key_values} past states. "
+ f"{'2 (key / value) for cross attention. ' if expected_num_past_key_values == 4 else ''}"
+ f"Got {len(past_key_value)} past key / value states"
+ )
+
+ self_attn_past_key_value = past_key_value[:2]
+ cross_attn_past_key_value = past_key_value[2:]
+ else:
+ self_attn_past_key_value, cross_attn_past_key_value = None, None
+
+ self_attention_outputs = self.layer[0](
+ hidden_states,
+ attention_mask=attention_mask,
+ position_bias=position_bias,
+ layer_head_mask=layer_head_mask,
+ past_key_value=self_attn_past_key_value,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ )
+ hidden_states, present_key_value_state = self_attention_outputs[:2]
+ attention_outputs = self_attention_outputs[2:] # Keep self-attention outputs and relative position weights
+
+ # clamp inf values to enable fp16 training
+ if hidden_states.dtype == torch.float16:
+ clamp_value = torch.where(
+ torch.isinf(hidden_states).any(),
+ torch.finfo(hidden_states.dtype).max - 1000,
+ torch.finfo(hidden_states.dtype).max,
+ )
+ hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+ do_cross_attention = self.is_decoder and encoder_hidden_states is not None
+ if do_cross_attention:
+ # the actual query length is unknown for cross attention
+ # if using past key value states. Need to inject it here
+ if present_key_value_state is not None:
+ query_length = present_key_value_state[0].shape[2]
+ else:
+ query_length = None
+
+ cross_attention_outputs = self.layer[1](
+ hidden_states,
+ key_value_states=encoder_hidden_states,
+ attention_mask=encoder_attention_mask,
+ position_bias=encoder_decoder_position_bias,
+ layer_head_mask=cross_attn_layer_head_mask,
+ past_key_value=cross_attn_past_key_value,
+ query_length=query_length,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ )
+ hidden_states = cross_attention_outputs[0]
+
+ # clamp inf values to enable fp16 training
+ if hidden_states.dtype == torch.float16:
+ clamp_value = torch.where(
+ torch.isinf(hidden_states).any(),
+ torch.finfo(hidden_states.dtype).max - 1000,
+ torch.finfo(hidden_states.dtype).max,
+ )
+ hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+ # Combine self attn and cross attn key value states
+ if present_key_value_state is not None:
+ present_key_value_state = present_key_value_state + cross_attention_outputs[1]
+
+ # Keep cross-attention outputs and relative position weights
+ attention_outputs = attention_outputs + cross_attention_outputs[2:]
+
+ # Apply Feed Forward layer
+ hidden_states = self.layer[-1](hidden_states)
+
+ # clamp inf values to enable fp16 training
+ if hidden_states.dtype == torch.float16:
+ clamp_value = torch.where(
+ torch.isinf(hidden_states).any(),
+ torch.finfo(hidden_states.dtype).max - 1000,
+ torch.finfo(hidden_states.dtype).max,
+ )
+ hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+ outputs = (hidden_states,)
+
+ if use_cache:
+ outputs = outputs + (present_key_value_state,) + attention_outputs
+ else:
+ outputs = outputs + attention_outputs
+
+ return outputs # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
+
+
+class T5ClassificationHead(nn.Module):
+ """Head for sentence-level classification tasks."""
+
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__()
+ self.dense = nn.Linear(config.d_model, config.d_model)
+ self.dropout = nn.Dropout(p=config.classifier_dropout)
+ self.out_proj = nn.Linear(config.d_model, config.num_labels)
+
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+ hidden_states = self.dropout(hidden_states)
+ hidden_states = self.dense(hidden_states)
+ hidden_states = torch.tanh(hidden_states)
+ hidden_states = self.dropout(hidden_states)
+ hidden_states = self.out_proj(hidden_states)
+ return hidden_states
+
+
+class T5PreTrainedModel(PreTrainedModel):
+ """
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+ models.
+ """
+
+ config_class = T5MIMOConfig
+ base_model_prefix = "transformer"
+ is_parallelizable = True
+ supports_gradient_checkpointing = True
+ _no_split_modules = ["T5Block"]
+ _keep_in_fp32_modules = ["wo"]
+
+ @property
+ def dummy_inputs(self):
+ input_ids = torch.tensor(DUMMY_INPUTS)
+ input_mask = torch.tensor(DUMMY_MASK)
+ dummy_inputs = {
+ "decoder_input_ids": input_ids,
+ "input_ids": input_ids,
+ "decoder_attention_mask": input_mask,
+ }
+ return dummy_inputs
+
+ def _init_weights(self, module):
+ """Initialize the weights"""
+ factor = self.config.initializer_factor # Used for testing weights initialization
+ if isinstance(module, T5LayerNorm):
+ module.weight.data.fill_(factor * 1.0)
+ elif isinstance(
+ module,
+ (T5MIMOModel, T5MIMOForConditionalGeneration, T5MIMOEncoderModel),
+ ):
+ # Mesh TensorFlow embeddings initialization
+ # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
+ module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)
+ if hasattr(module, "lm_head") and not self.config.tie_word_embeddings:
+ module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)
+ if hasattr(module, "qa_outputs"):
+ module.qa_outputs.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+ module.qa_outputs.bias.data.zero_()
+ elif isinstance(module, T5ClassificationHead):
+ module.dense.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+ if hasattr(module.dense, "bias") and module.dense.bias is not None:
+ module.dense.bias.data.zero_()
+ module.out_proj.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+ if hasattr(module.out_proj, "bias") and module.out_proj.bias is not None:
+ module.out_proj.bias.data.zero_()
+ elif isinstance(module, T5DenseActDense):
+ # Mesh TensorFlow FF initialization
+ # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
+ # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
+ module.wi.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+ if hasattr(module.wi, "bias") and module.wi.bias is not None:
+ module.wi.bias.data.zero_()
+ module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
+ if hasattr(module.wo, "bias") and module.wo.bias is not None:
+ module.wo.bias.data.zero_()
+ elif isinstance(module, T5DenseGatedActDense):
+ module.wi_0.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+ if hasattr(module.wi_0, "bias") and module.wi_0.bias is not None:
+ module.wi_0.bias.data.zero_()
+ module.wi_1.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+ if hasattr(module.wi_1, "bias") and module.wi_1.bias is not None:
+ module.wi_1.bias.data.zero_()
+ module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
+ if hasattr(module.wo, "bias") and module.wo.bias is not None:
+ module.wo.bias.data.zero_()
+ elif isinstance(module, T5Attention):
+ # Mesh TensorFlow attention initialization to avoid scaling before softmax
+ # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
+ d_model = self.config.d_model
+ key_value_proj_dim = self.config.d_kv
+ n_heads = self.config.num_heads
+ module.q.weight.data.normal_(mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5))
+ module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
+ module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
+ module.o.weight.data.normal_(mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5))
+ if module.has_relative_attention_bias:
+ module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5))
+
+ def _shift_right(self, input_ids):
+ decoder_start_token_id = self.config.decoder_start_token_id
+ pad_token_id = self.config.pad_token_id
+
+ if decoder_start_token_id is None:
+ raise ValueError(
+ "self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id. "
+ "See T5 docs for more information."
+ )
+
+ # shift inputs to the right
+ if is_torch_fx_proxy(input_ids):
+ # Item assignment is not supported natively for proxies.
+ shifted_input_ids = torch.full(input_ids.shape[:-1] + (1,), decoder_start_token_id)
+ shifted_input_ids = torch.cat([shifted_input_ids, input_ids[..., :-1]], dim=-1)
+ else:
+ shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+ shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
+ shifted_input_ids[..., 0] = decoder_start_token_id
+
+ if pad_token_id is None:
+ raise ValueError("self.model.config.pad_token_id has to be defined.")
+ # replace possible -100 values in labels by `pad_token_id`
+ shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+ return shifted_input_ids
+
+
+class T5Stack(T5PreTrainedModel):
+ def __init__(self, config, embed_tokens=None):
+ super().__init__(config)
+
+ self.embed_tokens = embed_tokens
+ self.is_decoder = config.is_decoder
+
+ self.block = nn.ModuleList(
+ [T5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)]
+ )
+ self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+ self.dropout = nn.Dropout(config.dropout_rate)
+
+ # Initialize weights and apply final processing
+ self.post_init()
+ # Model parallel
+ self.model_parallel = False
+ self.device_map = None
+ self.gradient_checkpointing = False
+
+ def parallelize(self, device_map=None):
+ warnings.warn(
+ "`T5Stack.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your model"
+ " with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'block.0': 0,"
+ " 'block.1': 1, ...}",
+ FutureWarning,
+ )
+ # Check validity of device_map
+ self.device_map = (
+ get_device_map(len(self.block), range(torch.cuda.device_count())) if device_map is None else device_map
+ )
+ assert_device_map(self.device_map, len(self.block))
+ self.model_parallel = True
+ self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
+ self.last_device = "cuda:" + str(max(self.device_map.keys()))
+ # Load onto devices
+ for k, v in self.device_map.items():
+ for layer in v:
+ cuda_device = "cuda:" + str(k)
+ self.block[layer] = self.block[layer].to(cuda_device)
+
+ # Set embed_tokens to first layer
+ self.embed_tokens = self.embed_tokens.to(self.first_device)
+ # Set final layer norm to last device
+ self.final_layer_norm = self.final_layer_norm.to(self.last_device)
+
+
+ def deparallelize(self):
+ warnings.warn(
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+ FutureWarning,
+ )
+ self.model_parallel = False
+ self.device_map = None
+ self.first_device = "cpu"
+ self.last_device = "cpu"
+ for i in range(len(self.block)):
+ self.block[i] = self.block[i].to("cpu")
+ self.embed_tokens = self.embed_tokens.to("cpu")
+ self.final_layer_norm = self.final_layer_norm.to("cpu")
+ torch.cuda.empty_cache()
+
+ def get_input_embeddings(self):
+ return self.embed_tokens
+
+ def set_input_embeddings(self, new_embeddings):
+ self.embed_tokens = new_embeddings
+
+ def forward(
+ self,
+ input_ids=None,
+ attention_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ inputs_embeds=None,
+ head_mask=None,
+ cross_attn_head_mask=None,
+ past_key_values=None,
+ use_cache=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ # Model parallel
+ if self.model_parallel:
+ torch.cuda.set_device(self.first_device)
+ self.embed_tokens = self.embed_tokens.to(self.first_device)
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
+ 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 input_ids is not None and inputs_embeds is not None:
+ err_msg_prefix = "decoder_" if self.is_decoder else ""
+ raise ValueError(
+ f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
+ )
+ elif input_ids is not None:
+ input_shape = input_ids.size()
+ # input_ids = input_ids.view(-1, input_shape[-1])
+ elif inputs_embeds is not None:
+ input_shape = inputs_embeds.size()[:-1]
+ else:
+ err_msg_prefix = "decoder_" if self.is_decoder else ""
+ raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds")
+
+ if inputs_embeds is None:
+ if self.embed_tokens is None:
+ raise ValueError("You have to initialize the model with valid token embeddings")
+ inputs_embeds = self.embed_tokens(input_ids)
+
+ if len(input_shape) == 3: 
+ batch_size, multivar_seqs ,seq_length = input_shape
+ else:
+ batch_size, seq_length = input_shape
+
+ # required mask seq length can be calculated via length of past
+ mask_seq_length = past_key_values[0][0].shape[2] + seq_length if past_key_values is not None else seq_length
+
+ if use_cache is True:
+ if not self.is_decoder:
+ raise ValueError(f"`use_cache` can only be set to `True` if {self} is used as a decoder")
+
+ # initialize past_key_values with `None` if past does not exist
+ if past_key_values is None:
+ past_key_values = [None] * len(self.block)
+
+ if attention_mask is None:
+ if len(input_shape) == 2:
+ attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
+ else:
+ attention_mask = torch.ones(batch_size, multivar_seqs, mask_seq_length, device=inputs_embeds.device)
+
+
+
+ # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+ # ourselves in which case we just need to make it broadcastable to all heads.
+ if len(input_shape) == 2:
+ extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape)
+ else:
+ extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape)
+ # permute from [batch_size, 1, multivar_seqs, seq_length] to [batch_size, multivar_seqs, 1, seq_length]
+ extended_attention_mask = extended_attention_mask.permute(0, 2, 1, 3)
+ # Now make it [batch_size, multivar_seqs, 1, 1, seq_length]
+ extended_attention_mask = extended_attention_mask.unsqueeze(3)
+
+ # If a 2D or 3D attention mask is provided for the cross-attention
+ # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+ if self.is_decoder and encoder_hidden_states is not None:
+
+ if len(encoder_hidden_states.size()) == 3 :
+ encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+ else: 
+ encoder_batch_size, multivar_dem, encoder_sequence_length, _ = encoder_hidden_states.size()
+ 
+ encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+ if encoder_attention_mask is None:
+ encoder_attention_mask = torch.ones(
+ encoder_hidden_shape, device=inputs_embeds.device, dtype=torch.long
+ )
+ if len(input_shape) == 2:
+ encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+ else:
+ encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+ multivar_dim = extended_attention_mask.shape[1]
+ encoder_extended_attention_mask = encoder_extended_attention_mask.unsqueeze(1)
+ encoder_extended_attention_mask = encoder_extended_attention_mask.permute(0, 3, 1, 2, 4)
+
+ else:
+ encoder_extended_attention_mask = None
+ 
+ 
+
+ if self.gradient_checkpointing and self.training:
+ if use_cache:
+ logger.warning_once(
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+ )
+ use_cache = False
+
+ # Prepare head mask if needed
+ head_mask = self.get_head_mask(head_mask, self.config.num_layers)
+ cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
+ present_key_value_states = () if use_cache else None
+ all_hidden_states = () if output_hidden_states else None
+ all_attentions = () if output_attentions else None
+ all_cross_attentions = () if (output_attentions and self.is_decoder) else None
+ position_bias = None
+ encoder_decoder_position_bias = None
+
+ hidden_states = self.dropout(inputs_embeds)
+
+ for i, (layer_module, past_key_value) in enumerate(zip(self.block, past_key_values)):
+ layer_head_mask = head_mask[i]
+ cross_attn_layer_head_mask = cross_attn_head_mask[i]
+ # Model parallel
+ if self.model_parallel:
+ torch.cuda.set_device(hidden_states.device)
+ # Ensure that attention_mask is always on the same device as hidden_states
+ if attention_mask is not None:
+ attention_mask = attention_mask.to(hidden_states.device)
+ if position_bias is not None:
+ position_bias = position_bias.to(hidden_states.device)
+ if encoder_hidden_states is not None:
+ encoder_hidden_states = encoder_hidden_states.to(hidden_states.device)
+ if encoder_extended_attention_mask is not None:
+ encoder_extended_attention_mask = encoder_extended_attention_mask.to(hidden_states.device)
+ if encoder_decoder_position_bias is not None:
+ encoder_decoder_position_bias = encoder_decoder_position_bias.to(hidden_states.device)
+ if layer_head_mask is not None:
+ layer_head_mask = layer_head_mask.to(hidden_states.device)
+ if cross_attn_layer_head_mask is not None:
+ cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(hidden_states.device)
+ if output_hidden_states:
+ all_hidden_states = all_hidden_states + (hidden_states,)
+
+ if self.gradient_checkpointing and self.training:
+ layer_outputs = self._gradient_checkpointing_func(
+ layer_module.forward,
+ hidden_states,
+ extended_attention_mask,
+ position_bias,
+ encoder_hidden_states,
+ encoder_extended_attention_mask,
+ encoder_decoder_position_bias,
+ layer_head_mask,
+ cross_attn_layer_head_mask,
+ None, # past_key_value is always None with gradient checkpointing
+ use_cache,
+ output_attentions,
+ )
+ else:
+ layer_outputs = layer_module(
+ hidden_states,
+ attention_mask=extended_attention_mask,
+ position_bias=position_bias,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_extended_attention_mask,
+ encoder_decoder_position_bias=encoder_decoder_position_bias,
+ layer_head_mask=layer_head_mask,
+ cross_attn_layer_head_mask=cross_attn_layer_head_mask,
+ past_key_value=past_key_value,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ )
+
+ # layer_outputs is a tuple with:
+ # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
+ if use_cache is False:
+ layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]
+
+ hidden_states, present_key_value_state = layer_outputs[:2]
+
+ # We share the position biases between the layers - the first layer store them
+ # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
+ # (cross-attention position bias), (cross-attention weights)
+ position_bias = layer_outputs[2]
+ if self.is_decoder and encoder_hidden_states is not None:
+ encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3]
+ # append next layer key value states
+ if use_cache:
+ present_key_value_states = present_key_value_states + (present_key_value_state,)
+
+ if output_attentions:
+ all_attentions = all_attentions + (layer_outputs[3],)
+ if self.is_decoder:
+ all_cross_attentions = all_cross_attentions + (layer_outputs[5],)
+
+ # Model Parallel: If it's the last layer for that device, put things on the next device
+ if self.model_parallel:
+ for k, v in self.device_map.items():
+ if i == v[-1] and "cuda:" + str(k) != self.last_device:
+ hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+ hidden_states = self.final_layer_norm(hidden_states)
+ hidden_states = self.dropout(hidden_states)
+
+ # Add last layer
+ if output_hidden_states:
+ all_hidden_states = all_hidden_states + (hidden_states,)
+
+ if not return_dict:
+ return tuple(
+ v
+ for v in [
+ hidden_states,
+ present_key_value_states,
+ all_hidden_states,
+ all_attentions,
+ all_cross_attentions,
+ ]
+ if v is not None
+ )
+ return BaseModelOutputWithPastAndCrossAttentions(
+ last_hidden_state=hidden_states,
+ past_key_values=present_key_value_states,
+ hidden_states=all_hidden_states,
+ attentions=all_attentions,
+ cross_attentions=all_cross_attentions,
+ )
+
+
+
+class T5MIMOModel(T5PreTrainedModel):
+ config_class = T5MIMOConfig
+
+ _keys_to_ignore_on_load_unexpected = [
+ "decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+ ]
+ _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]
+
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__(config)
+ self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+ encoder_config = copy.deepcopy(config)
+ encoder_config.is_decoder = False
+ encoder_config.use_cache = False
+ encoder_config.is_encoder_decoder = False
+ self.encoder = T5Stack(encoder_config, self.shared)
+
+ decoder_config = copy.deepcopy(config)
+ decoder_config.is_decoder = True
+ decoder_config.is_encoder_decoder = False
+ decoder_config.num_layers = config.num_decoder_layers
+ self.decoder = T5Stack(decoder_config, self.shared)
+
+ # Initialize weights and apply final processing
+ self.post_init()
+
+ # Model parallel
+ self.model_parallel = False
+ self.device_map = None
+
+
+ def parallelize(self, device_map=None):
+ warnings.warn(
+ "`T5Model.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your model"
+ " with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'encoder.block.0':"
+ " 0, 'encoder.block.1': 1, ...}",
+ FutureWarning,
+ )
+ self.device_map = (
+ get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
+ if device_map is None
+ else device_map
+ )
+ assert_device_map(self.device_map, len(self.encoder.block))
+ self.encoder.parallelize(self.device_map)
+ self.decoder.parallelize(self.device_map)
+ self.model_parallel = True
+
+
+ def deparallelize(self):
+ warnings.warn(
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+ FutureWarning,
+ )
+ self.encoder.deparallelize()
+ self.decoder.deparallelize()
+ self.encoder = self.encoder.to("cpu")
+ self.decoder = self.decoder.to("cpu")
+ self.model_parallel = False
+ self.device_map = None
+ torch.cuda.empty_cache()
+
+ def get_input_embeddings(self):
+ return self.shared
+
+ def set_input_embeddings(self, new_embeddings):
+ self.shared = new_embeddings
+ self.encoder.set_input_embeddings(new_embeddings)
+ self.decoder.set_input_embeddings(new_embeddings)
+
+ def _tie_weights(self):
+ if self.config.tie_word_embeddings:
+ self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared)
+ self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared)
+
+ def get_encoder(self):
+ return self.encoder
+
+ def get_decoder(self):
+ return self.decoder
+
+ def _prune_heads(self, heads_to_prune):
+ """
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+ class PreTrainedModel
+ """
+ for layer, heads in heads_to_prune.items():
+ self.encoder.layer[layer].attention.prune_heads(heads)
+
+ def forward(
+ self,
+ input_ids: Optional[torch.LongTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ decoder_input_ids: Optional[torch.LongTensor] = None,
+ decoder_attention_mask: Optional[torch.BoolTensor] = None,
+ head_mask: Optional[torch.FloatTensor] = None,
+ decoder_head_mask: Optional[torch.FloatTensor] = None,
+ cross_attn_head_mask: Optional[torch.Tensor] = None,
+ encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+ inputs_embeds: Optional[torch.Tensor] = None,
+ decoder_inputs_embeds: Optional[torch.Tensor] = None,
+ use_cache: Optional[bool] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:
+ r"""
+ Returns:
+
+ Example:
+
+ ```python
+ >>> from transformers import AutoTokenizer, T5Model
+
+ >>> tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")
+ >>> model = T5Model.from_pretrained("google-t5/t5-small")
+
+ >>> input_ids = tokenizer(
+ ... "Studies have been shown that owning a dog is good for you", return_tensors="pt"
+ ... ).input_ids # Batch size 1
+ >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids # Batch size 1
+
+ >>> # preprocess: Prepend decoder_input_ids with start token which is pad token for T5Model.
+ >>> # This is not needed for torch's T5ForConditionalGeneration as it does this internally using labels arg.
+ >>> decoder_input_ids = model._shift_right(decoder_input_ids)
+
+ >>> # forward pass
+ >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+ >>> last_hidden_states = outputs.last_hidden_state
+ ```"""
+ 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
+
+ # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+ if head_mask is not None and decoder_head_mask is None:
+ if self.config.num_layers == self.config.num_decoder_layers:
+ decoder_head_mask = head_mask
+
+ # Encode if needed (training, first prediction pass)
+ if encoder_outputs is None:
+ encoder_outputs = self.encoder(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ inputs_embeds=inputs_embeds,
+ head_mask=head_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+ elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+ encoder_outputs = BaseModelOutput(
+ 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,
+ )
+
+ hidden_states = encoder_outputs[0]
+
+ # Set device for model parallelism
+ if self.model_parallel:
+ torch.cuda.set_device(self.decoder.first_device)
+ hidden_states = hidden_states.to(self.decoder.first_device)
+ if decoder_input_ids is not None:
+ decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
+ if attention_mask is not None:
+ attention_mask = attention_mask.to(self.decoder.first_device)
+ if decoder_attention_mask is not None:
+ decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)
+
+ # Decode
+ decoder_outputs = self.decoder(
+ input_ids=decoder_input_ids,
+ attention_mask=decoder_attention_mask,
+ inputs_embeds=decoder_inputs_embeds,
+ past_key_values=past_key_values,
+ encoder_hidden_states=hidden_states,
+ encoder_attention_mask=attention_mask,
+ head_mask=decoder_head_mask,
+ cross_attn_head_mask=cross_attn_head_mask,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ if not return_dict:
+ return decoder_outputs + encoder_outputs
+
+ return Seq2SeqModelOutput(
+ 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,
+ )
+
+
+
+class T5MIMOForConditionalGeneration(T5PreTrainedModel):
+ config_class = T5MIMOConfig
+
+ _keys_to_ignore_on_load_unexpected = [
+ "decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+ ]
+ _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
+
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__(config)
+ self.model_dim = config.d_model
+
+ self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+ encoder_config = copy.deepcopy(config)
+ encoder_config.is_decoder = False
+ encoder_config.use_cache = False
+ encoder_config.is_encoder_decoder = False
+ self.encoder = T5Stack(encoder_config, self.shared)
+
+ decoder_config = copy.deepcopy(config)
+ decoder_config.is_decoder = True
+ decoder_config.is_encoder_decoder = False
+ decoder_config.num_layers = config.num_decoder_layers
+ self.decoder = T5Stack(decoder_config, self.shared)
+
+
+ self.conv_block = MultivariateConvBlock(config)
+ self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+ # Initialize weights and apply final processing
+ self.post_init()
+
+ # Model parallel
+ self.model_parallel = False
+ self.device_map = None
+
+
+ def parallelize(self, device_map=None):
+ warnings.warn(
+ "`T5ForConditionalGeneration.parallelize` is deprecated and will be removed in v5 of Transformers, you"
+ " should load your model with `device_map='balanced'` in the call to `from_pretrained`. You can also"
+ " provide your own `device_map` but it needs to be a dictionary module_name to device, so for instance"
+ " {'encoder.block.0': 0, 'encoder.block.1': 1, ...}",
+ FutureWarning,
+ )
+ self.device_map = (
+ get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
+ if device_map is None
+ else device_map
+ )
+ assert_device_map(self.device_map, len(self.encoder.block))
+ self.encoder.parallelize(self.device_map)
+ self.decoder.parallelize(self.device_map)
+ self.lm_head = self.lm_head.to(self.decoder.first_device)
+ self.model_parallel = True
+
+
+ def deparallelize(self):
+ warnings.warn(
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+ FutureWarning,
+ )
+ self.encoder.deparallelize()
+ self.decoder.deparallelize()
+ self.encoder = self.encoder.to("cpu")
+ self.decoder = self.decoder.to("cpu")
+ self.lm_head = self.lm_head.to("cpu")
+ self.model_parallel = False
+ self.device_map = None
+ torch.cuda.empty_cache()
+
+ def get_input_embeddings(self):
+ return self.shared
+
+ def set_input_embeddings(self, new_embeddings):
+ self.shared = new_embeddings
+ self.encoder.set_input_embeddings(new_embeddings)
+ self.decoder.set_input_embeddings(new_embeddings)
+
+ def _tie_weights(self):
+ if self.config.tie_word_embeddings:
+ self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared)
+ self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared)
+
+ def set_output_embeddings(self, new_embeddings):
+ self.lm_head = new_embeddings
+
+ def get_output_embeddings(self):
+ return self.lm_head
+
+ def get_encoder(self):
+ return self.encoder
+
+ def get_decoder(self):
+ return self.decoder
+
+ def forward(
+ self,
+ input_ids: Optional[torch.LongTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ decoder_input_ids: Optional[torch.LongTensor] = None,
+ decoder_attention_mask: Optional[torch.BoolTensor] = None,
+ head_mask: Optional[torch.FloatTensor] = None,
+ decoder_head_mask: Optional[torch.FloatTensor] = None,
+ cross_attn_head_mask: Optional[torch.Tensor] = None,
+ encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+ inputs_embeds: Optional[torch.FloatTensor] = None,
+ decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
+ labels: Optional[torch.LongTensor] = None,
+ use_cache: Optional[bool] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
+ r"""
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+ Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
+ config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
+ labels in `[0, ..., config.vocab_size]`
+
+ Returns:
+
+ Examples:
+
+ ```python
+ >>> from transformers import AutoTokenizer, T5ForConditionalGeneration
+
+ >>> tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")
+ >>> model = T5ForConditionalGeneration.from_pretrained("google-t5/t5-small")
+
+ >>> # training
+ >>> input_ids = tokenizer("The <extra_id_0> walks in <extra_id_1> park", return_tensors="pt").input_ids
+ >>> labels = tokenizer("<extra_id_0> cute dog <extra_id_1> the <extra_id_2>", return_tensors="pt").input_ids
+ >>> outputs = model(input_ids=input_ids, labels=labels)
+ >>> loss = outputs.loss
+ >>> logits = outputs.logits
+
+ >>> # inference
+ >>> input_ids = tokenizer(
+ ... "summarize: studies have shown that owning a dog is good for you", return_tensors="pt"
+ ... ).input_ids # Batch size 1
+ >>> outputs = model.generate(input_ids)
+ >>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+ >>> # studies have shown that owning a dog is good for you.
+ ```"""
+ 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
+
+ # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+ if head_mask is not None and decoder_head_mask is None:
+ if self.config.num_layers == self.config.num_decoder_layers:
+ decoder_head_mask = head_mask
+
+ # Encode if needed (training, first prediction pass)
+ if encoder_outputs is None:
+ # Convert encoder inputs in embeddings if needed
+ encoder_outputs = self.encoder(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ inputs_embeds=inputs_embeds,
+ head_mask=head_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+ elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+ encoder_outputs = BaseModelOutput(
+ 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,
+ )
+
+ hidden_states = encoder_outputs[0]
+
+ if self.model_parallel:
+ torch.cuda.set_device(self.decoder.first_device)
+
+ if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
+ # get decoder inputs from shifting lm labels to the right
+ decoder_input_ids = self._shift_right(labels)
+
+ # Set device for model parallelism
+ if self.model_parallel:
+ torch.cuda.set_device(self.decoder.first_device)
+ hidden_states = hidden_states.to(self.decoder.first_device)
+ if decoder_input_ids is not None:
+ decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
+ if attention_mask is not None:
+ attention_mask = attention_mask.to(self.decoder.first_device)
+ if decoder_attention_mask is not None:
+ decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)
+
+ # Decode
+ decoder_outputs = self.decoder(
+ input_ids=decoder_input_ids,
+ attention_mask=decoder_attention_mask,
+ inputs_embeds=decoder_inputs_embeds,
+ past_key_values=past_key_values,
+ encoder_hidden_states=hidden_states,
+ encoder_attention_mask=attention_mask,
+ head_mask=decoder_head_mask,
+ cross_attn_head_mask=cross_attn_head_mask,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ sequence_output = decoder_outputs[0]
+
+ # Set device for model parallelism
+ if self.model_parallel:
+ torch.cuda.set_device(self.encoder.first_device)
+ self.lm_head = self.lm_head.to(self.encoder.first_device)
+ sequence_output = sequence_output.to(self.lm_head.weight.device)
+
+ if self.config.tie_word_embeddings:
+ # Rescale output before projecting on vocab
+ # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+ sequence_output = sequence_output * (self.model_dim**-0.5)
+
+ sequence_output = self.conv_block(sequence_output)
+ lm_logits = self.lm_head(sequence_output)
+
+ loss = None
+ if labels is not None:
+ loss_fct = CrossEntropyLoss(ignore_index=-100)
+ # move labels to correct device to enable PP
+ labels = labels.to(lm_logits.device)
+ if len(labels.shape) == 2:
+ loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
+ else:
+ loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.reshape(-1))
+ # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
+
+ if not return_dict:
+ output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
+ return ((loss,) + output) if loss is not None else output
+
+ return Seq2SeqLMOutput(
+ loss=loss,
+ logits=lm_logits,
+ 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 prepare_inputs_for_generation(
+ self,
+ input_ids,
+ past_key_values=None,
+ attention_mask=None,
+ head_mask=None,
+ decoder_head_mask=None,
+ decoder_attention_mask=None,
+ cross_attn_head_mask=None,
+ use_cache=None,
+ encoder_outputs=None,
+ **kwargs,
+ ):
+ # cut decoder_input_ids if past_key_values is used
+ if past_key_values is not None:
+ past_length = past_key_values[0][0].shape[2]
+
+ # Some generation methods already pass only the last input ID
+ if input_ids.shape[1] > past_length:
+ remove_prefix_length = past_length
+ else:
+ # Default to old behavior: keep only final ID
+ remove_prefix_length = input_ids.shape[1] - 1
+
+ input_ids = input_ids[:, remove_prefix_length:]
+
+ return {
+ "decoder_input_ids": input_ids,
+ "past_key_values": past_key_values,
+ "encoder_outputs": encoder_outputs,
+ "attention_mask": attention_mask,
+ "head_mask": head_mask,
+ "decoder_head_mask": decoder_head_mask,
+ "decoder_attention_mask": decoder_attention_mask,
+ "cross_attn_head_mask": cross_attn_head_mask,
+ "use_cache": use_cache,
+ }
+
+ def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
+ return self._shift_right(labels)
+
+ def _reorder_cache(self, past_key_values, beam_idx):
+ # if decoder past is not included in output
+ # speedy decoding is disabled and no need to reorder
+ if past_key_values is None:
+ logger.warning("You might want to consider setting `use_cache=True` to speed up decoding")
+ return past_key_values
+
+ reordered_decoder_past = ()
+ for layer_past_states in past_key_values:
+ # get the correct batch idx from layer past batch dim
+ # batch dim of `past` is at 2nd position
+ reordered_layer_past_states = ()
+ for layer_past_state in layer_past_states:
+ # need to set correct `past` for each of the four key / value states
+ reordered_layer_past_states = reordered_layer_past_states + (
+ layer_past_state.index_select(0, beam_idx.to(layer_past_state.device)),
+ )
+
+ if reordered_layer_past_states[0].shape != layer_past_states[0].shape:
+ raise ValueError(
+ f"reordered_layer_past_states[0] shape {reordered_layer_past_states[0].shape} and layer_past_states[0] shape {layer_past_states[0].shape} mismatched"
+ )
+ if len(reordered_layer_past_states) != len(layer_past_states):
+ raise ValueError(
+ f"length of reordered_layer_past_states {len(reordered_layer_past_states)} and length of layer_past_states {len(layer_past_states)} mismatched"
+ )
+
+ reordered_decoder_past = reordered_decoder_past + (reordered_layer_past_states,)
+ return reordered_decoder_past
+
+
+
+class T5MIMOEncoderModel(T5PreTrainedModel):
+ _tied_weights_keys = ["encoder.embed_tokens.weight"]
+ _keys_to_ignore_on_load_unexpected = [r"decoder"]
+
+ def __init__(self, config: T5MIMOConfig):
+ super().__init__(config)
+ self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+ encoder_config = copy.deepcopy(config)
+ encoder_config.use_cache = False
+ encoder_config.is_encoder_decoder = False
+ self.encoder = T5Stack(encoder_config, self.shared)
+
+ # Initialize weights and apply final processing
+ self.post_init()
+
+ # Model parallel
+ self.model_parallel = False
+ self.device_map = None
+
+ def parallelize(self, device_map=None):
+ warnings.warn(
+ "`T5EncoderModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
+ " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'block.0': 0,"
+ " 'block.1': 1, ...}",
+ FutureWarning,
+ )
+ self.device_map = (
+ get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
+ if device_map is None
+ else device_map
+ )
+ assert_device_map(self.device_map, len(self.encoder.block))
+ self.encoder.parallelize(self.device_map)
+ self.model_parallel = True
+
+ def deparallelize(self):
+ warnings.warn(
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
+ FutureWarning,
+ )
+ self.encoder.deparallelize()
+ self.encoder = self.encoder.to("cpu")
+ self.model_parallel = False
+ self.device_map = None
+ torch.cuda.empty_cache()
+
+ def get_input_embeddings(self):
+ return self.shared
+
+ def set_input_embeddings(self, new_embeddings):
+ self.shared = new_embeddings
+ self.encoder.set_input_embeddings(new_embeddings)
+
+ def _tie_weights(self):
+ if self.config.tie_word_embeddings:
+ self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared)
+
+ def get_encoder(self):
+ return self.encoder
+
+ def _prune_heads(self, heads_to_prune):
+ """
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+ class PreTrainedModel
+ """
+ for layer, heads in heads_to_prune.items():
+ self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)
+
+ def forward(
+ self,
+ input_ids: Optional[torch.LongTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ head_mask: Optional[torch.FloatTensor] = None,
+ inputs_embeds: Optional[torch.FloatTensor] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:
+ r"""
+ Returns:
+
+ Example:
+
+ ```python
+ >>> from transformers import AutoTokenizer, T5EncoderModel
+
+ >>> tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")
+ >>> model = T5EncoderModel.from_pretrained("google-t5/t5-small")
+ >>> input_ids = tokenizer(
+ ... "Studies have been shown that owning a dog is good for you", return_tensors="pt"
+ ... ).input_ids # Batch size 1
+ >>> outputs = model(input_ids=input_ids)
+ >>> last_hidden_states = outputs.last_hidden_state
+ ```"""
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ encoder_outputs = self.encoder(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ inputs_embeds=inputs_embeds,
+ head_mask=head_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ return encoder_outputs
+
+
+