AmberYifan commited on
Commit
c660edb
1 Parent(s): 38a779f

Upload 6 files

Browse files
configuration_phi3_small.py ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
3
+ # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ from typing import Any, Dict, List, Optional, Union
17
+
18
+ from transformers.configuration_utils import PretrainedConfig
19
+ from transformers.utils import logging
20
+
21
+ from functools import cached_property
22
+
23
+ """ Phi3Small model configuration """
24
+ logger = logging.get_logger(__name__)
25
+
26
+
27
+ def next_mult(x, y):
28
+ return (x + y - 1) // y * y
29
+
30
+ class Phi3SmallConfig(PretrainedConfig):
31
+ """
32
+ This is the configuration class to store the configuration of a `Phi3Small` model. It is used to
33
+ instantiate a Phi-3-small model according to the specified arguments, defining the model architecture.
34
+ Instantiating a configuration with the defaults will yield a similar configuration to that of the Phi-3-small
35
+ [phi3](https://arxiv.org/pdf/2404.14219) architecture.
36
+
37
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
38
+ documentation from [`PretrainedConfig`] for more information.
39
+
40
+
41
+ Args:
42
+ vocab_size (`int`, *optional*, defaults to 100352):
43
+ Vocabulary size of the Phi3Small model. Defines the number of different tokens that can be represented by the
44
+ `inputs_ids` passed when calling `Phi3Small`.
45
+ max_position_embeddings (`int`, *optional*, defaults to 8192):
46
+ The maximum sequence length that this model might safely be used with.
47
+ rope_embedding_base (`float`, *optional*, defaults to 10^6):
48
+ The base value for the RoPE (Relative Position Encoding) embedding.
49
+ rope_position_scale (`float`, *optional*, defaults to 1.0):
50
+ The scale factor for the RoPE position encoding.
51
+ rope_scaling (`Optional[Dict[str, Union[float, List[float], int]]]`, *optional*, defaults to None):
52
+ The scaling configuration used for LongRoPE.
53
+ hidden_size (`int`, *optional*, defaults to 4096):
54
+ The size of the hidden layers in the model.
55
+ num_hidden_layers (`int`, *optional*, defaults to 32):
56
+ The number of layers in the model.
57
+ num_attention_heads (`int`, *optional*, defaults to 32):
58
+ The number of query heads in the model.
59
+ num_key_value_heads (`int`, *optional*, defaults to 8):
60
+ The number of key-value heads in the model.
61
+ hidden_act (`str`, *optional*, defaults to "gegelu"):
62
+ The activation function used in the model.
63
+ gegelu_limit (`float`, *optional*, defaults to 20.0):
64
+ The limit value for the GELU activation function (for numerical stability).
65
+ gegelu_pad_to_256 (`bool`, *optional*, defaults to True):
66
+ Whether to pad the intermediate size to a multiple of 256 (for faster matmul ops).
67
+ ff_dim_multiplier (`Optional[int]`, *optional*, defaults to None):
68
+ The dimension multiplier for the feed-forward layers.
69
+ ff_intermediate_size (`Optional[int]`, *optional*, defaults to 14336):
70
+ The intermediate size for the feed-forward layers.
71
+ One of `ff_dim_multiplier` or `ff_intermediate_size` must be specified.
72
+ blocksparse_homo_head_pattern (`bool`, *optional*, defaults to False):
73
+ Whether to use a homogeneous head pattern for block-sparse attention.
74
+ blocksparse_block_size (`int`, *optional*, defaults to 64):
75
+ The block size for block-sparse attention.
76
+ blocksparse_num_local_blocks (`int`, *optional*, defaults to 16):
77
+ The number of local blocks for block-sparse attention.
78
+ The local window used in blocksparse equals `blocksparse_num_local_blocks * blocksparse_block_size`
79
+ blocksparse_vert_stride (`int`, *optional*, defaults to 8):
80
+ The vertical stride for block-sparse attention.
81
+ blocksparse_triton_kernel_block_size (`int`, *optional*, defaults to 64):
82
+ The kernel block size for block-sparse attention.
83
+ dense_attention_every_n_layers (`Optional[int]`, *optional*, defaults to 2):
84
+ The frequency of all dense attention layers in the model
85
+ embedding_dropout_prob (`float`, *optional*, defaults to 0.1):
86
+ The dropout probability for the embedding layer.
87
+ attention_dropout_prob (`float`, *optional*, defaults to 0.0):
88
+ The dropout probability for the attention layers.
89
+ ffn_dropout_prob (`float`, *optional*, defaults to 0.1):
90
+ The dropout probability for the feed-forward layers.
91
+ layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
92
+ The epsilon value for layer normalization.
93
+ initializer_range (`float`, *optional*, defaults to 0.02):
94
+ The range for weight initialization.
95
+ mup_use_scaling (`bool`, *optional*, defaults to True):
96
+ Whether to use scaling for MuP parameters (see: https://arxiv.org/abs/2203.03466).
97
+ mup_width_multiplier (`bool`, *optional*, defaults to 8.0):
98
+ The width multiplier for MuP.
99
+ mup_embedding_multiplier (`bool`, *optional*, defaults to 10.0):
100
+ The embedding multiplier for MuP.
101
+ mup_attn_multiplier (`bool`, *optional*, defaults to 1.0):
102
+ The attention multiplier for MuP.
103
+ use_cache (`bool`, *optional*, defaults to True):
104
+ Whether to use cache for the model.
105
+ bos_token_id (`int`, *optional*, defaults to 100257):
106
+ The token ID for the beginning of sentence.
107
+ eos_token_id (`int`, *optional*, defaults to 100257):
108
+ The token ID for the end of sentence.
109
+ reorder_and_upcast_attn (`bool`, *optional*, defaults to False):
110
+ Whether to reorder and upcast attention.
111
+ pad_sequence_to_multiple_of_64 (`bool`, *optional*, defaults to True):
112
+ Whether to pad the sequence length to a multiple of 64.
113
+ **kwargs:
114
+ Additional keyword arguments.
115
+
116
+ Example:
117
+
118
+ ```python
119
+ >>> from transformers import Phi3SmallConfig, Phi3SmallModel
120
+
121
+ >>> # Initializing a Phi3Small configuration
122
+ >>> configuration = Phi3SmallConfig()
123
+
124
+ >>> # Initializing a model (with random weights) from the configuration
125
+ >>> model = Phi3SmallModel(configuration)
126
+
127
+ >>> # Accessing the model configuration
128
+ >>> configuration = model.config
129
+ ```
130
+ """
131
+
132
+ model_type = "phi3small"
133
+ keys_to_ignore_at_inference = ["past_key_values"]
134
+
135
+
136
+ def __init__(
137
+ self,
138
+ # General information about the model
139
+ vocab_size: int =100352,
140
+ max_position_embeddings: int = 8192,
141
+ # RoPE Related Parameters
142
+ rope_embedding_base: float = 10**6,
143
+ rope_position_scale: float = 1.0,
144
+ rope_scaling: Optional[Dict[str, Union[float, List[float], int]]] = None,
145
+ # General Model Parameters
146
+ hidden_size: int = 4096,
147
+ num_hidden_layers: int = 32,
148
+ # KV Shared Attention Configurations
149
+ num_attention_heads: int = 32,
150
+ num_key_value_heads: int = 8,
151
+ # GEGELU Related Parameters
152
+ hidden_act: str = "gegelu",
153
+ gegelu_limit: float = 20.0,
154
+ gegelu_pad_to_256: bool = True,
155
+ ff_dim_multiplier: Optional[int] = None,
156
+ ff_intermediate_size: Optional[int] = 14336,
157
+ # Block Sparse Attention Parameters
158
+ blocksparse_homo_head_pattern: bool = False,
159
+ blocksparse_block_size: int = 64,
160
+ blocksparse_num_local_blocks: int = 16,
161
+ blocksparse_vert_stride: int = 8,
162
+ blocksparse_triton_kernel_block_size: int = 64,
163
+ # Frequency of block-sparsity
164
+ dense_attention_every_n_layers: Optional[int] = 2,
165
+ # Reegularization parameters
166
+ embedding_dropout_prob: float =0.1,
167
+ attention_dropout_prob: float = 0.0,
168
+ ffn_dropout_prob: float = 0.1,
169
+ layer_norm_epsilon=1e-5,
170
+ initializer_range=0.02,
171
+ # MuP parameters
172
+ mup_use_scaling: bool = True,
173
+ mup_width_multiplier: bool = 8.0,
174
+ mup_embedding_multiplier: bool = 10.0,
175
+ mup_attn_multiplier: bool =1.0,
176
+ use_cache=True,
177
+ # The model does not have a bos token id
178
+ # However, in order for some of the downstream libraries to not break
179
+ # we set this to be the same as the eos_token_id
180
+ bos_token_id: int = 100257,
181
+ eos_token_id: int = 100257,
182
+ reorder_and_upcast_attn=False,
183
+ # Configuration to pad sequence length to a multiple of 64
184
+ pad_sequence_to_multiple_of_64: bool = True,
185
+ **kwargs,
186
+ ):
187
+ self.vocab_size = vocab_size
188
+ self.max_position_embeddings = max_position_embeddings
189
+ self.rope_embedding_base = rope_embedding_base
190
+ self.rope_position_scale = rope_position_scale
191
+ self.rope_scaling = rope_scaling
192
+ self.hidden_size = hidden_size
193
+ # QK Shared Attention
194
+ self.num_hidden_layers = num_hidden_layers
195
+ self.num_attention_heads = num_attention_heads
196
+ self.num_key_value_heads = num_key_value_heads
197
+ # Block Sparse Attention Pattern
198
+ self.blocksparse_homo_head_pattern = blocksparse_homo_head_pattern
199
+ self.blocksparse_block_size = blocksparse_block_size
200
+ self.blocksparse_num_local_blocks = blocksparse_num_local_blocks
201
+ self.blocksparse_vert_stride = blocksparse_vert_stride
202
+ self.blocksparse_triton_kernel_block_size = blocksparse_triton_kernel_block_size
203
+ # Frequency of block sparsity
204
+ self.dense_attention_every_n_layers = dense_attention_every_n_layers
205
+ # Activation function
206
+ self.hidden_act = hidden_act
207
+ self.gegelu_limit = gegelu_limit
208
+ self.gegelu_pad_to_256 = gegelu_pad_to_256
209
+ self.ff_dim_multiplier = ff_dim_multiplier
210
+ self.ff_intermediate_size = ff_intermediate_size
211
+ if self.ff_dim_multiplier is None and self.ff_intermediate_size is None:
212
+ raise ValueError(f"Cannot have both {self.ff_dim_multiplier} and {self.ff_intermediate_size} as None")
213
+ if self.ff_dim_multiplier is not None and self.ff_intermediate_size is not None:
214
+ raise ValueError(f"Cannot specify both {self.ff_dim_multiplier} and {self.ff_intermediate_size}.")
215
+ # General regularization
216
+ self.embedding_dropout_prob = embedding_dropout_prob
217
+ self.attention_dropout_prob = attention_dropout_prob
218
+ self.ffn_dropout_prob = ffn_dropout_prob
219
+ self.layer_norm_epsilon = layer_norm_epsilon
220
+ self.initializer_range = initializer_range
221
+ # MuP parameters
222
+ self.mup_use_scaling = mup_use_scaling
223
+ self.mup_width_multiplier = mup_width_multiplier
224
+ self.mup_embedding_multiplier = mup_embedding_multiplier
225
+ self.mup_attn_multiplier = mup_attn_multiplier
226
+ self.use_cache = use_cache
227
+
228
+ self.reorder_and_upcast_attn = reorder_and_upcast_attn
229
+ self.pad_sequence_to_multiple_of_64 = pad_sequence_to_multiple_of_64
230
+
231
+ self.bos_token_id = bos_token_id
232
+ self.eos_token_id = eos_token_id
233
+
234
+ super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
235
+
236
+ @cached_property
237
+ def dummy_token_indices(self) -> List[int]:
238
+ # Importing here to avoid circular imports
239
+ from .tokenization_phi3_small import Phi3SmallTokenizer
240
+ tokenizer = Phi3SmallTokenizer()
241
+ return tokenizer.dummy_token_indices
242
+
243
+ @property
244
+ def intermediate_size(self) -> int:
245
+ if self.ff_intermediate_size is not None:
246
+ return self.ff_intermediate_size
247
+ intermediate_size = (self.ff_dim_multiplier) * (self.hidden_size // 3) * 2
248
+ if self.gegelu_pad_to_256:
249
+ intermediate_size = next_mult(intermediate_size, 256)
250
+ return intermediate_size
modeling_phi3_small.py ADDED
@@ -0,0 +1,1140 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ from typing import Any, Dict, Optional, List, Tuple, Union
3
+ import torch
4
+ import torch.nn as nn
5
+ import torch.nn.functional as F
6
+
7
+
8
+ from einops import rearrange
9
+
10
+ from transformers.modeling_outputs import SequenceClassifierOutputWithPast, CausalLMOutputWithPast, BaseModelOutputWithPast
11
+ from transformers.modeling_utils import PreTrainedModel
12
+ from transformers.utils import logging
13
+
14
+ from transformers.cache_utils import Cache, DynamicCache
15
+
16
+ from .triton_flash_blocksparse_attn import BlockSparseParams
17
+ from .triton_blocksparse_attention_layer import BlockSparseAttentionLayer
18
+ from .positional_embedding import RotaryEmbedding
19
+
20
+ from .configuration_phi3_small import Phi3SmallConfig
21
+
22
+ # Flash Attention Related Imports
23
+ is_flash_attention_available = False
24
+ try:
25
+ import flash_attn
26
+ if int(flash_attn.__version__.split('.')[0]) < 2:
27
+ from flash_attn.flash_attn_interface import (
28
+ flash_attn_func,
29
+ flash_attn_unpadded_kvpacked_func as flash_attn_varlen_kvpacked_func,
30
+ )
31
+
32
+ # rename `max_seqlen`
33
+ def flash_attn_varlen_qkvpacked_func(qkv, cu_seqlens, max_seqlen, dropout_p=0.0, **kwargs):
34
+ return flash_attn_func(qkv, cu_seqlens, dropout_p=dropout_p, max_s=max_seqlen, **kwargs)
35
+
36
+ else:
37
+ from flash_attn.flash_attn_interface import (
38
+ flash_attn_varlen_kvpacked_func,
39
+ )
40
+ from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
41
+ is_flash_attention_available = True
42
+ except ImportError:
43
+ pass
44
+
45
+ logger = logging.get_logger(__name__)
46
+
47
+ LegacyCache = Tuple[Tuple[torch.FloatTensor]]
48
+
49
+ # Taken from https://github.com/allenai/allennlp/blob/main/allennlp/nn/util.py
50
+ def info_value_of_dtype(dtype: torch.dtype):
51
+ """
52
+ Returns the `finfo` or `iinfo` object of a given PyTorch data type. Does not allow torch.bool.
53
+ """
54
+ if dtype == torch.bool:
55
+ raise TypeError("Does not support torch.bool")
56
+ elif dtype.is_floating_point:
57
+ return torch.finfo(dtype)
58
+ else:
59
+ return torch.iinfo(dtype)
60
+
61
+
62
+ # Taken from https://github.com/allenai/allennlp/blob/main/allennlp/nn/util.py
63
+ def min_value_of_dtype(dtype: torch.dtype):
64
+ """
65
+ Returns the minimum value of a given PyTorch data type. Does not allow torch.bool.
66
+ """
67
+ return info_value_of_dtype(dtype).min
68
+
69
+ # Copied from transformers.models.llama.modeling_llama._get_unpad_data
70
+ def _get_unpad_data(attention_mask):
71
+ seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
72
+ indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
73
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
74
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
75
+ return (
76
+ indices,
77
+ cu_seqlens,
78
+ max_seqlen_in_batch,
79
+ )
80
+
81
+
82
+ @torch.jit.script
83
+ def quick_gelu(x):
84
+ return x * torch.sigmoid(1.702 * x)
85
+
86
+
87
+ @torch.jit.script
88
+ def gegelu(input, limit: Optional[float] = None):
89
+ a_gelu, a_linear = input[..., ::2], input[..., 1::2]
90
+ if limit is not None:
91
+ a_gelu = torch.where(
92
+ torch.isinf(a_gelu), a_gelu, a_gelu.clamp(min=None, max=limit)
93
+ )
94
+ a_linear = torch.where(
95
+ torch.isinf(a_linear), a_linear, a_linear.clamp(min=-limit, max=limit)
96
+ )
97
+ out_gelu = quick_gelu(a_gelu)
98
+ return out_gelu * (a_linear + 1)
99
+
100
+ def collapse_first_n_dims(x: torch.Tensor, n: int) -> torch.Tensor:
101
+ """
102
+ Collapse the first `n` dimensions of a tensor into a single dimension.
103
+
104
+ Args:
105
+ x (torch.Tensor): The input tensor.
106
+ n (int): The number of dimensions to collapse.
107
+
108
+ Returns:
109
+ torch.Tensor: The output tensor.
110
+ """
111
+ return x.view(-1, *x.shape[n:])
112
+
113
+ def pad_tensor_to_next_mult_of(
114
+ tensor: torch.Tensor,
115
+ dim: int,
116
+ n: int,
117
+ ) -> Tuple[torch.Tensor, int]:
118
+ """
119
+ Pads a tensor along a specified dimension to the next multiple of a given number.
120
+
121
+ Args:
122
+ tensor (torch.Tensor): The input tensor.
123
+ dim (int): The dimension along which to pad the tensor.
124
+ n (int): The number to pad the tensor to the next multiple of.
125
+
126
+ Returns:
127
+ Tuple[torch.Tensor, int]: A tuple containing the padded tensor and the amount of padding added.
128
+ """
129
+ residual = tensor.size(dim) % n
130
+ if residual == 0:
131
+ return tensor, 0
132
+ padding = n - residual
133
+ padding_tensor = torch.zeros((*tensor.size()[:dim], padding, *tensor.size()[dim + 1:]), device=tensor.device, dtype=tensor.dtype)
134
+ return torch.cat([tensor, padding_tensor], dim=dim), padding
135
+
136
+ def strip_padding_from_tensor(
137
+ tensor: torch.Tensor,
138
+ dim: int,
139
+ residual: int,
140
+ ) -> torch.Tensor:
141
+ """
142
+ Removes padding from a tensor along a specified dimension.
143
+
144
+ Args:
145
+ tensor (torch.Tensor): The input tensor.
146
+ dim (int): The dimension along which to remove padding.
147
+ residual (int): The amount of padding to remove.
148
+
149
+ Returns:
150
+ torch.Tensor: The tensor with padding removed along the specified dimension.
151
+ """
152
+ return torch.narrow(tensor, dim, 0, tensor.size(dim) - residual)
153
+
154
+ class Phi3SmallMLP(nn.Module):
155
+ def __init__(self, config: Phi3SmallConfig):
156
+ super().__init__()
157
+ self.config = config
158
+ assert self.config.hidden_act == "gegelu", "Only `gegelu` is supported for the Phi-3-small model .."
159
+ self.hidden_size = config.hidden_size
160
+ self.gegelu_limit = config.gegelu_limit
161
+ self.intermediate_size = config.intermediate_size
162
+
163
+ self.up_proj = nn.Linear(self.hidden_size, 2 * self.intermediate_size)
164
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size)
165
+ self.dropout = nn.Dropout(config.ffn_dropout_prob)
166
+
167
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
168
+ return self.dropout(
169
+ self.down_proj(
170
+ gegelu(self.up_proj(x), limit=self.gegelu_limit)
171
+ )
172
+ )
173
+
174
+
175
+ class Phi3SmallSelfAttention(nn.Module):
176
+ def __init__(self, config: Phi3SmallConfig, layer_idx: Optional[int] = None) -> None:
177
+ super().__init__()
178
+ self.config = config
179
+ self.layer_idx = layer_idx
180
+ if layer_idx is None:
181
+ logger.warning_once(
182
+ f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
183
+ "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
184
+ "when creating this class."
185
+ )
186
+
187
+ self.hidden_size = config.hidden_size
188
+ # Number of Query Heads
189
+ self.num_heads = config.num_attention_heads
190
+ self.head_dim = self.hidden_size // self.num_heads
191
+ # Number of Key Value Heads
192
+ self.num_key_value_heads = config.num_key_value_heads
193
+ self.num_q_per_kv = self.num_heads // self.num_key_value_heads
194
+ self.max_position_embeddings = config.max_position_embeddings
195
+ self.rope_embedding_base = config.rope_embedding_base
196
+ self.rope_position_scale = config.rope_position_scale
197
+ self.is_causal = True
198
+
199
+ self.attention_dropout_rate = config.attention_dropout_prob
200
+
201
+ norm_factor = None
202
+ if config.mup_use_scaling:
203
+ norm_factor = self.head_dim / config.mup_attn_multiplier
204
+ else:
205
+ norm_factor = math.sqrt(self.head_dim)
206
+ self.softmax_scale = 1.0 / norm_factor
207
+
208
+ self.query_key_value = nn.Linear(self.hidden_size, (self.num_heads + 2 * self.num_key_value_heads) * self.head_dim)
209
+ self.dense = nn.Linear(self.hidden_size, self.hidden_size)
210
+
211
+ self.blocksparse_params = None
212
+ # layer_idx is 0 indexed because that's what the KV Cache expects.
213
+ if self.config.dense_attention_every_n_layers and ((self.layer_idx + 1) % self.config.dense_attention_every_n_layers == 0):
214
+ logger.info(
215
+ f"Layer {layer_idx + 1} is using dense attention since it is divisible by "
216
+ f"{self.config.dense_attention_every_n_layers}"
217
+ )
218
+ assert is_flash_attention_available, "Flash Attention is not available, but is needed for dense attention"
219
+ else:
220
+ # BlockSparse related Parameters
221
+ self.blocksparse_params = BlockSparseParams.from_config(config)
222
+
223
+ if self.blocksparse:
224
+ active_head_range = None
225
+ """
226
+ ... note(bapatra)::
227
+
228
+ In case of tensor parallelism and while using the heterogeneous head patterns,
229
+ the active head range needs to be modified based on the tensor parallel rank
230
+ and the tensor parallel world size.
231
+
232
+ This is because in the case of heterogeneous head patterns, the kernel needs to know
233
+ which head is on which device, so that it can pick the corresponding blocksparse head
234
+ pattern correctly.
235
+
236
+ Example:
237
+ ```python
238
+
239
+ if not self.blocksparse_params.homo_head_pattern:
240
+ tp_rank = torch.distributed.get_rank() % tp_world_size
241
+ num_heads_per_partition = num_heads // tp_world_size
242
+ active_head_range = (tp_rank * num_heads_per_partition, (tp_rank + 1) * num_heads_per_partition)
243
+
244
+ ```
245
+
246
+ """
247
+
248
+ self._blocksparse_layer = BlockSparseAttentionLayer(
249
+ n_heads=self.num_heads,
250
+ max_seq_len=self.max_position_embeddings,
251
+ sparse_block_size=self.blocksparse_params.block_size,
252
+ local_blocks=self.blocksparse_params.num_local_blocks,
253
+ vert_stride=self.blocksparse_params.vert_stride,
254
+ kernel_block_size=self.blocksparse_params.kernel_block_size,
255
+ homo_head=self.blocksparse_params.homo_head_pattern,
256
+ active_head_range=active_head_range,
257
+ )
258
+ self.rotary_emb = RotaryEmbedding.from_config(config)
259
+
260
+
261
+ @property
262
+ def blocksparse(self):
263
+ return self.blocksparse_params is not None
264
+
265
+ def _split_heads(self, mixed_x_layer: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
266
+ bs, sq, _ = mixed_x_layer.size()
267
+ r"""
268
+ The main idea is that we group tensors as
269
+ [bs, sq, (q00, q01, ... q0m, k0, v0), (q10, q11, ... q1m, k1, v1), ... (qn0, qn1, ... qnm, kn, vn)]
270
+ That ways, when the MP column sharding happens, this tensor will be sharded keeping all the
271
+ queries and keys intact. In order to get the correct qkv, we first break into groups, and then
272
+ index into the groups.
273
+ """
274
+
275
+ intermediate_shape = (bs, sq, -1, (self.num_q_per_kv + 2), self.head_dim)
276
+ mixed_x_layer = mixed_x_layer.view(*intermediate_shape)
277
+ q = mixed_x_layer[:, :, :, :-2]
278
+ k = mixed_x_layer[:, :, :, [-2]]
279
+ v = mixed_x_layer[:, :, :, [-1]]
280
+ q, k, v = [
281
+ rearrange(
282
+ x,
283
+ "bs sq group nh hn -> bs sq (group nh) hn"
284
+ ) for x in (q, k, v)
285
+ ]
286
+ return q, k, v
287
+
288
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._unpad_input
289
+ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
290
+ batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
291
+
292
+
293
+ indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
294
+
295
+ key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
296
+ value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
297
+
298
+ if query_length == kv_seq_len:
299
+ query_layer = index_first_axis(
300
+ query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
301
+ )
302
+ cu_seqlens_q = cu_seqlens_k
303
+ max_seqlen_in_batch_q = max_seqlen_in_batch_k
304
+ indices_q = indices_k
305
+ elif query_length == 1:
306
+ max_seqlen_in_batch_q = 1
307
+ cu_seqlens_q = torch.arange(
308
+ batch_size + 1, dtype=torch.int32, device=query_layer.device
309
+ ) # There is a memcpy here, that is very bad.
310
+ indices_q = cu_seqlens_q[:-1]
311
+ query_layer = query_layer.squeeze(1)
312
+ else:
313
+ # The -q_len: slice assumes left padding.
314
+ attention_mask = attention_mask[:, -query_length:]
315
+ query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
316
+
317
+ return (
318
+ query_layer,
319
+ key_layer,
320
+ value_layer,
321
+ indices_q,
322
+ (cu_seqlens_q, cu_seqlens_k),
323
+ (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
324
+ )
325
+
326
+ def _apply_blocksparse_attention(
327
+ self,
328
+ q: torch.Tensor,
329
+ k: torch.Tensor,
330
+ v: torch.Tensor,
331
+ attention_mask: Optional[torch.LongTensor],
332
+ return_attention_probs: bool = False,
333
+ ) -> torch.Tensor:
334
+ """
335
+ Applies blocksparse attention to the input tensors.
336
+
337
+ Args:
338
+ q (torch.Tensor): The query tensor of shape (bs, nqp, seq_len, hn).
339
+ k (torch.Tensor): The key tensor of shape (bs, nkp, seq_len, hn).
340
+ v (torch.Tensor): The value tensor of shape (bs, nkp, seq_len, hn).
341
+ attention_mask (Optional[torch.LongTensor]): The attention mask tensor of shape (bs, seq_len).
342
+ return_attention_probs (bool, optional): Whether to return attention probabilities. Defaults to False.
343
+
344
+ Returns:
345
+ torch.Tensor: The context layer tensor of shape (bs, nqp, seq_len, hn).
346
+ """
347
+ assert not return_attention_probs, "return_attention_probs is not supported for blocksparse attention"
348
+ q, k, v = q.contiguous(), k.contiguous(), v.contiguous()
349
+ # shape: (bs, nqp, seq_len, hn)
350
+ if torch.is_grad_enabled():
351
+ # Training or non-batched inference
352
+ context_layer = self._blocksparse_layer(
353
+ q=q, k=k, v=v, sm_scale=self.softmax_scale
354
+ )
355
+ elif attention_mask is None:
356
+ if q.size(0) != 1:
357
+ logger.warning_once(
358
+ "You are attempting to do batched inference without passing the attention mask.\n"
359
+ "This is okay if you are running loglikelihood requests. However, if you want to do generation, "
360
+ "this probably won't work as expected. Please pass the attention mask to the forward function."
361
+ )
362
+ context_layer = self._blocksparse_layer(
363
+ q=q, k=k, v=v, sm_scale=self.softmax_scale
364
+ )
365
+ else:
366
+ """
367
+ Shapes of tensors are as follows:
368
+ q: (bs, nqp, seq_len, hdim)
369
+ k: (bs, nkp, seq_len, hdim)
370
+ v: (bs, nkp, seq_len, hdim)
371
+ We first need to transpose the shapes to fit what the
372
+ kernel needs, and the reinvert it back at the end of the operations
373
+ """
374
+ assert attention_mask.ndim == 2, "The kernel, like flash-attention-2, only supports 2d attention masks ..."
375
+ left_paddings = attention_mask.shape[1] - attention_mask.sum(dim=-1)
376
+ # shape: (bs, seq_len, nqp, hdim)
377
+ q = q.transpose(1, 2).contiguous()
378
+ # shape: (bs, seq_len, nkp, hdim)
379
+ k = k.transpose(1, 2).contiguous()
380
+ # shape: (bs, seq_len, nkp, hdim)
381
+ v = v.transpose(1, 2).contiguous()
382
+ context_layer = self._blocksparse_layer(
383
+ q=q, k=k, v=v, sm_scale=self.softmax_scale, left_paddings=left_paddings.to(torch.int32)
384
+ )
385
+ # shape: (bs, nqp, seq_len, hdim)
386
+ context_layer = context_layer.transpose(1, 2).contiguous()
387
+ return context_layer
388
+
389
+ def _apply_dense_attention(
390
+ self,
391
+ q: torch.Tensor,
392
+ k: torch.Tensor,
393
+ v: torch.Tensor,
394
+ attention_mask: torch.Tensor,
395
+ return_attention_probs: bool = False,
396
+ ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
397
+ """
398
+ Apply dense attention
399
+
400
+ Args:
401
+ q (torch.Tensor):
402
+ The query tensor, shape: (bs, num_query_heads, seq_len, head_size)
403
+ k (torch.Tensor):
404
+ The key tensor, shape: (bs, num_query_heads, seq_len, head_size)
405
+ v (torch.Tensor):
406
+ The value tensor, shape: (bs, num_query_heads, seq_len, head_size)
407
+
408
+ return_attention_probs (bool, optional):
409
+ Return the attention probabilities. Defaults to False.
410
+
411
+ Returns:
412
+ Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
413
+ Return the output of the attention aggregation. If `return_attention_probs` is True, then
414
+ also return the attention probabilities
415
+
416
+ .. note::
417
+ Right now, am assuming the expansion for the query key values is already done
418
+ outside. But ideally, since Flash attention handles the GQA correctly, we can
419
+ avoid doing that.
420
+
421
+ """
422
+ attention_dropout_prob = self.attention_dropout_rate if self.training else 0.0
423
+ # Get into the correct shape for the Flash Attention API
424
+ # shape: (bs, seq_len, nqp, hn)
425
+ q = q.transpose(1, 2).contiguous()
426
+ query_length = q.size(1)
427
+ # shape: (bs, seq_len, npq, hn)
428
+ k = k.transpose(1, 2).contiguous()
429
+ # shape: (bs, seq_len, npq, hn)
430
+ v = v.transpose(1, 2).contiguous()
431
+
432
+ if attention_mask is not None:
433
+ causal = q.size(2) == k.size(2)
434
+ batch_size = q.shape[0]
435
+ flat_q, flat_k, flat_v, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
436
+ q, k, v, attention_mask, query_length
437
+ )
438
+ cu_seqlens_q, cu_seqlens_k = cu_seq_lens
439
+ max_seqlen_q, max_seqlen_k = max_seq_lens
440
+ flat_kv = torch.cat((flat_k.unsqueeze(1), flat_v.unsqueeze(1)), dim=1)
441
+ attn_output_unpad = flash_attn_varlen_kvpacked_func(
442
+ q=flat_q,
443
+ kv=flat_kv,
444
+ cu_seqlens_q=cu_seqlens_q,
445
+ cu_seqlens_k=cu_seqlens_k,
446
+ max_seqlen_q=max_seqlen_q,
447
+ max_seqlen_k=max_seqlen_k,
448
+ dropout_p=attention_dropout_prob,
449
+ softmax_scale=self.softmax_scale,
450
+ causal=causal,
451
+ return_attn_probs=return_attention_probs
452
+ )
453
+ attention_output = pad_input(
454
+ attn_output_unpad, indices_q, batch_size, query_length
455
+ )
456
+ else:
457
+ kv = torch.cat((k.unsqueeze(2), v.unsqueeze(2)), dim=2)
458
+ cu_seqlens_q = torch.arange(
459
+ 0, (q.size(0) + 1), device=q.device, dtype=torch.int32
460
+ ) * q.size(1)
461
+ cu_seqlens_kv = torch.arange(
462
+ 0, (kv.size(0) + 1), device=kv.device, dtype=torch.int32
463
+ ) * kv.size(1)
464
+ max_seqlen_q = q.size(1)
465
+ max_seqlen_k = kv.size(1)
466
+ attention_output = flash_attn_varlen_kvpacked_func(
467
+ q=collapse_first_n_dims(q, 2),
468
+ kv=collapse_first_n_dims(kv, 2),
469
+ cu_seqlens_q=cu_seqlens_q,
470
+ cu_seqlens_k=cu_seqlens_kv,
471
+ max_seqlen_q=max_seqlen_q,
472
+ max_seqlen_k=max_seqlen_k,
473
+ dropout_p=attention_dropout_prob,
474
+ softmax_scale=self.softmax_scale,
475
+ causal=q.size(1) == kv.size(1),
476
+ return_attn_probs=return_attention_probs
477
+ )
478
+ if return_attention_probs:
479
+ (context_layer, attn_probs) = attention_output
480
+ context_layer = context_layer.view(q.size(0), q.size(1), -1, q.size(3)).transpose(1, 2).contiguous()
481
+ return (context_layer, attn_probs)
482
+ context_layer = attention_output
483
+ context_layer = context_layer.view(q.size(0), q.size(1), -1, q.size(3)).transpose(1, 2).contiguous()
484
+ return context_layer
485
+
486
+
487
+ def expand_kv_to_q_size(self, kv: torch.Tensor, num_q_per_kv: int) -> torch.Tensor:
488
+ """
489
+ Expand the key-value tensor to match the size of the query tensor.
490
+
491
+ Args:
492
+ kv (torch.Tensor): The key-value tensor of shape (bsz, nkp, 2, seq_len, hdim).
493
+ num_q_per_kv (int): The number of queries per key-value.
494
+
495
+ Returns:
496
+ torch.Tensor: The expanded key-value tensor of shape (bsz, nqp, 2, seq_len, hdim).
497
+ Where nqp = num_q_per_kv * nkp
498
+
499
+ .. note(bapatra)::
500
+ Right now, I am using a repeat_interleave to expand the kv to the size of q.
501
+ This incurs a memory penalty, since the tensors are actually copied.
502
+ TODO: If this does yield benefits, then potentially we can use the re-written
503
+ flash attention kernel that can handle GQA.
504
+ """
505
+
506
+ repeats = torch.tensor([num_q_per_kv] * kv.size(1)).to(kv.device)
507
+ total = repeats.sum()
508
+ expanded_kv = torch.repeat_interleave(
509
+ kv,
510
+ repeats=repeats,
511
+ dim=1,
512
+ output_size=total
513
+ )
514
+ return expanded_kv
515
+
516
+ def forward(
517
+ self,
518
+ hidden_states: torch.Tensor,
519
+ attention_mask: Optional[torch.Tensor] = None,
520
+ position_ids: Optional[torch.LongTensor] = None,
521
+ past_key_values: Optional[Cache] = None,
522
+ output_attentions: bool = False,
523
+ use_cache: bool = False,
524
+ **kwargs,
525
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
526
+ """
527
+ The forward function of the Self Attention Layer.
528
+
529
+ Args:
530
+ hidden_states (torch.Tensor):
531
+ The input tensor of shape (bs, q_len, h).
532
+ attention_mask (Optional[torch.Tensor], optional):
533
+ The attention mask tensor of shape (bs, seq_len). This is the 2D attention mask tensor as is standard in the flash-attention
534
+ kernel.
535
+ Defaults to None.
536
+ position_ids (Optional[torch.LongTensor], optional):
537
+ The position ids tensor of shape (bs, q_len). Defaults to None. Unused by the function.
538
+ past_key_value (Optional[Cache], optional):
539
+ The previous kv cache values. Defaults to None.
540
+ output_attentions (bool, optional):
541
+ Whether to return the attention scores. Defaults to False.
542
+ .. note::
543
+ For the blocksparse attention kernel, we do not support returning the attention scores.
544
+ use_cache (bool, optional):
545
+ Whether to use the cache for storing the kv. Defaults to False.
546
+
547
+ Returns:
548
+ Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
549
+ The output tensor of shape (bs, q_len, h),
550
+ the attention scores tensor of shape (bs, nqp, q_len, seq_len) if `output_attentions` is True,
551
+ and the updated cache values if `use_cache` is True.
552
+
553
+ Notations:
554
+ ------------
555
+ bs: batch size
556
+ sq_len: sequence length of the entire sequence
557
+ q_len: sequence length of the query
558
+ cache_sq: sequence length in the cache
559
+ If there is no cache then cache_sq = 0
560
+ and sq_len = q_len
561
+ otherwise sq_len = q_len + cache_sq
562
+ h: hidden size
563
+ nq: number of query heads
564
+ nkv: number of key heads
565
+ hn: hidden size per head
566
+ hn = h // nq
567
+ nqp: number of query heads (per MP partition)
568
+ nqp = nq // (num mp partitions)
569
+ nkvp: number of key-value heads (per MP partition)
570
+ nkvp = nk // (num mp partitions)
571
+
572
+ """
573
+ # shape: (bs, q_len, h)
574
+ bsz, q_len, _ = hidden_states.size()
575
+
576
+ # shape: (bs, q_len, (nqp + 2 * nkvp) * hn)
577
+ mixed_x_layer = self.query_key_value(hidden_states)
578
+ # shape: (bs, q_len, nqp, hn), shape: (bs, q_len, nkvp, hn), shape: (bs, q_len, nkvp, hn)
579
+ q, k, v = self._split_heads(mixed_x_layer)
580
+
581
+ # shape: (bs, qnp, q_len, hn)
582
+ query_states = q.permute(0, 2, 1, 3).contiguous()
583
+ # shape: (bs, nkvp, q_len, hn)
584
+ key_states = k.permute(0, 2, 1, 3).contiguous()
585
+ # shape: (bs, nkvp, q_len, hn)
586
+ value_states = v.permute(0, 2, 1, 3).contiguous()
587
+
588
+ kv_seq_len = key_states.shape[-2]
589
+ if past_key_values is not None:
590
+ if self.layer_idx is None:
591
+ raise ValueError(
592
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
593
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
594
+ "with a layer index."
595
+ )
596
+ if self.rotary_emb is not None:
597
+ seqlen_offset = past_key_values.get_usable_length(kv_seq_len, layer_idx=self.layer_idx)
598
+ # shape: (bs, nqp, q_len, hn), shape: (bs, nkvp, q_len, hn)
599
+ query_states, key_states = self.rotary_emb(
600
+ query_states, key_states, seq_dimension=2, seqlen_offset=seqlen_offset
601
+ )
602
+ key_states, value_states = past_key_values.update(key_states=key_states, value_states=value_states, layer_idx=self.layer_idx)
603
+ else:
604
+ # In this case seq_len = q_len and cache_sq = 0
605
+ if self.rotary_emb is not None:
606
+ # shape: (bs, nqp, seq_len, hn), shape: (bs, nkvp, seq_len, hn)
607
+ query_states, key_states = self.rotary_emb(query_states, key_states, seq_dimension=2)
608
+
609
+ # shape: (bs, nkvp, 2, seq_len, hn)
610
+ kv_states = torch.cat((key_states.unsqueeze(2), value_states.unsqueeze(2)), dim=2)
611
+ # shape: (bs, nqp, 2, seq_len, hn)
612
+ expanded_kv_states = self.expand_kv_to_q_size(kv_states, num_q_per_kv=self.num_q_per_kv)
613
+ # shape: (bs, nqp, seq_len, hn), shape: (bs, nqp, seq_len, hn)
614
+ expanded_key_states, expanded_value_states = expanded_kv_states[:, :, 0], expanded_kv_states[:, :, 1]
615
+ if self.blocksparse:
616
+ attn_function_output = self._apply_blocksparse_attention(
617
+ q=query_states,
618
+ k=expanded_key_states,
619
+ v=expanded_value_states,
620
+ attention_mask=attention_mask,
621
+ return_attention_probs=output_attentions
622
+ )
623
+ else:
624
+ attn_function_output = self._apply_dense_attention(
625
+ q=query_states,
626
+ k=expanded_key_states,
627
+ v=expanded_value_states,
628
+ attention_mask=attention_mask,
629
+ return_attention_probs=output_attentions
630
+ )
631
+
632
+ attn_weights = None
633
+ if output_attentions:
634
+ attn_output, attn_weights = attn_function_output
635
+ else:
636
+ # shape: (bs, nqp, seq_len, hn)
637
+ attn_output = attn_function_output
638
+ # shape: (bs, seq_len, nqp, hn)
639
+ attn_output = attn_output.transpose(1, 2).contiguous()
640
+
641
+ # shape: (bs, seq_len, h)
642
+ attn_output = attn_output.view(bsz, q_len, -1)
643
+ attn_output = self.dense(attn_output)
644
+ return attn_output, attn_weights, past_key_values
645
+
646
+
647
+ class Phi3SmallDecoderLayer(nn.Module):
648
+ def __init__(self, config: Phi3SmallConfig, layer_idx: int):
649
+ super().__init__()
650
+ self.hidden_size = config.hidden_size
651
+ self.self_attn = Phi3SmallSelfAttention(config, layer_idx)
652
+ self.mlp = Phi3SmallMLP(config)
653
+
654
+ self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
655
+ self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
656
+
657
+ def forward(
658
+ self,
659
+ hidden_states: torch.Tensor,
660
+ attention_mask: Optional[torch.Tensor] = None,
661
+ position_ids: Optional[torch.LongTensor] = None,
662
+ past_key_values: Optional[Cache] = None,
663
+ output_attentions: Optional[bool] = None,
664
+ use_cache: Optional[bool] = None,
665
+ **kwargs,
666
+ ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Cache]]:
667
+ residual = hidden_states
668
+ hidden_states = self.input_layernorm(hidden_states)
669
+
670
+ # Self Attention
671
+ hidden_states, self_attn_weights, present_key_values = self.self_attn(
672
+ hidden_states=hidden_states,
673
+ attention_mask=attention_mask,
674
+ position_ids=position_ids,
675
+ past_key_values=past_key_values,
676
+ output_attentions=output_attentions,
677
+ use_cache=use_cache,
678
+ )
679
+ hidden_states = residual + hidden_states
680
+
681
+ # Fully Connected
682
+ residual = hidden_states
683
+ hidden_states = self.post_attention_layernorm(hidden_states)
684
+ hidden_states = self.mlp(hidden_states)
685
+ hidden_states = residual + hidden_states
686
+
687
+ outputs = (hidden_states,)
688
+
689
+ if output_attentions:
690
+ outputs += (self_attn_weights,)
691
+
692
+ if use_cache:
693
+ outputs += (present_key_values,)
694
+
695
+ return outputs
696
+
697
+
698
+
699
+ class Phi3SmallPreTrainedModel(PreTrainedModel):
700
+ config_class = Phi3SmallConfig
701
+ base_model_prefix = "model"
702
+ supports_gradient_checkpointing = True
703
+ _no_split_modules = ["Phi3SmallDecoderLayer"]
704
+ skip_keys_device_placement = "past_key_values"
705
+ _supports_flash_attn_2 = True
706
+ _supports_sdpa = False
707
+ _supports_cache_class = True
708
+
709
+ def _init_weights(self, module: nn.Module):
710
+ std = self.config.initializer_range
711
+ if isinstance(module, nn.Linear):
712
+ # Slightly different from the TF version which uses truncated_normal for initialization
713
+ # cf https://github.com/pytorch/pytorch/pull/5617
714
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
715
+ elif isinstance(module, nn.Embedding):
716
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
717
+ if module.padding_idx is not None:
718
+ module.weight.data[module.padding_idx].zero_()
719
+ elif isinstance(module, nn.LayerNorm):
720
+ module.bias.data.zero_()
721
+ module.weight.data.fill_(1.0)
722
+
723
+ # The output projection on the decoder attention layer as well as the down_proj in the MLP are scaled
724
+ # differently (dubbed `output_layer_init_method` in the Megatron code). This is replicated here
725
+ for name, p in module.named_parameters():
726
+ if any(x in name for x in ("c_proj.weight", "down_proj.weight", "o_proj.weight")):
727
+ # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
728
+ p.data.normal_(mean=0.0, std=(self.config.initializer_range / math.sqrt(2 * self.config.num_hidden_layers)))
729
+
730
+
731
+ class Phi3SmallModel(Phi3SmallPreTrainedModel):
732
+
733
+ def __init__(self, config):
734
+ super().__init__(config)
735
+ self.config = config
736
+
737
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
738
+
739
+ # Embedding Dropout
740
+ self.embedding_dropout = nn.Dropout(config.embedding_dropout_prob)
741
+
742
+ # MuP Embedding scaling
743
+ self.mup_embedding_multiplier = config.mup_embedding_multiplier
744
+
745
+ self.layers = nn.ModuleList([Phi3SmallDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
746
+
747
+ self.final_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
748
+
749
+ self.gradient_checkpointing = False
750
+
751
+ # Initialize weights and apply final processing
752
+ self.post_init()
753
+
754
+ def get_input_embeddings(self):
755
+ return self.embed_tokens
756
+
757
+ def set_input_embeddings(self, value):
758
+ self.embed_tokens = value
759
+
760
+ @property
761
+ def pad_sequence_to_multiple_of_64(self):
762
+ # We only need to do this for the backward pass. So only required
763
+ # when we are in the context of generating gradients
764
+ return self.config.pad_sequence_to_multiple_of_64 and torch.is_grad_enabled()
765
+
766
+ def forward(
767
+ self,
768
+ input_ids: torch.LongTensor = None,
769
+ attention_mask: Optional[torch.Tensor] = None,
770
+ position_ids: Optional[torch.LongTensor] = None,
771
+ past_key_values: Optional[Union[Cache, LegacyCache]] = None,
772
+ inputs_embeds: Optional[torch.FloatTensor] = None,
773
+ use_cache: Optional[bool] = None,
774
+ output_attentions: Optional[bool] = None,
775
+ output_hidden_states: Optional[bool] = None,
776
+ return_dict: Optional[bool] = None,
777
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
778
+
779
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
780
+ output_hidden_states = (
781
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
782
+ )
783
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
784
+
785
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
786
+
787
+ if input_ids is not None and inputs_embeds is not None:
788
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
789
+ elif input_ids is not None:
790
+ batch_size, seq_length = input_ids.shape
791
+ elif inputs_embeds is not None:
792
+ batch_size, seq_length, _ = inputs_embeds.shape
793
+ else:
794
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
795
+
796
+ if self.gradient_checkpointing and self.training:
797
+ if use_cache:
798
+ logger.warning_once(
799
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
800
+ )
801
+ use_cache = False
802
+
803
+ past_key_values_length = 0
804
+
805
+ if use_cache:
806
+ use_legacy_cache = not isinstance(past_key_values, Cache)
807
+ if use_legacy_cache:
808
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
809
+ past_key_values_length = past_key_values.get_usable_length(seq_length)
810
+
811
+ if position_ids is None:
812
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
813
+ position_ids = torch.arange(
814
+ past_key_values_length, past_key_values_length + seq_length, dtype=torch.long, device=device
815
+ )
816
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
817
+ else:
818
+ position_ids = position_ids.view(-1, seq_length).long()
819
+
820
+ if attention_mask is not None:
821
+ if batch_size <= 0:
822
+ raise ValueError("batch_size has to be defined and > 0")
823
+
824
+ if inputs_embeds is None:
825
+ inputs_embeds = self.embed_tokens(input_ids)
826
+ inputs_embeds = self.embedding_dropout(inputs_embeds)
827
+
828
+ if self.mup_embedding_multiplier is not None and self.mup_embedding_multiplier > 0.0:
829
+ inputs_embeds = inputs_embeds * self.mup_embedding_multiplier
830
+
831
+ residual = 0
832
+ if self.pad_sequence_to_multiple_of_64:
833
+ # note(bapatra): Since we don't particularly use the position_ids and the attention mask
834
+ # we don't need to pad them
835
+ inputs_embeds, residual = pad_tensor_to_next_mult_of(tensor=inputs_embeds, dim=1, n=64)
836
+
837
+ hidden_states = inputs_embeds
838
+
839
+ # decoder layers
840
+ all_hidden_states = () if output_hidden_states else None
841
+ all_self_attns = () if output_attentions else None
842
+ next_decoder_cache = None
843
+
844
+ for decoder_layer in self.layers:
845
+ if output_hidden_states:
846
+ all_hidden_states += (hidden_states,)
847
+
848
+ if self.gradient_checkpointing and self.training:
849
+ layer_outputs = self._gradient_checkpointing_func(
850
+ decoder_layer.__call__,
851
+ hidden_states,
852
+ attention_mask,
853
+ position_ids,
854
+ past_key_values,
855
+ output_attentions,
856
+ use_cache,
857
+ )
858
+ else:
859
+ layer_outputs = decoder_layer(
860
+ hidden_states,
861
+ attention_mask=attention_mask,
862
+ position_ids=position_ids,
863
+ past_key_values=past_key_values,
864
+ output_attentions=output_attentions,
865
+ use_cache=use_cache,
866
+ )
867
+ hidden_states = layer_outputs[0]
868
+
869
+ if use_cache:
870
+ # Following the Mistral schema for layer return values
871
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
872
+ if output_attentions:
873
+ all_self_attns += (layer_outputs[1],)
874
+
875
+ hidden_states = self.final_layernorm(hidden_states)
876
+
877
+ if residual > 0:
878
+ hidden_states = strip_padding_from_tensor(tensor=hidden_states, dim=1, residual=residual)
879
+
880
+ # add hidden states from the last decoder layer
881
+ if output_hidden_states:
882
+ all_hidden_states += (hidden_states,)
883
+
884
+ next_cache = None
885
+ if use_cache:
886
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
887
+
888
+ if not return_dict:
889
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
890
+ return BaseModelOutputWithPast(
891
+ last_hidden_state=hidden_states,
892
+ past_key_values=next_cache,
893
+ hidden_states=all_hidden_states,
894
+ attentions=all_self_attns,
895
+ )
896
+
897
+
898
+ class Phi3SmallForCausalLM(Phi3SmallPreTrainedModel):
899
+ _tied_weights_keys = ["lm_head.weight"]
900
+
901
+ def __init__(self, config):
902
+ super().__init__(config)
903
+ self.model = Phi3SmallModel(config)
904
+ self.vocab_size = config.vocab_size
905
+ self.lm_head = nn.Linear(config.hidden_size, self.vocab_size, bias=False)
906
+ self.mup_width_multiplier = config.mup_width_multiplier
907
+
908
+ # Create the mask for the dummy tokens in the vocabulary
909
+ dummy_token_indices = config.dummy_token_indices
910
+ dummy_tokens_mask = torch.zeros(self.vocab_size).bool()
911
+ dummy_tokens_mask[dummy_token_indices] = True
912
+ # shape: (vocab_size,)
913
+ self.register_buffer("dummy_tokens_mask", dummy_tokens_mask, persistent=False)
914
+
915
+ # Initialize weights and apply final processing
916
+ self.post_init()
917
+
918
+ def get_input_embeddings(self):
919
+ return self.model.embed_tokens
920
+
921
+ def set_input_embeddings(self, value):
922
+ self.model.embed_tokens = value
923
+
924
+ def get_output_embeddings(self):
925
+ return self.lm_head
926
+
927
+ def set_output_embeddings(self, value):
928
+ self.lm_head = value
929
+
930
+ def set_decoder(self, decoder):
931
+ self.model = decoder
932
+
933
+ def get_decoder(self):
934
+ return self.model
935
+
936
+ def forward(
937
+ self,
938
+ input_ids: torch.LongTensor = None,
939
+ attention_mask: Optional[torch.Tensor] = None,
940
+ position_ids: Optional[torch.LongTensor] = None,
941
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
942
+ inputs_embeds: Optional[torch.FloatTensor] = None,
943
+ labels: Optional[torch.LongTensor] = None,
944
+ use_cache: Optional[bool] = None,
945
+ output_attentions: Optional[bool] = None,
946
+ output_hidden_states: Optional[bool] = None,
947
+ return_dict: Optional[bool] = None,
948
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
949
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
950
+ output_hidden_states = (
951
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
952
+ )
953
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
954
+
955
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
956
+ outputs = self.model(
957
+ input_ids=input_ids,
958
+ attention_mask=attention_mask,
959
+ position_ids=position_ids,
960
+ past_key_values=past_key_values,
961
+ inputs_embeds=inputs_embeds,
962
+ use_cache=use_cache,
963
+ output_attentions=output_attentions,
964
+ output_hidden_states=output_hidden_states,
965
+ return_dict=return_dict,
966
+ )
967
+
968
+ hidden_states = outputs[0]
969
+ logits = self.lm_head(hidden_states)
970
+ logits = logits.float()
971
+ if self.mup_width_multiplier:
972
+ logits = logits / self.mup_width_multiplier
973
+ logits = logits.masked_fill(self.dummy_tokens_mask, min_value_of_dtype(logits.dtype))
974
+
975
+ loss = None
976
+ if labels is not None:
977
+ # Shift so that tokens < n predict n
978
+ shift_logits = logits[..., :-1, :].contiguous()
979
+ shift_labels = labels[..., 1:].contiguous()
980
+ # Flatten the tokens
981
+ loss_fct = nn.CrossEntropyLoss()
982
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
983
+ shift_labels = shift_labels.view(-1)
984
+ # Enable model parallelism
985
+ shift_labels = shift_labels.to(shift_logits.device)
986
+ loss = loss_fct(shift_logits, shift_labels)
987
+
988
+ if not return_dict:
989
+ output = (logits,) + outputs[1:]
990
+ return (loss,) + output if loss is not None else output
991
+
992
+ return CausalLMOutputWithPast(
993
+ loss=loss,
994
+ logits=logits,
995
+ past_key_values=outputs.past_key_values,
996
+ hidden_states=outputs.hidden_states,
997
+ attentions=outputs.attentions,
998
+ )
999
+
1000
+ def prepare_inputs_for_generation(
1001
+ self,
1002
+ input_ids: torch.LongTensor,
1003
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1004
+ attention_mask: Optional[torch.FloatTensor] = None,
1005
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1006
+ **kwargs
1007
+ ) -> Dict[str, Any]:
1008
+ # only last token for inputs_ids if past is defined in kwargs
1009
+ if past_key_values:
1010
+ input_ids = input_ids[:, -1].unsqueeze(-1)
1011
+
1012
+ position_ids = kwargs.get("position_ids", None)
1013
+
1014
+ if attention_mask is not None and position_ids is None:
1015
+ # create position_ids on the fly for batch generation
1016
+ position_ids = attention_mask.long().cumsum(-1) - 1
1017
+ position_ids.masked_fill_(attention_mask == 0, 1)
1018
+ if past_key_values:
1019
+ position_ids = position_ids[:, -1].unsqueeze(-1)
1020
+ else:
1021
+ position_ids = None
1022
+
1023
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1024
+ if inputs_embeds is not None and past_key_values is None:
1025
+ model_inputs = {"inputs_embeds": inputs_embeds}
1026
+ else:
1027
+ model_inputs = {"input_ids": input_ids}
1028
+
1029
+ model_inputs.update(
1030
+ {
1031
+ "past_key_values": past_key_values,
1032
+ "use_cache": kwargs.get("use_cache"),
1033
+ "position_ids": position_ids,
1034
+ "attention_mask": attention_mask,
1035
+ }
1036
+ )
1037
+ return model_inputs
1038
+
1039
+
1040
+ # Copied from transformers.models.mistral.modeling_mistral.MistralForSequenceClassification with Mistral -> Phi3Small
1041
+ class Phi3SmallForSequenceClassification(Phi3SmallPreTrainedModel):
1042
+ def __init__(self, config):
1043
+ super().__init__(config)
1044
+ self.num_labels = config.num_labels
1045
+ self.model = Phi3SmallModel(config)
1046
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1047
+
1048
+ # Initialize weights and apply final processing
1049
+ self.post_init()
1050
+
1051
+ def get_input_embeddings(self):
1052
+ return self.model.embed_tokens
1053
+
1054
+ def set_input_embeddings(self, value):
1055
+ self.model.embed_tokens = value
1056
+
1057
+
1058
+ def forward(
1059
+ self,
1060
+ input_ids: torch.LongTensor = None,
1061
+ attention_mask: Optional[torch.Tensor] = None,
1062
+ position_ids: Optional[torch.LongTensor] = None,
1063
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1064
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1065
+ labels: Optional[torch.LongTensor] = None,
1066
+ use_cache: Optional[bool] = None,
1067
+ output_attentions: Optional[bool] = None,
1068
+ output_hidden_states: Optional[bool] = None,
1069
+ return_dict: Optional[bool] = None,
1070
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1071
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1072
+
1073
+ transformer_outputs = self.model(
1074
+ input_ids,
1075
+ attention_mask=attention_mask,
1076
+ position_ids=position_ids,
1077
+ past_key_values=past_key_values,
1078
+ inputs_embeds=inputs_embeds,
1079
+ use_cache=use_cache,
1080
+ output_attentions=output_attentions,
1081
+ output_hidden_states=output_hidden_states,
1082
+ return_dict=return_dict,
1083
+ )
1084
+ hidden_states = transformer_outputs[0]
1085
+ logits = self.score(hidden_states)
1086
+
1087
+ if input_ids is not None:
1088
+ batch_size = input_ids.shape[0]
1089
+ else:
1090
+ batch_size = inputs_embeds.shape[0]
1091
+
1092
+ if self.config.pad_token_id is None and batch_size != 1:
1093
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1094
+ if self.config.pad_token_id is None:
1095
+ sequence_lengths = -1
1096
+ else:
1097
+ if input_ids is not None:
1098
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1099
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1100
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1101
+ sequence_lengths = sequence_lengths.to(logits.device)
1102
+ else:
1103
+ sequence_lengths = -1
1104
+
1105
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1106
+
1107
+ loss = None
1108
+ if labels is not None:
1109
+ labels = labels.to(logits.device)
1110
+ if self.config.problem_type is None:
1111
+ if self.num_labels == 1:
1112
+ self.config.problem_type = "regression"
1113
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1114
+ self.config.problem_type = "single_label_classification"
1115
+ else:
1116
+ self.config.problem_type = "multi_label_classification"
1117
+
1118
+ if self.config.problem_type == "regression":
1119
+ loss_fct = nn.MSELoss()
1120
+ if self.num_labels == 1:
1121
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1122
+ else:
1123
+ loss = loss_fct(pooled_logits, labels)
1124
+ elif self.config.problem_type == "single_label_classification":
1125
+ loss_fct = nn.CrossEntropyLoss()
1126
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1127
+ elif self.config.problem_type == "multi_label_classification":
1128
+ loss_fct = nn.BCEWithLogitsLoss()
1129
+ loss = loss_fct(pooled_logits, labels)
1130
+ if not return_dict:
1131
+ output = (pooled_logits,) + transformer_outputs[1:]
1132
+ return ((loss,) + output) if loss is not None else output
1133
+
1134
+ return SequenceClassifierOutputWithPast(
1135
+ loss=loss,
1136
+ logits=pooled_logits,
1137
+ past_key_values=transformer_outputs.past_key_values,
1138
+ hidden_states=transformer_outputs.hidden_states,
1139
+ attentions=transformer_outputs.attentions,
1140
+ )
positional_embedding.py ADDED
@@ -0,0 +1,288 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Orginally Taken verbatim from xformers library
3
+ https://github.com/facebookresearch/xformers/blob/bcb707576c6a80eaf850aa80e8643d3497ec2bc4/xformers/components/positional_embedding/rotary.py
4
+
5
+ The difference is that xformers seems to assume the inputs to be
6
+ (bs, head, seq_len, dim) while we assume (bs, seq_len, head, dim)
7
+
8
+ """
9
+ # Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
10
+ #
11
+ # This source code is licensed under the BSD license found in the
12
+ # LICENSE file in the root directory of this source tree.
13
+
14
+
15
+ # CREDITS: This implementation is inspired by GPT-NeoX https://github.com/EleutherAI/gpt-neox
16
+ # NOTE: Almost the same right now, moving parts to Triton is the next step
17
+
18
+ import math
19
+ from typing import List, Optional, Tuple, Dict, Union
20
+
21
+ import torch
22
+ import dataclasses
23
+ from transformers.utils import logging
24
+
25
+ from transformers import PretrainedConfig
26
+
27
+ is_dacite_available = False
28
+ try:
29
+ import dacite
30
+ is_dacite_available = True
31
+ except ImportError:
32
+ pass
33
+
34
+ logger = logging.get_logger(__name__)
35
+
36
+ @dataclasses.dataclass
37
+ class LongRopeConfig(object):
38
+ short_factor: List[float]
39
+ long_factor: List[float]
40
+ original_max_position_embeddings: int
41
+ type: str = "longrope"
42
+ short_mscale: float = -1
43
+ long_mscale: float = -1
44
+
45
+
46
+ def __post_init__(self):
47
+ assert self.type in ("longrope", "su"), f"Invalid type {self.type} for LongRopeConfig. Expected longrope / su"
48
+
49
+
50
+ @classmethod
51
+ def from_dict(cls, config_dict: Dict[str, Union[float, List[float], int]]) -> "LongRopeConfig":
52
+ if is_dacite_available:
53
+ # Preferred since we can also type check the input
54
+ return dacite.from_dict(data_class=cls, data=config_dict)
55
+ kwargs = {}
56
+ for field in dataclasses.fields(cls):
57
+ if field.name in config_dict:
58
+ if field.init:
59
+ kwargs[field.name] = config_dict[field.name]
60
+ else:
61
+ raise ValueError(f"Field {field.name} is not initiable")
62
+ else:
63
+ if field.default is dataclasses.MISSING:
64
+ raise ValueError(f"Field {field.name} is required")
65
+ extra_keys = set(config_dict.keys()) - set(kwargs.keys())
66
+ if len(extra_keys) > 0:
67
+ for key in extra_keys:
68
+ logger.error(f"Unrecognized key {key} in config_dict")
69
+ raise ValueError(f"Unrecognized keys in config_dict")
70
+ return cls(**kwargs)
71
+
72
+ def rotate_half(x):
73
+ x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
74
+ return torch.cat((-x2, x1), dim=x1.ndim - 1)
75
+
76
+
77
+
78
+ @torch.jit.script
79
+ def apply_rotary_pos_emb(x, cos, sin, seq_dimension: int):
80
+ # NOTE: This could probably be moved to Triton
81
+
82
+ if seq_dimension == 0:
83
+ cos = cos[: x.shape[0], None, None, :]
84
+ sin = sin[: x.shape[0], None, None, :]
85
+ elif seq_dimension == 1:
86
+ # Handle a possible sequence length mismatch in between q and k
87
+ cos = cos[None, : x.shape[1], None, :]
88
+ sin = sin[None, : x.shape[1], None, :]
89
+ elif seq_dimension == 2:
90
+ cos = cos[None, None, : x.shape[2], :]
91
+ sin = sin[None, None, : x.shape[2], :]
92
+
93
+ return (x * cos) + (rotate_half(x) * sin)
94
+
95
+
96
+
97
+ class RotaryEmbedding(torch.nn.Module):
98
+ """
99
+ Adapted from the xformers library
100
+
101
+ The rotary position embeddings from RoFormer_ (Su et. al).
102
+ A crucial insight from the method is that the query and keys are
103
+ transformed by rotation matrices which depend on the relative positions.
104
+ Other implementations are available in the Rotary Transformer repo_ and in
105
+ GPT-NeoX_, GPT-NeoX was an inspiration
106
+ .. _RoFormer: https://arxiv.org/abs/2104.09864
107
+ .. _repo: https://github.com/ZhuiyiTechnology/roformer
108
+ .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
109
+ .. warning: Please note that this embedding is not registered on purpose, as it is transformative
110
+ (it does not create the embedding dimension) and will likely be picked up (imported) on a ad-hoc basis
111
+
112
+ # Arguments
113
+ :param dim_mode: head dimention
114
+ :param max_seq_len:
115
+ :param default_seq_dimension: which dim is the sequence length
116
+ :param dtype: cos/sin dtype
117
+ :param use_fused_kernel: if to use customized fused kernel.
118
+ Note: if used, q, k will be modified inplace. Ok for both forward & backward.
119
+ """
120
+
121
+ def __init__(
122
+ self,
123
+ dim_model: int,
124
+ *,
125
+ max_seq_len: Optional[int] = None,
126
+ dtype: Optional[torch.dtype] = None,
127
+ base=10000,
128
+ position_scale=1,
129
+ device: Optional[torch.device] = None,
130
+ longrope_config: Optional[LongRopeConfig] = None,
131
+ ):
132
+ super().__init__()
133
+ self.base = base
134
+ self.dim_model = dim_model
135
+ self.max_seq_len = max_seq_len
136
+ self.longrope_config = longrope_config
137
+
138
+ if self.is_longrope:
139
+ # Keep the maximum range vector, and slice from it as needed
140
+ self.register_buffer(
141
+ "range_vector",
142
+ torch.arange(max_seq_len, device=device, dtype=torch.float32),
143
+ persistent=False
144
+ )
145
+ self.register_buffer(
146
+ "short_factors",
147
+ torch.tensor(self.longrope_config.short_factor, dtype=torch.float32),
148
+ persistent=False
149
+ )
150
+ self.register_buffer(
151
+ "long_factors",
152
+ torch.tensor(self.longrope_config.long_factor, dtype=torch.float32),
153
+ persistent=False
154
+ )
155
+ else:
156
+ # Generate and save the inverse frequency buffer (non trainable)
157
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim_model, 2).float().to(device) / self.dim_model))
158
+ self.register_buffer("inv_freq", inv_freq)
159
+
160
+ self.position_scale = position_scale
161
+
162
+ if not self.is_longrope:
163
+ dtype = dtype or torch.get_default_dtype()
164
+ self._set_cos_sin_cache(
165
+ seq_len=max_seq_len,
166
+ device=self.inv_freq.device,
167
+ dtype=dtype,
168
+ )
169
+ @property
170
+ def is_longrope(self):
171
+ return self.longrope_config is not None
172
+
173
+ @property
174
+ def original_max_seq_len(self):
175
+ if self.longrope_config is not None:
176
+ return self.longrope_config.original_max_position_embeddings
177
+ logger.warning_once(
178
+ (
179
+ "``original_max_seq_len'' is being accessed, but longrope_config has not been set. "
180
+ "Please only do this if you are sure about the context."
181
+ )
182
+ )
183
+ return self.max_seq_len
184
+
185
+ def get_range_vector(self, seq_len: int, device: torch.device):
186
+ if self.is_longrope:
187
+ assert seq_len < self.range_vector.shape[0], f"Found seq_len {seq_len} greater than max_seq_len {self.range_vector.shape[0]}"
188
+ if self.range_vector.device != device:
189
+ self.range_vector = self.range_vector.to(device)
190
+ return self.range_vector[:seq_len]
191
+ return torch.arange(seq_len, device=device, dtype=torch.float32)
192
+
193
+
194
+ def _calc_mscale(self, scale: torch.Tensor) -> torch.Tensor:
195
+ if scale <= 1.0:
196
+ return 1.0
197
+ return math.sqrt(1 + math.log(scale) / math.log(self.original_max_seq_len))
198
+
199
+ def _set_cos_sin_cache(
200
+ self,
201
+ seq_len: int,
202
+ device: Optional[torch.device] = None,
203
+ dtype: Optional[torch.dtype] = None,
204
+ ) -> None:
205
+ dtype = dtype or torch.get_default_dtype()
206
+ self.max_seq_len_cached = seq_len
207
+ t = (torch.arange(self.max_seq_len_cached, device=device, dtype=torch.float32) * self.position_scale).type_as(self.inv_freq)
208
+ device_type = device.type if device is not None else "cpu"
209
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
210
+ with torch.autocast(device_type=device_type, enabled=False):
211
+ # shape: (seq_len, dim_model // 2)
212
+ freqs = torch.outer(t, self.inv_freq)
213
+ # shape: (seq_len, dim_model)
214
+ emb = torch.cat((freqs, freqs), dim=-1)
215
+ cos = emb.cos()
216
+ sin = emb.sin()
217
+ self.register_buffer("cos_cached", cos.to(dtype), persistent=False)
218
+ self.register_buffer("sin_cached", sin.to(dtype), persistent=False)
219
+
220
+ def forward(
221
+ self, q: torch.Tensor,
222
+ k: torch.Tensor,
223
+ seq_dimension: int = 1,
224
+ seqlen_offset: int = 0,
225
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
226
+ """q, k does not include `seqlen_offset`
227
+ q: Either (bs, seq_len, num_heads, head_dim) or (seq_len, bs, num_heads, head_dim)
228
+ k: Either (bs, seq_len, num_heads, head_dim) or (seq_len, bs, num_heads, head_dim)
229
+ """
230
+ if seq_dimension < 0:
231
+ seq_dimension = k.ndim + seq_dimension
232
+ assert seq_dimension in (0, 1, 2)
233
+ seq_len = k.shape[seq_dimension] + seqlen_offset
234
+
235
+ if self.is_longrope:
236
+ if seq_len > self.original_max_seq_len:
237
+ t = self.get_range_vector(seq_len, device=q.device)
238
+ rescale_factors = self.long_factors.to(q.device)
239
+ long_mscale = self.longrope_config.long_mscale
240
+ mscale = long_mscale if long_mscale > 0 else self._calc_mscale(self.max_seq_len / self.original_max_seq_len)
241
+ else:
242
+ t = self.get_range_vector(self.original_max_seq_len, device=q.device)
243
+ rescale_factors = self.short_factors.to(q.device)
244
+ short_mscale = self.longrope_config.short_mscale
245
+ mscale = short_mscale if short_mscale > 0 else 1.0
246
+ assert rescale_factors.shape == (self.dim_model // 2, ), (
247
+ f"misaligned shape for LongRoPE rescale factors:\n"
248
+ f"\tExpected {(self.dim_model // 2, )}, got {rescale_factors.shape}."
249
+ )
250
+ inv_freq = 1.0 / (rescale_factors * (self.base ** (torch.arange(0, self.dim_model, 2).float().to(q.device) / self.dim_model)))
251
+ device_type = q.device.type if q.device is not None else "cpu"
252
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
253
+ with torch.autocast(device_type=device_type, enabled=False):
254
+ freqs = torch.outer(t, inv_freq)
255
+ emb = torch.cat((freqs, freqs), dim=-1)
256
+ cos = emb.cos() * mscale
257
+ sin = emb.sin() * mscale
258
+ cos_cached = cos.to(q.dtype)
259
+ sin_cached = sin.to(q.dtype)
260
+ else:
261
+ if seq_len > self.max_seq_len_cached:
262
+ self._set_cos_sin_cache(
263
+ seq_len=seq_len,
264
+ device=k.device,
265
+ dtype=k.dtype,
266
+ )
267
+ cos_cached = self.cos_cached
268
+ sin_cached = self.sin_cached
269
+ return (
270
+ apply_rotary_pos_emb(
271
+ q, cos_cached[seqlen_offset:seq_len], sin_cached[seqlen_offset:seq_len], seq_dimension=seq_dimension
272
+ ).to(q.dtype),
273
+ apply_rotary_pos_emb(
274
+ k, cos_cached[seqlen_offset:seq_len], sin_cached[seqlen_offset:seq_len], seq_dimension=seq_dimension
275
+ ).to(k.dtype),
276
+ )
277
+
278
+ @classmethod
279
+ def from_config(cls, config: PretrainedConfig) -> "RotaryEmbedding":
280
+ kwargs = dict(
281
+ dim_model=config.hidden_size // config.num_attention_heads,
282
+ max_seq_len=config.max_position_embeddings,
283
+ base=config.rope_embedding_base,
284
+ position_scale=config.rope_position_scale,
285
+ )
286
+ if config.rope_scaling is not None:
287
+ kwargs["longrope_config"] = LongRopeConfig.from_dict(config.rope_scaling)
288
+ return cls(**kwargs)
tokenization_phi3_small.py ADDED
@@ -0,0 +1,313 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Adapted from https://huggingface.co/Qwen/Qwen-7B-Chat/blob/main/tokenization_qwen.py
2
+ import os
3
+ from typing import Collection, List, Optional, Dict, Set, Tuple, Union
4
+
5
+ from functools import cached_property
6
+
7
+ import base64
8
+
9
+ from transformers import PreTrainedTokenizer, AddedToken, AutoConfig
10
+ from transformers.models.auto.tokenization_auto import get_tokenizer_config
11
+ import tiktoken
12
+
13
+
14
+ """
15
+ This tokenizer is almost identical to tiktoken.get_encoding("cl100k_base")
16
+ with a few additional special tokens to support the ChatML format.
17
+
18
+ TODO(bapatra): Right now, I do not save the special tokens to the vocab file.
19
+ Maybe in the future, that would be useful? Can add that support later.
20
+
21
+ """
22
+
23
+ def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
24
+ with open(tiktoken_bpe_file, "rb") as f:
25
+ contents = f.read()
26
+ return {
27
+ base64.b64decode(token): int(rank)
28
+ for token, rank in (line.split() for line in contents.splitlines() if line)
29
+ }
30
+
31
+ # On the megatron codebase, we pad vocabularies to ensure matrix multiplication is fast.
32
+ # this in turn causes some indices to be empty. We account for these empty indices by adding
33
+ # dummy tokens to the tokenizer.
34
+
35
+ EFFECTIVE_PADDED_VOCAB_SIZE = 100352
36
+ ACTUAL_VOCAB_SIZE = 100276
37
+
38
+
39
+ DUMMY_TOKENS = {
40
+ f"<|dummy_id_{11 + offset}|>": 100276 + offset
41
+ for offset in range(1, EFFECTIVE_PADDED_VOCAB_SIZE - ACTUAL_VOCAB_SIZE)
42
+ }
43
+
44
+ SPECIAL_TOKENS = {
45
+ # tiktoken.get_encoding("cl100k_base")._special_tokens
46
+ '<|endoftext|>': 100257,
47
+ '<|fim_prefix|>': 100258,
48
+ '<|fim_middle|>': 100259,
49
+ '<|fim_suffix|>': 100260,
50
+ # Special tokens for post-training
51
+ "<|system|>": 100261,
52
+ "<|user|>": 100262,
53
+ "<|assistant|>": 100263,
54
+ # Dummy unused tokens
55
+ "<|dummy_id_0|>": 100264,
56
+ "<|dummy_id_1|>": 100265,
57
+ # Special tokens for post-training continued
58
+ "<|end|>": 100266,
59
+ # Some dummy tokens, so that tokenization is contiguous and does not cause issues
60
+ # Note that the 100256th token of tiktoken.get_encoding("cl100k_base") does not
61
+ # actually map to anything. So we use a dummy token here.
62
+ "<|dummy_id_2|>": 100256,
63
+ # Likewise, tokens from 100267 to 100275 are also unused
64
+ "<|dummy_id_3|>": 100267,
65
+ "<|dummy_id_4|>": 100268,
66
+ "<|dummy_id_5|>": 100269,
67
+ "<|dummy_id_6|>": 100270,
68
+ "<|dummy_id_7|>": 100271,
69
+ "<|dummy_id_8|>": 100272,
70
+ "<|dummy_id_9|>": 100273,
71
+ "<|dummy_id_10|>": 100274,
72
+ "<|dummy_id_11|>": 100275,
73
+ # The final end of prompt token
74
+ # (unused, but present as a part of tiktoken.get_encoding("cl100k_base")._special_tokens)
75
+ '<|endofprompt|>': 100276,
76
+ # Dummy tokens to account for padding of the tokenizer
77
+ # We pad to ensure tensor cores are used for vocab multiplication
78
+ **DUMMY_TOKENS
79
+ }
80
+
81
+ class Phi3SmallTokenizer(PreTrainedTokenizer):
82
+ vocab_files_names = {
83
+ "vocab_file": "cl100k_base.tiktoken"
84
+ }
85
+
86
+ model_input_names: List[str] = ["input_ids", "attention_mask"]
87
+ padding_side = "left"
88
+
89
+ def __init__(
90
+ self,
91
+ vocab_file: Optional[str] = None,
92
+ errors: str = "replace",
93
+ **kwargs
94
+ ) -> None:
95
+ # PreTrainedTokenizer's init calls _add_tokens, which in turn checks
96
+ # if the token is present in `self.special_tokens``. Hence instantiating it here.
97
+ # The way Qwen gets around this is by checking against SPECIAL_TOKENS
98
+ # But I think it's better to check against the objects own `special_tokens`
99
+ # in case we eventually want to allow the tokenizer to have special tokens.
100
+ self.special_tokens = SPECIAL_TOKENS
101
+
102
+ super().__init__(**kwargs)
103
+ self.errors = errors
104
+
105
+ base = tiktoken.get_encoding("cl100k_base")
106
+ if vocab_file is None:
107
+ self.mergeable_ranks: Dict[bytes, int] = base._mergeable_ranks
108
+ else:
109
+ self.mergeable_ranks = _load_tiktoken_bpe(vocab_file)
110
+
111
+ self.pat_str = base._pat_str
112
+
113
+ enc = tiktoken.Encoding(
114
+ name="phi3small",
115
+ pat_str=self.pat_str,
116
+ mergeable_ranks=self.mergeable_ranks,
117
+ special_tokens=self.special_tokens,
118
+ )
119
+ self.tokenizer = enc
120
+
121
+ self.decoder: Dict[int, bytes] = {
122
+ v: k for k, v in self.mergeable_ranks.items()
123
+ }
124
+ self.decoder.update({v: k for k, v in self.special_tokens.items()})
125
+
126
+ self.eod_id = self.tokenizer.eot_token
127
+ self._eos_token = self._convert_id_to_token(self.eod_id)
128
+
129
+ # Setting the bos_token to be the same as the eos_token
130
+ # Note that this is **not** the correct thing to do, and is done
131
+ # just so that some of the downstream libraries do not break.
132
+ self._bos_token = self._eos_token
133
+
134
+ # Assign the special tokens to class variables
135
+ self.system_id = self.special_tokens["<|system|>"]
136
+ self.user_id = self.special_tokens["<|user|>"]
137
+ self.assistant_id = self.special_tokens["<|assistant|>"]
138
+ self.end_id = self.special_tokens["<|end|>"]
139
+
140
+ @cached_property
141
+ def dummy_token_indices(self) -> List[int]:
142
+ # There are some additional special tokens in the cl100k_base tokenizer
143
+ # that we do not use. Hence, we also consider them to be dummy tokens.
144
+ additional_tokens = [
145
+ "<|fim_prefix|>",
146
+ "<|fim_middle|>",
147
+ "<|fim_suffix|>",
148
+ "<|endofprompt|>"
149
+ ]
150
+ dummy_token_indices = [index for token, index in self.special_tokens.items() if "dummy_id" in token]
151
+ dummy_token_indices.extend([self.special_tokens[token] for token in additional_tokens])
152
+ return sorted(dummy_token_indices)
153
+
154
+ def __getstate__(self):
155
+ state = self.__dict__.copy()
156
+ del state["tokenizer"]
157
+ return state
158
+
159
+ def __setstate__(self, state):
160
+ self.__dict__ = state
161
+ enc = tiktoken.Encoding(
162
+ name="cl100k_im",
163
+ pat_str=self.pat_str,
164
+ mergeable_ranks=self.mergeable_ranks,
165
+ special_tokens=self.special_tokens,
166
+ )
167
+ self.tokenizer = enc
168
+
169
+ def __len__(self):
170
+ return self.tokenizer.n_vocab
171
+
172
+ @classmethod
173
+ def from_pretrained(
174
+ cls,
175
+ pretrained_model_name_or_path: Union[str, os.PathLike],
176
+ *init_inputs,
177
+ **kwargs,
178
+ ):
179
+ cls_kwargs = kwargs
180
+ # First try to load from the tokenization config if it exists
181
+ tokenization_config = get_tokenizer_config(pretrained_model_name_or_path, **kwargs)
182
+ if tokenization_config:
183
+ cls_kwargs = {
184
+ **tokenization_config,
185
+ **cls_kwargs
186
+ }
187
+ else:
188
+ config = AutoConfig.from_pretrained(pretrained_model_name_or_path, trust_remote_code=True)
189
+ cls_kwargs["model_max_length"] = config.max_position_embeddings
190
+ return cls(**cls_kwargs)
191
+
192
+ def get_vocab(self) -> Dict[Union[str, bytes], int]:
193
+ return {**self.mergeable_ranks, **self.special_tokens}
194
+
195
+ def convert_tokens_to_ids(
196
+ self,
197
+ tokens: Union[bytes, str, List[Union[bytes, str]]]
198
+ ) -> Union[int, List[int]]:
199
+ ids = []
200
+ if isinstance(tokens, (str, bytes)):
201
+ if tokens in self.special_tokens:
202
+ return self.special_tokens[tokens]
203
+ else:
204
+ return self.mergeable_ranks.get(tokens)
205
+ ids: List[int] = []
206
+ for token in tokens:
207
+ ids.append(self.convert_tokens_to_ids(token))
208
+ return ids
209
+
210
+ def _add_tokens(
211
+ self,
212
+ new_tokens: Union[List[str], List[AddedToken]],
213
+ special_tokens: bool = False,
214
+ ) -> int:
215
+ if not special_tokens and new_tokens:
216
+ raise ValueError("Only special tokens can be added to this tokenizer")
217
+ for token in new_tokens:
218
+ surface_form = token.content if isinstance(token, AddedToken) else token
219
+ if surface_form not in self.special_tokens:
220
+ raise ValueError(
221
+ "For now, we do not support unknown special tokens\n"
222
+ "In the future, if there is a need for this, we can add special tokens to the tokenizer\n"
223
+ "starting from rank 100261 - 100263 and then 100266 - 100275.\n"
224
+ "And finally, we can re-construct the enc object back\n"
225
+ )
226
+ return 0
227
+
228
+ def save_vocabulary(self, save_directory: str, **kwargs) -> Tuple[str]:
229
+ file_path = os.path.join(save_directory, "cl100k_base.tiktoken")
230
+ with open(file_path, "w") as f:
231
+ for token, rank in self.mergeable_ranks.items():
232
+ line = base64.b64encode(token).decode("utf-8") + " " + str(rank) + "\n"
233
+ f.write(line)
234
+ return (file_path,)
235
+
236
+ def tokenize(
237
+ self,
238
+ text: str,
239
+ allowed_special: Union[Set, str] = "all",
240
+ disallowed_special: Union[Collection, str] = (),
241
+ **kwargs
242
+ ) -> List[Union[bytes, str]]:
243
+ tokens: List[Union[bytes, str]] = []
244
+ for token_id in self.tokenizer.encode(
245
+ text, allowed_special=allowed_special, disallowed_special=disallowed_special
246
+ ):
247
+ tokens.append(self.decoder[token_id])
248
+ return tokens
249
+
250
+ def convert_tokens_to_string(self, tokens: List[Union[bytes, str]]) -> str:
251
+ """
252
+ Converts a sequence of tokens in a single string.
253
+ """
254
+ text = ""
255
+ temp = b""
256
+ for t in tokens:
257
+ if isinstance(t, str):
258
+ if temp:
259
+ text += temp.decode("utf-8", errors=self.errors)
260
+ temp = b""
261
+ text += t
262
+ elif isinstance(t, bytes):
263
+ temp += t
264
+ else:
265
+ raise TypeError("token should only be of type types or str")
266
+ if temp:
267
+ text += temp.decode("utf-8", errors=self.errors)
268
+ return text
269
+
270
+ @property
271
+ def vocab_size(self):
272
+ return self.tokenizer.n_vocab
273
+
274
+ @property
275
+ def eos_token_id(self) -> int:
276
+ return self.eod_id
277
+
278
+ def _convert_id_to_token(self, index: int) -> Union[bytes, str]:
279
+ """Converts an id to a token, special tokens included"""
280
+ if index in self.decoder:
281
+ return self.decoder[index]
282
+ raise ValueError("unknown ids")
283
+
284
+ def _convert_token_to_id(self, token: Union[bytes, str]) -> int:
285
+ """Converts a token to an id using the vocab, special tokens included"""
286
+ if token in self.special_tokens:
287
+ return self.special_tokens[token]
288
+ if token in self.mergeable_ranks:
289
+ return self.mergeable_ranks[token]
290
+ raise ValueError("unknown token")
291
+
292
+ def _tokenize(self, text: str, **kwargs):
293
+ """
294
+ Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
295
+ vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
296
+ Do NOT take care of added tokens.
297
+ """
298
+ raise NotImplementedError
299
+
300
+ def _decode(
301
+ self,
302
+ token_ids: Union[int, List[int]],
303
+ skip_special_tokens: bool = False,
304
+ errors: str = None,
305
+ **kwargs,
306
+ ) -> str:
307
+ if isinstance(token_ids, int):
308
+ token_ids = [token_ids]
309
+ if skip_special_tokens:
310
+ token_ids = [i for i in token_ids if i < self.eod_id]
311
+ return self.tokenizer.decode(token_ids, errors=errors or self.errors)
312
+
313
+
triton_blocksparse_attention_layer.py ADDED
@@ -0,0 +1,176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ from typing import Optional, Tuple, TypeVar
3
+ import torch.nn as nn
4
+ import torch
5
+ import triton
6
+
7
+ from functools import lru_cache
8
+
9
+
10
+ from .triton_flash_blocksparse_attn import get_local_strided_sparse_attention_op, _get_sparse_attn_mask, blocksparse_flash_attn_padded_fwd, blocksparse_flash_attn_varlen_fwd
11
+
12
+
13
+ Layout = Tuple[torch.LongTensor, torch.LongTensor]
14
+
15
+
16
+ def create_sparse_attn_mask(
17
+ n_heads: int,
18
+ max_seq_len: int,
19
+ max_seq_len_k: int,
20
+ dtype: torch.dtype,
21
+ device: torch.device,
22
+ BLOCK: int,
23
+ local_blocks: int,
24
+ vert_stride: int,
25
+ homo_head: bool,
26
+ return_dense: bool
27
+ ) -> Tuple[Layout, torch.Tensor, Optional[torch.Tensor]]:
28
+ layout, block_sparse_pattern, _ = _get_sparse_attn_mask(
29
+ n_heads=n_heads,
30
+ q_len=max_seq_len,
31
+ N_CTX=max_seq_len_k,
32
+ dtype=dtype,
33
+ device=device,
34
+ BLOCK=BLOCK,
35
+ local_blocks=local_blocks,
36
+ vert_stride=vert_stride,
37
+ homo_head=homo_head,
38
+ return_dense=return_dense
39
+ )
40
+ return layout, block_sparse_pattern
41
+
42
+
43
+ class BlockSparseAttentionLayer(nn.Module):
44
+ def __init__(
45
+ self,
46
+ n_heads: int,
47
+ max_seq_len: int,
48
+ sparse_block_size: int,
49
+ local_blocks: int,
50
+ vert_stride: int,
51
+ kernel_block_size: Optional[int] = None,
52
+ homo_head: bool = False,
53
+ active_head_range: Optional[Tuple[int]] = None
54
+ ) -> None:
55
+ super().__init__()
56
+
57
+ self.n_heads = n_heads
58
+ self.max_seq_len = max_seq_len
59
+ self.sparse_block_size = sparse_block_size
60
+ self.kernel_block_size = kernel_block_size or sparse_block_size
61
+ self.local_blocks = local_blocks
62
+ self.vert_stride = vert_stride
63
+ self.homo_head = homo_head
64
+ self.active_head_range = active_head_range
65
+
66
+ # Internal Parameters used by the layer
67
+ self._sparse_block_mask = None
68
+ self._sparse_layout = None
69
+ self._dtype = None
70
+ self._device = None
71
+
72
+ # TODO(bapatra): Ideally, I'd want to keep all the code for
73
+ # forward to be handled here, and not branch for training and inference.
74
+ # However, that refactor would need a lot of testing. For now, using the
75
+ # training op as is, and will refactor again later.
76
+
77
+ def prune_blocksparse_layout_to_heads(self, h_start: int, h_end: int) -> None:
78
+ self._sparse_block_mask = self._sparse_block_mask[h_start: h_end]
79
+ self._sparse_layout[0] = self._sparse_layout[0][h_start: h_end]
80
+ self._sparse_layout[1] = self._sparse_layout[1][h_start: h_end]
81
+
82
+ def _initialize_internals(
83
+ self,
84
+ dtype: torch.dtype,
85
+ device: torch.device
86
+ ) -> None:
87
+ self._dtype, self._device = dtype, device
88
+ self._sparse_layout, self._sparse_block_mask = create_sparse_attn_mask(
89
+ n_heads=self.n_heads,
90
+ max_seq_len=self.max_seq_len,
91
+ max_seq_len_k=self.max_seq_len,
92
+ dtype=dtype,
93
+ device=device,
94
+ BLOCK=self.sparse_block_size,
95
+ local_blocks=self.local_blocks,
96
+ vert_stride=self.vert_stride,
97
+ homo_head=self.homo_head,
98
+ return_dense=False,
99
+ )
100
+ if (not self.homo_head) and (self.active_head_range is not None):
101
+ assert len(self.active_head_range) == 2, "\"active_head_range\" should be a tuple of start/end index of the heads."
102
+ h_start, h_end = self.active_head_range
103
+ self.prune_blocksparse_layout_to_heads(h_start=h_start, h_end=h_end)
104
+
105
+ assert self.sparse_block_size % self.kernel_block_size == 0, f"The sparse block size must be a multiple of {self.kernel_block_size}. Found {self.sparse_block_size}."
106
+ assert self.kernel_block_size >=16 and math.log2(self.kernel_block_size) % 1 == 0, f"block_size must be power of 2 and at least 16, but {self.kernel_block_size} is given"
107
+ if self.sparse_block_size // self.kernel_block_size > 1:
108
+ _mul = self.sparse_block_size // self.kernel_block_size
109
+ # need to consider if block_m and block_n are different
110
+ self._sparse_block_mask = torch.kron(self._sparse_block_mask, self._sparse_block_mask.new_ones(_mul, _mul))
111
+ num_sparse_blocks = self._sparse_block_mask.size(-1)
112
+ block_causal_mask = torch.arange(0, num_sparse_blocks)[:, None] >= torch.arange(0, num_sparse_blocks)[None]
113
+ self._sparse_block_mask *= block_causal_mask.type_as(self._sparse_block_mask)
114
+
115
+
116
+ def forward(
117
+ self,
118
+ q: torch.Tensor,
119
+ k: torch.Tensor,
120
+ v: torch.Tensor,
121
+ sm_scale: float,
122
+ *,
123
+ # Arguments Related to Block Attention Inference
124
+ left_paddings: Optional[torch.LongTensor] = None,
125
+ seqlens: Optional[torch.LongTensor] = None,
126
+ # Arguements Related to Variable Length Inference
127
+ cu_seqlens_k: Optional[torch.LongTensor] = None,
128
+ cu_seqlens_q: Optional[torch.LongTensor] = None,
129
+ ) -> torch.Tensor:
130
+
131
+ if left_paddings is None and seqlens is None and cu_seqlens_k is None and cu_seqlens_q is None:
132
+ blocksparse_op = get_local_strided_sparse_attention_op(
133
+ n_heads=self.n_heads,
134
+ max_seq_len=self.max_seq_len,
135
+ sparse_block_size=self.sparse_block_size,
136
+ kernel_block_size=self.kernel_block_size,
137
+ local_blocks=self.local_blocks,
138
+ vert_stride=self.vert_stride,
139
+ homo_head=self.homo_head,
140
+ device=q.device,
141
+ inference=not self.training
142
+ )
143
+ return blocksparse_op(q, k, v, sm_scale)
144
+
145
+ assert not torch.is_grad_enabled(), "Variable Length Inference / Batched inference is not supported during training. Please run it in a torch.no_grad() context"
146
+ # First set internals if they have not been set
147
+ if self._sparse_block_mask is None or (self._dtype != q.dtype) or (self._device != q.device):
148
+ self._initialize_internals(dtype=q.dtype, device=q.device)
149
+
150
+ if k.dim() == 3:
151
+ assert cu_seqlens_k is not None
152
+ return blocksparse_flash_attn_varlen_fwd(
153
+ q=q,
154
+ k=k,
155
+ v=v,
156
+ cu_seqlens_k=cu_seqlens_k,
157
+ cu_seqlens_q=cu_seqlens_q,
158
+ sm_scale=sm_scale,
159
+ sparse_layout=self._sparse_layout,
160
+ block_size=self.kernel_block_size,
161
+ max_seqlen=self.max_seq_len,
162
+ )
163
+ if k.dim() == 4:
164
+ assert not (left_paddings is None and seqlens is None), "Either left_paddings or seqlens must be provided for batched inference."
165
+ return blocksparse_flash_attn_padded_fwd(
166
+ q=q,
167
+ k=k,
168
+ v=v,
169
+ sm_scale=sm_scale,
170
+ sparse_layout=self._sparse_layout,
171
+ left_paddings=left_paddings,
172
+ seqlens=seqlens,
173
+ block_size=self.kernel_block_size,
174
+ max_seqlen=self.max_seq_len,
175
+ )
176
+ raise ValueError('q/k/v must be either 3 dim for variable-length input or 4 dim for fixed-length.')
triton_flash_blocksparse_attn.py ADDED
@@ -0,0 +1,1943 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Author: Eric Lin (xihlin)
3
+ """
4
+ """
5
+ ... note(bapatra)::
6
+ This is written as one big file, instead of splitting into logical components because I was running into issues with transformers auto module
7
+ imports when splitting into different files. I've tried keeping the logical partitions demarkated with comment blocks, but it is not ideal.
8
+ In the future, would be really good to revisit this and refactor into a more readable file structure.
9
+
10
+ """
11
+ from typing import TypeVar
12
+ from functools import lru_cache
13
+ import math
14
+ import pytest
15
+ import torch
16
+ import numpy as np
17
+
18
+ import triton
19
+ import triton.language as tl
20
+
21
+ import os
22
+
23
+ import dataclasses
24
+
25
+ Phi3SmallConfig = TypeVar('Phi3SmallConfig')
26
+
27
+ # triton 2.0.0: fail at backward on A100, for the examples, if h_dim=128.
28
+
29
+ # Done
30
+ # 1. strided of qkv
31
+ # 2. seq len not power of 2
32
+ # 3. bf16 with Triton May, 2023
33
+
34
+ # TODO:
35
+ # 1. wip: support non-contiguous backward, also help reduce memory allocation in training (q, k, v split)
36
+ # 2. block sparse with different BLOCK_M, BLOCK_N?
37
+ # 3. for Lq not divided by BLOCK_M, BLOCK_N, only apply mask to K/V on last batch, still need to apply mask on Q.
38
+ # Attempt, fail to compile
39
+ # 4. For 2nd iter of inference, BLOCK_M=1, how to make things work? K/V maynot divided by BLOCK_N.
40
+ # 5. The inner loop can also be paralled via bigger num_stage(better) or on different thread-block (via m/L and atomic update, but this no-comm/sync between blocks)
41
+
42
+
43
+ ###########################################################
44
+ ################### Kernel Parameters #####################
45
+ ###########################################################
46
+
47
+ @dataclasses.dataclass
48
+ class BlockSparseParams(object):
49
+ block_size: int
50
+ kernel_block_size: int
51
+ num_local_blocks: int
52
+ vert_stride: int
53
+ homo_head_pattern: bool = False
54
+
55
+ @classmethod
56
+ def from_config(cls, config: Phi3SmallConfig) -> "BlockSparseParams":
57
+ return cls(
58
+ block_size=config.blocksparse_block_size,
59
+ kernel_block_size=config.blocksparse_triton_kernel_block_size,
60
+ num_local_blocks=config.blocksparse_num_local_blocks,
61
+ vert_stride=config.blocksparse_vert_stride,
62
+ homo_head_pattern=config.blocksparse_homo_head_pattern,
63
+ )
64
+
65
+
66
+ ###########################################################
67
+ ###########################################################
68
+
69
+ ###########################################################
70
+ ################### Utility Functions #####################
71
+ ###########################################################
72
+
73
+ # helper functions for 3D sparse pattern
74
+ # these function are not optimized and very inefficient. Avoid calling them too frequent.
75
+ # currently, it is only called within `get_local_strided_sparse_attention_op`, which is cached.
76
+ def dense_to_crow_col(x):
77
+ ''' Turning a 2D/3D torch tensor (x) to CSR rows/cols indexing.
78
+ param:
79
+ TODO:
80
+ 1. improve efficiency, is it faster if done in CPU, or customize a cuda kernel for it?
81
+ NOTE: col_indices padded -1
82
+ '''
83
+ pad = -1
84
+ dim = x.dim()
85
+ assert x.dim() in (2, 3)
86
+ if x.dim() == 2:
87
+ x = x[None]
88
+ x = [xi.to_sparse_csr() for xi in x]
89
+ crows = torch.vstack([xi.crow_indices() for xi in x])
90
+ cols = [xi.col_indices() for xi in x]
91
+ max_cols = max(len(xi) for xi in cols)
92
+ cols = [torch.cat([xi, pad + xi.new_zeros(max_cols - xi.shape[0])]) for xi in cols]
93
+ cols = torch.vstack(cols)
94
+ if dim == 2:
95
+ crows = crows[0]
96
+ cols = cols[0]
97
+ return crows, cols
98
+
99
+
100
+ def crow_col_to_dense(crows, cols, dtype=torch.float16):
101
+ dim = crows.dim()
102
+ if dim == 1:
103
+ crows = crows[None]
104
+ cols = cols[None]
105
+ device = crows.device
106
+ crows, cols = crows.cpu(), cols.cpu() # faster in cpu
107
+ shape = (crows.shape[0], crows.shape[1] - 1, cols.max() + 1)
108
+ x = torch.zeros(shape, dtype=dtype)
109
+ for i in range(shape[0]):
110
+ for j in range(shape[1]):
111
+ x[i, j, cols[i, crows[i, j]:crows[i, j+1]]] = 1
112
+ if dim == 1:
113
+ x = x[0]
114
+ return x.to(device)
115
+
116
+
117
+ def dense_to_ccol_row(x):
118
+ '''Similar, but to CSC format
119
+ '''
120
+ x = x.transpose(-2, -1)
121
+ return dense_to_crow_col(x)
122
+
123
+
124
+ def ccol_row_to_dense(ccol, rows, dtype=torch.float16):
125
+ return crow_col_to_dense(ccol, rows, dtype).permute(0, 2, 1).contiguous()
126
+
127
+
128
+ def _get_sparse_attn_mask_homo_head(q_len, N_CTX, dtype, device, BLOCK=128, local_blocks=4, vert_stride=4, return_dense=False):
129
+ '''
130
+ :return: a tuple of 3:
131
+ - tuple of crow_indices, col_indices representation of CSR format.
132
+ - block dense mask
133
+ - all token dense mask (be aware that it can be OOM if it is too big) if `return_dense==True`, otherwise, None
134
+ '''
135
+ with torch.no_grad():
136
+ N_BLOCK = triton.cdiv(N_CTX, BLOCK)
137
+ q_pos = torch.arange(N_BLOCK)[:, None]
138
+ k_pos = torch.arange(N_BLOCK)[None]
139
+ mask_vert_strided = (torch.arange(N_BLOCK) + 1) % vert_stride == 0
140
+ block_mask_dense = ((q_pos >= k_pos) & ((q_pos - k_pos < local_blocks) | mask_vert_strided)).to(device).to(dtype)
141
+ N_BLOCK_Q = triton.cdiv(q_len, BLOCK)
142
+ block_mask_dense_output = block_mask_dense[-N_BLOCK_Q:].contiguous().to_sparse_csr()
143
+ if return_dense:
144
+ mask_dense = torch.kron(block_mask_dense, block_mask_dense.new_ones((BLOCK, BLOCK)))
145
+ causal_mask = torch.tril(torch.ones(N_CTX, N_CTX)).type_as(mask_dense)[-q_len:]
146
+ mask_dense = mask_dense[-q_len:, :N_CTX] * causal_mask
147
+ return (block_mask_dense_output.crow_indices(), block_mask_dense_output.col_indices()), block_mask_dense, mask_dense
148
+ else:
149
+ return (block_mask_dense_output.crow_indices(), block_mask_dense_output.col_indices()), block_mask_dense, None
150
+
151
+
152
+ def _get_sparse_attn_mask(n_heads, q_len, N_CTX, dtype, device, BLOCK=128, local_blocks=4, vert_stride=4, homo_head=True, return_dense=False):
153
+ '''
154
+ :return: a tuple of 3:
155
+ - tuple of crow_indices, col_indices representation of CSR format.
156
+ - block dense mask
157
+ - all token dense mask (be aware that it can be OOM if it is too big) if `return_dense==True`, otherwise, None
158
+ '''
159
+ if homo_head:
160
+ with torch.no_grad():
161
+ (crow, col), block_mask_dense, mask_dense = _get_sparse_attn_mask_homo_head(q_len, N_CTX, dtype, device, BLOCK, local_blocks, vert_stride, return_dense)
162
+ crow = crow[None].expand(n_heads, crow.shape[0])
163
+ col = col[None].expand(n_heads, col.shape[0])
164
+ if return_dense:
165
+ mask_dense = mask_dense[None].expand(n_heads, *mask_dense.shape)
166
+ return (crow, col), block_mask_dense, mask_dense
167
+
168
+ with torch.no_grad():
169
+ N_BLOCK = triton.cdiv(N_CTX, BLOCK)
170
+ q_pos = torch.arange(N_BLOCK)[None, :, None]
171
+ k_pos = torch.arange(N_BLOCK)[None, None]
172
+ head_sliding_step = max(1, int(vert_stride / n_heads)) # if vert_stride <= n_heads, rotating the heads
173
+ mask_vert_strided = [(torch.arange(N_BLOCK) + h * head_sliding_step + 1) % vert_stride == 0 for h in range(n_heads)]
174
+ mask_vert_strided = torch.vstack(mask_vert_strided).unsqueeze(1)
175
+ block_mask_dense = ((q_pos >= k_pos) & ((q_pos - k_pos < local_blocks) | mask_vert_strided)).to(device).to(dtype)
176
+ N_BLOCK_Q = triton.cdiv(q_len, BLOCK)
177
+ block_mask_dense_output = block_mask_dense[:, -N_BLOCK_Q:]
178
+ if return_dense:
179
+ mask_dense = torch.kron(block_mask_dense, block_mask_dense.new_ones((BLOCK, BLOCK)))
180
+ causal_mask = torch.tril(torch.ones(N_CTX, N_CTX)).type_as(mask_dense)[-q_len:]
181
+ mask_dense = mask_dense[..., -q_len:, :N_CTX] * causal_mask[None]
182
+ return dense_to_crow_col(block_mask_dense_output), block_mask_dense, mask_dense
183
+ else:
184
+ return dense_to_crow_col(block_mask_dense_output), block_mask_dense, None
185
+
186
+
187
+ def get_sparse_attn_mask(q, N_CTX, *args, **kwargs):
188
+ return _get_sparse_attn_mask(q.size(1), q.size(2), N_CTX, q.dtype, q.device, *args, **kwargs)
189
+
190
+ ###########################################################
191
+ ###########################################################
192
+
193
+ ###########################################################
194
+ ###################### Training Kernels ###################
195
+ ###########################################################
196
+
197
+ # TODO: only apply loading/saving mask on the last iteration for EVEN_N_BLOCK, useful for 1st iteration of inference.
198
+ # Experiment failed inside loop.
199
+ # Another idea: only on saving? load even out of boundary(will it causes illegal access error)?
200
+ @triton.jit
201
+ def _fwd_kernel(
202
+ Q, K, V, sm_scale,
203
+ layout_crow_ptr,
204
+ layout_col_ptr,
205
+ layout_crow_stride_h, layout_crow_stride_m,
206
+ layout_col_stride_h, layout_col_stride_m,
207
+ TMP, L, M, # NOTE: TMP is a scratchpad buffer to workaround a compiler bug. TMP, L, M are assumed to have contiguous layouts
208
+ Out,
209
+ stride_qz, stride_qh, stride_qm, stride_qd,
210
+ stride_kz, stride_kh, stride_kn, stride_kd,
211
+ stride_vz, stride_vh, stride_vn, stride_vd,
212
+ stride_oz, stride_oh, stride_om, stride_od,
213
+ Z, H, N_CTX,
214
+ PAST_LEN,
215
+ Q_ROUNDED_LEN,
216
+ BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
217
+ BLOCK_N: tl.constexpr,
218
+ EVEN_M_BLOCK: tl.constexpr,
219
+ EVEN_N_BLOCK: tl.constexpr,
220
+ INFERENCE: tl.constexpr,
221
+ NUM_DBLOCKS: tl.constexpr,
222
+ ):
223
+ Q_LEN = N_CTX - PAST_LEN
224
+ start_m = tl.program_id(0)
225
+ off_hz = tl.program_id(1)
226
+ off_h = off_hz % H
227
+ off_z = off_hz // H
228
+ Q += off_z * stride_qz + off_h * stride_qh
229
+ K += off_z * stride_kz + off_h * stride_kh
230
+ V += off_z * stride_vz + off_h * stride_vh
231
+ # initialize offsets
232
+ offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
233
+ offs_n = tl.arange(0, BLOCK_N)
234
+ offs_d = tl.arange(0, BLOCK_DMODEL)
235
+ off_q = offs_m[:, None] * stride_qm + offs_d[None, :] * stride_qd
236
+ # off_k = offs_n[:, None] * stride_kn + offs_d[None, :] * stride_kd
237
+ off_k = offs_n[None, :] * stride_kn + offs_d[:, None] * stride_kd
238
+ off_v = offs_n[:, None] * stride_vn + offs_d[None, :] * stride_vd
239
+ # Initialize pointers to Q, K, V
240
+ q_ptrs = Q + off_q
241
+ k_ptrs = K + off_k
242
+ v_ptrs = V + off_v
243
+ # initialize pointer to m and l
244
+ t_ptrs = TMP + off_hz * Q_ROUNDED_LEN + offs_m
245
+ m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
246
+ l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
247
+ acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
248
+ if NUM_DBLOCKS >= 2:
249
+ acc2 = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
250
+
251
+ # load q: it will stay in SRAM throughout
252
+ if EVEN_M_BLOCK:
253
+ q = tl.load(q_ptrs)
254
+ if NUM_DBLOCKS >= 2:
255
+ q2 = tl.load(q_ptrs + BLOCK_DMODEL * stride_qd)
256
+ else:
257
+ q = tl.load(q_ptrs, mask=offs_m[:, None] < Q_LEN)
258
+ if NUM_DBLOCKS >= 2:
259
+ q2 = tl.load(q_ptrs + BLOCK_DMODEL * stride_qd, mask=offs_m[:, None] < Q_LEN)
260
+
261
+ layout_ptr = layout_crow_ptr + off_h * layout_crow_stride_h + start_m * layout_crow_stride_m
262
+ start_l = tl.load(layout_ptr).to(tl.int32)
263
+ end_l = tl.load(layout_ptr + layout_crow_stride_m).to(tl.int32)
264
+
265
+ # loop over k, v and update accumulator
266
+ for col_idx_idx in range(start_l, end_l):
267
+ col_idx = tl.load(layout_col_ptr + off_h * layout_col_stride_h + col_idx_idx * layout_col_stride_m).to(tl.int32)
268
+ start_n = col_idx * BLOCK_N
269
+ # -- compute qk ----
270
+ if EVEN_N_BLOCK:
271
+ k = tl.load(k_ptrs + start_n * stride_kn)
272
+ else:
273
+ k = tl.load(k_ptrs + start_n * stride_kn, mask=offs_n[None, :] + start_n < N_CTX)
274
+ qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
275
+ qk += tl.dot(q, k)
276
+
277
+ if NUM_DBLOCKS >= 2:
278
+ if EVEN_N_BLOCK:
279
+ k = tl.load(k_ptrs + start_n * stride_kn + BLOCK_DMODEL * stride_kd)
280
+ else:
281
+ k = tl.load(k_ptrs + start_n * stride_kn + BLOCK_DMODEL * stride_kd, mask=offs_n[None, :] + start_n < N_CTX)
282
+ qk += tl.dot(q2, k)
283
+
284
+ qk *= sm_scale
285
+ qk += tl.where(offs_m[:, None] + PAST_LEN >= (start_n + offs_n[None, :]), 0, float('-inf'))
286
+ # -- compute m_ij, p, l_ij
287
+ m_ij = tl.max(qk, 1)
288
+ p = tl.exp(qk - m_ij[:, None])
289
+ l_ij = tl.sum(p, 1)
290
+ # -- update m_i and l_i
291
+ m_i_new = tl.maximum(m_i, m_ij)
292
+ alpha = tl.exp(m_i - m_i_new)
293
+ beta = tl.exp(m_ij - m_i_new)
294
+ l_i_new = alpha * l_i + beta * l_ij
295
+ # -- update output accumulator --
296
+ # scale p
297
+ p_scale = beta / l_i_new
298
+ p = p * p_scale[:, None]
299
+ # scale acc
300
+ acc_scale = l_i / l_i_new * alpha
301
+ # tl.store(t_ptrs, acc_scale)
302
+ # acc_scale = tl.load(t_ptrs) # BUG: have to store and immediately load
303
+ acc = acc * acc_scale[:, None]
304
+ if NUM_DBLOCKS >= 2:
305
+ acc2 = acc2 * acc_scale[:, None]
306
+ p = p.to(Q.dtype.element_ty)
307
+ # update acc
308
+ if EVEN_N_BLOCK:
309
+ v = tl.load(v_ptrs + start_n * stride_vn)
310
+ else:
311
+ v = tl.load(v_ptrs + start_n * stride_vn, mask=offs_n[:, None] + start_n < N_CTX)
312
+ acc += tl.dot(p, v)
313
+
314
+ if NUM_DBLOCKS >= 2:
315
+ if EVEN_N_BLOCK:
316
+ v = tl.load(v_ptrs + start_n * stride_vn + BLOCK_DMODEL * stride_vd)
317
+ else:
318
+ v = tl.load(v_ptrs + start_n * stride_vn + BLOCK_DMODEL * stride_vd, mask=offs_n[:, None] + start_n < N_CTX)
319
+ acc2 += tl.dot(p, v)
320
+
321
+ # update m_i and l_i
322
+ l_i = l_i_new
323
+ m_i = m_i_new
324
+
325
+ # rematerialize offsets to save registers
326
+ # start_m = tl.program_id(0)
327
+ # offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
328
+ # write back l and m
329
+ if not INFERENCE:
330
+ l_ptrs = L + off_hz * N_CTX + offs_m
331
+ m_ptrs = M + off_hz * N_CTX + offs_m
332
+ if EVEN_M_BLOCK:
333
+ tl.store(l_ptrs, l_i)
334
+ tl.store(m_ptrs, m_i)
335
+ else:
336
+ tl.store(l_ptrs, l_i, mask=offs_m < Q_LEN)
337
+ tl.store(m_ptrs, m_i, mask=offs_m < Q_LEN)
338
+ # initialize pointers to output
339
+ # offs_n = tl.arange(0, BLOCK_DMODEL)
340
+ off_o = off_z * stride_oz + off_h * stride_oh + offs_m[:, None] * stride_om + offs_d[None, :] * stride_od
341
+ out_ptrs = Out + off_o
342
+ tl.store(out_ptrs, acc, mask=offs_m[:, None] < Q_LEN)
343
+ if NUM_DBLOCKS >= 2:
344
+ tl.store(out_ptrs + BLOCK_DMODEL * stride_od, acc2, mask=offs_m[:, None] < Q_LEN)
345
+
346
+
347
+ ## backward
348
+ @triton.heuristics(
349
+ {
350
+ 'EVEN_M_BLOCK': lambda kwargs: kwargs['N_CTX'] % kwargs['BLOCK_M'] == 0,
351
+ }
352
+ )
353
+ @triton.jit
354
+ def _bwd_preprocess(
355
+ Out, DO, L, # assume contiguous for Out, DO, L, NewDO, Delta layout.
356
+ NewDO, Delta,
357
+ N_CTX,
358
+ BLOCK_M: tl.constexpr, D_HEAD: tl.constexpr,
359
+ EVEN_M_BLOCK: tl.constexpr,
360
+ ):
361
+ off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
362
+ off_d = tl.arange(0, D_HEAD)
363
+ # load
364
+ if EVEN_M_BLOCK:
365
+ o = tl.load(Out + off_m[:, None] * D_HEAD + off_d[None, :]).to(tl.float32)
366
+ do = tl.load(DO + off_m[:, None] * D_HEAD + off_d[None, :]).to(tl.float32)
367
+ else:
368
+ o = tl.load(Out + off_m[:, None] * D_HEAD + off_d[None, :], mask=off_m[:, None] < N_CTX).to(tl.float32)
369
+ do = tl.load(DO + off_m[:, None] * D_HEAD + off_d[None, :], mask=off_m[:, None] < N_CTX).to(tl.float32)
370
+ denom = tl.load(L + off_m).to(tl.float32)
371
+ # compute
372
+ do = do / denom[:, None]
373
+ delta = tl.sum(o * do, axis=1)
374
+ # write-back
375
+ if EVEN_M_BLOCK:
376
+ tl.store(NewDO + off_m[:, None] * D_HEAD + off_d[None, :], do)
377
+ else:
378
+ tl.store(NewDO + off_m[:, None] * D_HEAD + off_d[None, :], do, mask=off_m[:, None] < N_CTX)
379
+ tl.store(Delta + off_m, delta)
380
+
381
+
382
+ # Does not suuport unequal seqlen(q) and seqlen(k)
383
+ @triton.heuristics(
384
+ {
385
+ 'EVEN_M_BLOCK': lambda kwargs: kwargs['N_CTX'] % kwargs['BLOCK_M'] == 0,
386
+ 'EVEN_N_BLOCK': lambda kwargs: kwargs['N_CTX'] % kwargs['BLOCK_N'] == 0,
387
+ }
388
+ )
389
+ @triton.jit
390
+ def _bwd_kernel(
391
+ Q, K, V, sm_scale,
392
+ layout_ccol_ptr,
393
+ layout_row_ptr,
394
+ layout_ccol_stride_h, layout_ccol_stride_m,
395
+ layout_row_stride_h, layout_row_stride_m,
396
+ Out, DO, # assume contigous: Out, Do, DQ, DK, DV, L, M, D, seq(q) == seq(k), with stride_oz, stride_oh, stride_om, stride_od,
397
+ DQ, DK, DV,
398
+ L, M,
399
+ D,
400
+ stride_qz, stride_qh, stride_qm, stride_qd,
401
+ stride_kz, stride_kh, stride_kn, stride_kd,
402
+ stride_vz, stride_vh, stride_vn, stride_vd,
403
+ stride_oz, stride_oh, stride_om, stride_od,
404
+ # stride_dz, stride_dh, stride_dm, stride_dd,
405
+ Z, H, N_CTX,
406
+ num_block,
407
+ BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
408
+ BLOCK_N: tl.constexpr,
409
+ EVEN_M_BLOCK: tl.constexpr,
410
+ EVEN_N_BLOCK: tl.constexpr,
411
+ NUM_DBLOCKS: tl.constexpr,
412
+ ):
413
+ start_n = tl.program_id(0)
414
+ off_hz = tl.program_id(1)
415
+ off_z = off_hz // H
416
+ off_h = off_hz % H
417
+ # offset pointers for batch/head
418
+ Q += off_z * stride_qz + off_h * stride_qh
419
+ K += off_z * stride_kz + off_h * stride_kh
420
+ V += off_z * stride_vz + off_h * stride_vh
421
+ DO += off_z * stride_oz + off_h * stride_oh
422
+ DQ += off_z * stride_oz + off_h * stride_oh
423
+ DK += off_z * stride_oz + off_h * stride_oh
424
+ DV += off_z * stride_oz + off_h * stride_oh
425
+ # Look like this loop can be parallelled
426
+ # for start_n in range(0, num_block):
427
+
428
+ offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
429
+ offs_m = tl.arange(0, BLOCK_M)
430
+ offs_d = tl.arange(0, BLOCK_DMODEL)
431
+ # initialize pointers to value-like data
432
+ k_ptrs = K + (offs_n[:, None] * stride_kn + offs_d[None, :] * stride_kd)
433
+ v_ptrs = V + (offs_n[:, None] * stride_vn + offs_d[None, :] * stride_vd)
434
+
435
+ # pointer to row-wise quantities in value-like data
436
+ D_ptrs = D + off_hz * N_CTX
437
+ m_ptrs = M + off_hz * N_CTX
438
+ # initialize dv amd dk
439
+ dv = tl.zeros([BLOCK_N, BLOCK_DMODEL], dtype=tl.float32)
440
+ dk = tl.zeros([BLOCK_N, BLOCK_DMODEL], dtype=tl.float32)
441
+ # k and v stay in SRAM throughout
442
+ if EVEN_N_BLOCK:
443
+ k = tl.load(k_ptrs)
444
+ v = tl.load(v_ptrs)
445
+ else:
446
+ k = tl.load(k_ptrs, mask=offs_n[:, None] < N_CTX)
447
+ v = tl.load(v_ptrs, mask=offs_n[:, None] < N_CTX)
448
+
449
+ if NUM_DBLOCKS >= 2:
450
+ dv2 = tl.zeros([BLOCK_N, BLOCK_DMODEL], dtype=tl.float32)
451
+ dk2 = tl.zeros([BLOCK_N, BLOCK_DMODEL], dtype=tl.float32)
452
+ if EVEN_N_BLOCK:
453
+ k2 = tl.load(k_ptrs + BLOCK_DMODEL * stride_kd)
454
+ v2 = tl.load(v_ptrs + BLOCK_DMODEL * stride_vd)
455
+ else:
456
+ k2 = tl.load(k_ptrs + BLOCK_DMODEL * stride_kd, mask=offs_n[:, None] < N_CTX)
457
+ v2 = tl.load(v_ptrs + BLOCK_DMODEL * stride_vd, mask=offs_n[:, None] < N_CTX)
458
+
459
+ # loop over rows
460
+
461
+ layout_ptr = layout_ccol_ptr + off_h * layout_ccol_stride_h + start_n * layout_ccol_stride_m
462
+ start_l = tl.load(layout_ptr).to(tl.int32)
463
+ end_l = tl.load(layout_ptr + layout_ccol_stride_m).to(tl.int32)
464
+
465
+ for row_idx_idx in range(start_l, end_l):
466
+ row_idx = tl.load(layout_row_ptr + off_h * layout_row_stride_h + row_idx_idx * layout_row_stride_m).to(tl.int32)
467
+ start_m = row_idx * BLOCK_M
468
+
469
+ # offs_qm = start_m + tl.arange(0, BLOCK_M)
470
+ offs_m_curr = start_m + offs_m
471
+ q_ptrs = Q + (offs_m_curr[:, None] * stride_qm + offs_d[None, :] * stride_qd)
472
+ do_ptrs = DO + (offs_m_curr[:, None] * stride_om + offs_d[None, :] * stride_od)
473
+ dq_ptrs = DQ + (offs_m_curr[:, None] * stride_om + offs_d[None, :] * stride_od)
474
+
475
+ # load q, k, v, do on-chip
476
+ if EVEN_M_BLOCK:
477
+ q = tl.load(q_ptrs)
478
+ else:
479
+ q = tl.load(q_ptrs, mask=offs_m_curr[:, None] < N_CTX)
480
+ # re-compute p = softmax(qk, dim=-1).T
481
+ # NOTE: `do` is pre-divided by `l`; no normalization here
482
+ qk = tl.dot(q, tl.trans(k))
483
+
484
+ if NUM_DBLOCKS >= 2:
485
+ if EVEN_M_BLOCK:
486
+ q2 = tl.load(q_ptrs + BLOCK_DMODEL * stride_qd)
487
+ else:
488
+ q2 = tl.load(q_ptrs + BLOCK_DMODEL * stride_qd, mask=offs_m_curr[:, None] < N_CTX)
489
+ qk += tl.dot(q2, tl.trans(k2))
490
+
491
+ qk += tl.where(offs_m_curr[:, None] >= (offs_n[None, :]), 0, float('-inf'))
492
+
493
+ if EVEN_M_BLOCK:
494
+ m = tl.load(m_ptrs + offs_m_curr)
495
+ else:
496
+ m = tl.load(m_ptrs + offs_m_curr, mask=offs_m_curr < N_CTX)
497
+ p = tl.exp(qk * sm_scale - m[:, None])
498
+
499
+ # compute dv
500
+ if EVEN_M_BLOCK:
501
+ do = tl.load(do_ptrs)
502
+ else:
503
+ do = tl.load(do_ptrs, mask=offs_m_curr[:, None] < N_CTX)
504
+
505
+ if NUM_DBLOCKS >= 2:
506
+ if EVEN_M_BLOCK:
507
+ do2 = tl.load(do_ptrs + BLOCK_DMODEL * stride_od)
508
+ else:
509
+ do2 = tl.load(do_ptrs + BLOCK_DMODEL * stride_od, mask=offs_m_curr[:, None] < N_CTX)
510
+
511
+ dv += tl.dot(tl.trans(p.to(Q.dtype.element_ty)), do)
512
+
513
+ if NUM_DBLOCKS >= 2:
514
+ dv2 += tl.dot(tl.trans(p.to(Q.dtype.element_ty)), do2)
515
+
516
+ # compute dp = dot(v, do)
517
+ if EVEN_M_BLOCK:
518
+ Di = tl.load(D_ptrs + offs_m_curr)
519
+ else:
520
+ Di = tl.load(D_ptrs + offs_m_curr, mask=offs_m_curr < N_CTX)
521
+ dp = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) - Di[:, None]
522
+ dp += tl.dot(do, tl.trans(v))
523
+
524
+ if NUM_DBLOCKS >= 2:
525
+ dp += tl.dot(do2, tl.trans(v2))
526
+
527
+ # compute ds = p * (dp - delta[:, None])
528
+ ds = p * dp * sm_scale
529
+ # compute dk = dot(ds.T, q)
530
+ dk += tl.dot(tl.trans(ds.to(Q.dtype.element_ty)), q)
531
+ if NUM_DBLOCKS >= 2:
532
+ dk2 += tl.dot(tl.trans(ds.to(Q.dtype.element_ty)), q2)
533
+
534
+ # # compute dq
535
+ dq = tl.dot(ds.to(Q.dtype.element_ty), k)
536
+ if EVEN_M_BLOCK:
537
+ tl.atomic_add(dq_ptrs, dq)
538
+ else:
539
+ tl.atomic_add(dq_ptrs, dq, mask=offs_m_curr[:, None] < N_CTX)
540
+
541
+ if NUM_DBLOCKS >= 2:
542
+ dq2 = tl.dot(ds.to(Q.dtype.element_ty), k2)
543
+ dq_ptrs2 = dq_ptrs + BLOCK_DMODEL * stride_od
544
+ if EVEN_M_BLOCK:
545
+ tl.atomic_add(dq_ptrs2, dq2)
546
+ else:
547
+ tl.atomic_add(dq_ptrs2, dq2, mask=offs_m_curr[:, None] < N_CTX)
548
+
549
+ # write-back
550
+ dv_ptrs = DV + (offs_n[:, None] * stride_om + offs_d[None, :] * stride_od)
551
+ dk_ptrs = DK + (offs_n[:, None] * stride_om + offs_d[None, :] * stride_od)
552
+ if EVEN_N_BLOCK:
553
+ tl.store(dv_ptrs, dv)
554
+ tl.store(dk_ptrs, dk)
555
+ else:
556
+ tl.store(dv_ptrs, dv, mask=offs_n[:, None] < N_CTX)
557
+ tl.store(dk_ptrs, dk, mask=offs_n[:, None] < N_CTX)
558
+
559
+ if NUM_DBLOCKS >= 2:
560
+ dv_ptrs2 = dv_ptrs + BLOCK_DMODEL * stride_od
561
+ dk_ptrs2 = dk_ptrs + BLOCK_DMODEL * stride_od
562
+ if EVEN_N_BLOCK:
563
+ tl.store(dv_ptrs2, dv2)
564
+ tl.store(dk_ptrs2, dk2)
565
+ else:
566
+ tl.store(dv_ptrs2, dv2, mask=offs_n[:, None] < N_CTX)
567
+ tl.store(dk_ptrs2, dk2, mask=offs_n[:, None] < N_CTX)
568
+
569
+
570
+
571
+ def _forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale, BLOCK_M, BLOCK_N, num_warps=None, num_stages=1, inference=None, out=None):
572
+ '''
573
+ :param q, k, v: [batch, n_heads, seq_len, model_dim]. len of q is allowed to be different than k/v.
574
+ :param layout_crow_indices, layout_col_indices: same as CSR.crow_indices, and CSR.col_indices used to preresent a sparse tensor.
575
+ Each element represent a block, i.e, all elements in a block to be attentdd, or not attended at all..
576
+ '''
577
+ assert q.shape[-1] == k.shape[-1] == v.shape[-1]
578
+ assert k.shape[2] == v.shape[2]
579
+ o = out if out is not None else torch.empty_like(q).contiguous()
580
+ grid = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1])
581
+
582
+ q_rounded_len = grid[0] * BLOCK_M
583
+ tmp = torch.empty((q.shape[0] * q.shape[1], q_rounded_len), device=q.device, dtype=torch.float32)
584
+
585
+ if inference is None:
586
+ inference = (not q.requires_grad) and (not k.requires_grad) and (not v.requires_grad)
587
+
588
+ if inference:
589
+ L, m = tmp, tmp # no need to use create new tensor
590
+ else:
591
+ L = torch.empty((q.shape[0] * q.shape[1], q_rounded_len), device=q.device, dtype=torch.float32)
592
+ m = torch.empty((q.shape[0] * q.shape[1], q_rounded_len), device=q.device, dtype=torch.float32)
593
+
594
+ if layout_col_indices.dim() == 1:
595
+ layout_crow_indices = layout_crow_indices[None].expand(q.shape[1] , -1)
596
+ layout_col_indices = layout_col_indices[None].expand(q.shape[1] , -1)
597
+
598
+ assert q.shape[-1] in [64, 128]
599
+ BLOCK_DMODEL = 64
600
+
601
+ if num_warps is None:
602
+ MIN_D = min(BLOCK_M, BLOCK_N, BLOCK_DMODEL)
603
+ num_warps = max(1, 2 ** int(math.log2(MIN_D / 16)))
604
+ # print(f'> {BLOCK_M=}, {BLOCK_N=}, {BLOCK_DMODEL=}, {num_warps=}, {num_stages=}')
605
+ else:
606
+ assert math.log2(num_warps) % 1 == 0, f'''"num_warps" should be power of 2, but got {num_warps}.'''
607
+
608
+ ## For debugging:
609
+ # print(f'>> {q.shape=}, {k.shape=}, {BLOCK_M=}, {BLOCK_N=}, {num_warps=}, {BLOCK_DMODEL=}, {q.stride()=}, {k.stride()=}')
610
+ # print(f'>> {layout_crow_indices=}\n{layout_col_indices=}\n {layout_crow_indices.stride()=}, {layout_crow_indices.stride()=}')
611
+ # print(f'> {q.shape=}, {k.shape=}, {layout_crow_indices.shape}, {layout_col_indices.shape}, {layout_crow_indices.stride()}, \
612
+ # {layout_col_indices.stride()}, {layout_crow_indices=}, {layout_col_indices=}')
613
+
614
+ _fwd_kernel[grid](
615
+ q, k, v, sm_scale,
616
+ layout_crow_indices,
617
+ layout_col_indices,
618
+ layout_crow_indices.stride(0), layout_crow_indices.stride(1),
619
+ layout_col_indices.stride(0), layout_col_indices.stride(1),
620
+ tmp, L, m,
621
+ o,
622
+ q.stride(0), q.stride(1), q.stride(2), q.stride(3),
623
+ k.stride(0), k.stride(1), k.stride(2), k.stride(3),
624
+ v.stride(0), v.stride(1), v.stride(2), v.stride(3),
625
+ o.stride(0), o.stride(1), o.stride(2), o.stride(3),
626
+ q.shape[0], q.shape[1], k.shape[2],
627
+ k.shape[2] - q.shape[2],
628
+ q_rounded_len,
629
+ BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N,
630
+ BLOCK_DMODEL=BLOCK_DMODEL,
631
+ EVEN_M_BLOCK=q.shape[2] % BLOCK_M == 0,
632
+ EVEN_N_BLOCK=k.shape[2] % BLOCK_N == 0 ,
633
+ INFERENCE=inference,
634
+ NUM_DBLOCKS=q.shape[-1] // BLOCK_DMODEL,
635
+ num_warps=num_warps,
636
+ num_stages=num_stages,
637
+ )
638
+ if inference:
639
+ L, m = None, None
640
+
641
+ ctx.save_for_backward(q, k, v, o, L, m, layout_crow_indices, layout_col_indices)
642
+ ctx.BLOCK_M = BLOCK_M
643
+ ctx.BLOCK_N = BLOCK_N
644
+ ctx.BLOCK_DMODEL = BLOCK_DMODEL
645
+ # ctx.BLOCK = BLOCK
646
+ ctx.grid = grid
647
+ ctx.sm_scale = sm_scale
648
+ ctx.num_warps = num_warps
649
+ ctx.num_stages = num_stages
650
+ return o
651
+
652
+
653
+ def _backward(ctx, do, layout_ccol_indices, layout_row_indices, dq=None, dk=None, dv=None):
654
+ # q, k, v, o, l, m = ctx.saved_tensors
655
+ q, k, v, o, l, m, layout_crow_indices, layout_col_indices = ctx.saved_tensors
656
+
657
+ ## this following too slow to do online, so get it from inputs, which is cached.
658
+ # layout_ccol_indices, layout_row_indices = dense_to_ccol_row(crow_col_to_dense(ctx.layout_crow_indices, ctx.layout_col_indices))
659
+ # layout_ccol_indices, layout_row_indices = dense_to_ccol_row(crow_col_to_dense(layout_crow_indices, layout_col_indices))
660
+
661
+ if not do.is_contiguous():
662
+ do = do.contiguous()
663
+ ## for debugging
664
+ # print(f'----> do is not contiguous: {do.stride()=}')
665
+ # raise ValueError(f'>>>> output grad is not contiguous: {do.stride()=}')
666
+
667
+ if not o.is_contiguous():
668
+ # TODO: currently only work with contiguous q/k/v.
669
+ raise ValueError(f'--> output is not contiguous: {o.stride()=}. This is maybe caused by q/k/v not being contiguous.')
670
+
671
+
672
+ if layout_ccol_indices.dim() == 1:
673
+ layout_ccol_indices = layout_ccol_indices[None].expand(q.shape[1], -1)
674
+ layout_row_indices = layout_row_indices[None].expand(q.shape[1], -1)
675
+
676
+ # do = do.contiguous()
677
+ dq = dq if dq is not None else torch.zeros_like(q, dtype=torch.float32)
678
+ dk = dk if dk is not None else torch.empty_like(k)
679
+ dv =dv if dv is not None else torch.empty_like(v)
680
+ do_scaled = torch.empty_like(do)
681
+ delta = torch.empty_like(l)
682
+
683
+ assert o.stride() == dq.stride() == dk.stride() == dv.stride() == do_scaled.stride()
684
+
685
+ _bwd_preprocess[(ctx.grid[0] * ctx.grid[1], )](
686
+ o, do, l,
687
+ do_scaled, delta,
688
+ k.shape[2],
689
+ BLOCK_M=ctx.BLOCK_M, D_HEAD=q.shape[-1],
690
+ )
691
+
692
+ grid = (triton.cdiv(q.shape[2], ctx.BLOCK_N), ctx.grid[1])
693
+
694
+ _bwd_kernel[grid](
695
+ q, k, v, ctx.sm_scale,
696
+ layout_ccol_indices,
697
+ layout_row_indices,
698
+ layout_ccol_indices.stride(0), layout_ccol_indices.stride(1),
699
+ layout_row_indices.stride(0), layout_row_indices.stride(1),
700
+ o, do_scaled,
701
+ dq, dk, dv,
702
+ l, m,
703
+ delta,
704
+ q.stride(0), q.stride(1), q.stride(2), q.stride(3),
705
+ k.stride(0), k.stride(1), k.stride(2), k.stride(3),
706
+ v.stride(0), v.stride(1), v.stride(2), v.stride(3),
707
+ o.stride(0), o.stride(1), o.stride(2), o.stride(3),
708
+ q.shape[0], q.shape[1], q.shape[2],
709
+ ctx.grid[0],
710
+ BLOCK_M=ctx.BLOCK_M,
711
+ BLOCK_N=ctx.BLOCK_N,
712
+ BLOCK_DMODEL=ctx.BLOCK_DMODEL,
713
+ NUM_DBLOCKS=q.shape[-1] // ctx.BLOCK_DMODEL,
714
+ num_warps=ctx.num_warps,
715
+ num_stages=1,
716
+ )
717
+ return dq, dk, dv, None, None, None
718
+
719
+
720
+ class _sparse_attention(torch.autograd.Function):
721
+
722
+ @staticmethod
723
+ def forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale):
724
+ BLOCK = 128
725
+ # shape constraints
726
+ return _forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale, BLOCK, BLOCK)
727
+
728
+ @staticmethod
729
+ def backward(ctx, do):
730
+ # q, k, v, o, l, m = ctx.saved_tensors
731
+ q, k, v, o, l, m, layout_crow_indices, layout_col_indices = ctx.saved_tensors
732
+ # TODO: the following is very inefficient.
733
+ # layout_ccol_indices, layout_row_indices = dense_to_ccol_row(crow_col_to_dense(ctx.layout_crow_indices, ctx.layout_col_indices))
734
+ layout_ccol_indices, layout_row_indices = dense_to_ccol_row(crow_col_to_dense(layout_crow_indices, layout_col_indices))
735
+ return _backward(ctx, do, layout_ccol_indices, layout_row_indices)
736
+
737
+
738
+
739
+ # suppressed
740
+ class _sparse_attention_inference(_sparse_attention):
741
+ # TODO: does not work now, as BLOCK_M cannot be <1, as shape for tl.dot cannot be smaller than 16.
742
+ @staticmethod
743
+ def forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale):
744
+ BLOCK = 128
745
+ return _forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale, 1, BLOCK)
746
+
747
+
748
+
749
+ def sparse_attention_factory(BLOCK_M=128, BLOCK_N=128, **kwargs):
750
+ class _sparse_attention_config(_sparse_attention):
751
+ @staticmethod
752
+ def forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale):
753
+ # shape constraints
754
+ return _forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale, BLOCK_M, BLOCK_N,
755
+ **kwargs
756
+ )
757
+ return _sparse_attention_config.apply
758
+
759
+
760
+ @lru_cache(maxsize=8)
761
+ def get_local_strided_sparse_attention_op(
762
+ n_heads: int,
763
+ max_seq_len:int,
764
+ sparse_block_size: int=128,
765
+ local_blocks: int=4,
766
+ vert_stride: int=4,
767
+ homo_head: bool=False,
768
+ dtype=torch.bfloat16,
769
+ device='cuda',
770
+ active_head_range=None,
771
+ verbose=True,
772
+ **kwargs):
773
+ '''
774
+ :param n_heads: total number of attention heads (regardless of tensor/model parallel)
775
+ :param max_seq_len: max sequence length. Need to be bigger or equal to the length of sequences.
776
+ :param sparse_block_size: sparse block size. Default to 128
777
+ :param local_blocks: number of nearest block to attend to. Default to 4, i.e., attention to previous 4xblock_size tokens.
778
+ :param vert_stride: Default to 4. Meaning
779
+ :param homo_head: if all head shared the same pattern.
780
+ :param active_head_range: tuple of start & end of the heads, e..g, (8, 16). Default to use all heads.
781
+ Mainly for tensor/model parallelization where heads are splitted to different GPUs.
782
+ '''
783
+
784
+ if verbose:
785
+ print((f'> new block_sparse_attn op constructed with config: '
786
+ f'{n_heads=}, {max_seq_len=}, {sparse_block_size=}, {local_blocks=}, '
787
+ f'{vert_stride=}, {homo_head=}, {active_head_range=}, {kwargs=}'))
788
+ # assert math.log2(max_seq_len) % 2 == 0, f"max_seq_len should be power of 2 to be more efficient"
789
+ _, block_sparse_pattern, _ = _get_sparse_attn_mask(n_heads, max_seq_len, max_seq_len, dtype, device,
790
+ BLOCK=sparse_block_size, local_blocks=local_blocks,
791
+ vert_stride=vert_stride, homo_head=homo_head,
792
+ return_dense=False)
793
+ if (not homo_head) and (active_head_range is not None):
794
+ assert isinstance(active_head_range, tuple)
795
+ assert len(active_head_range) == 2, '"active_head_range" should be a tuple of start/end index of the heads.'
796
+ h_start, h_end = active_head_range
797
+ block_sparse_pattern = block_sparse_pattern[h_start:h_end]
798
+ # print(block_sparse_pattern)
799
+ return get_sparse_attn_op(block_sparse_pattern, sparse_block_size, **kwargs)
800
+
801
+
802
+ def get_sparse_attn_op(
803
+ sparse_pattern: torch.tensor,
804
+ sparse_block_size: int=128,
805
+ kernel_block_size=128,
806
+ qkv_format='q,k,v',
807
+ **kwargs):
808
+ '''
809
+ Ccreate a block-sparse op with fixed layout. This is to avoid the need to of create CSR layout and convert it to CSC layout everytime,
810
+ which is very inefficient (use python loops on CPU. PyTorch 1.13 supports CSR->CSC, may help.)
811
+
812
+ :param sparse_pattern: sparse pattern of the blocks. Should be `num_blocks(q) x num_blocks(k)` or `n_heads x num_blocks x num_blocks`.
813
+ This tensor should have lower-triangular matrices on the last 2 dimensions for causal attention
814
+ :param sparse_block_size: sparse block size. Default to 128
815
+ :param kernel_block_size: the tile/block size to launch a triton instance. Default to None, i.e., same as `sparse_block_size`
816
+ :param qkv_format: Choices=['q,k,v', 'q, kv', 'qkv'], i.e., separated q,k,v, or kv packed, or qkv packed. Currently, only 'q,k,v' is supported.
817
+
818
+ :param kwargs: keyward arguments passed to `_forward`
819
+ '''
820
+ # assert qkv_format in ('q,k,v', 'q, kv', 'qkv') # to save from running `concat` at forward/backward
821
+
822
+ assert qkv_format == 'q,k,v'
823
+
824
+ if kernel_block_size is None:
825
+ kernel_block_size = sparse_block_size
826
+ else:
827
+ assert sparse_block_size % kernel_block_size == 0, f"The sparse block size must be a multiple of {kernel_block_size}."
828
+ assert kernel_block_size >=16 and math.log2(kernel_block_size) % 1 == 0, f"block_size must be power of 2 and at least 16, but {kernel_block_size} is given"
829
+
830
+
831
+ # print(f'>> {sparse_pattern.shape=}')
832
+ # print(f'{sparse_pattern=}')
833
+ if sparse_block_size // kernel_block_size > 1:
834
+ _mul = sparse_block_size // kernel_block_size
835
+ # need to consider if block_m and block_n are different
836
+ sparse_pattern = torch.kron(sparse_pattern, sparse_pattern.new_ones(_mul, _mul))
837
+ num_sparse_blocks = sparse_pattern.size(-1)
838
+ block_causal_mask = torch.arange(0, num_sparse_blocks)[:, None] >= torch.arange(0, num_sparse_blocks)[None]
839
+ sparse_pattern *= block_causal_mask.type_as(sparse_pattern)
840
+ # print(f'>> after: {sparse_pattern.shape=}')
841
+ # print(f'{sparse_pattern=}')
842
+
843
+ BLOCK_N = kernel_block_size
844
+ NUM_BLOCK = sparse_pattern.size(-1)
845
+ MAX_SEQ_LEN = kernel_block_size * NUM_BLOCK
846
+
847
+ grand_layout_crow_indices, grand_layout_col_indices = dense_to_crow_col(sparse_pattern)
848
+ # sparse csc layout for backward
849
+ grand_layout_ccol_indices, grand_layout_row_indices = dense_to_ccol_row(sparse_pattern)
850
+
851
+
852
+ # cache GPU backward layout. limit the size to avoid OOM as time goes.
853
+ # For inference, one only needs to cache one block as sequence length always increases
854
+ # Therefore, this cache needs to be reconstructed per every `block_size`-steps.
855
+ # For training/finetune, set to 8 to increase cache hit.
856
+ # Given an input, the block_len will be the same for all layers, so cache is very helpful.
857
+
858
+ max_cache_size = 1 if kwargs.get('inference', False) else 8
859
+
860
+ @lru_cache(maxsize=max_cache_size)
861
+ def get_backward_layout_by_block_len(block_len):
862
+ assert block_len <= NUM_BLOCK
863
+ if block_len == NUM_BLOCK:
864
+ return (grand_layout_ccol_indices, grand_layout_row_indices)
865
+ return dense_to_ccol_row(sparse_pattern[..., :block_len, :block_len])
866
+
867
+ # for debugging
868
+ # if not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0:
869
+ # print(f'> {sparse_pattern.cpu().tolist()=}')
870
+ # print('----')
871
+ # print(f'> {grand_layout_crow_indices.cpu().tolist()=}\n{grand_layout_col_indices.cpu().tolist()=}')
872
+
873
+
874
+ # q, k, v separated
875
+ class _q_k_v_sparse_attention(torch.autograd.Function):
876
+ @staticmethod
877
+ def forward(ctx, q, k, v, sm_scale):
878
+ # assert q.shape[2] == 1 or q.shape[2] == k.shape[2]
879
+ # shape constraints
880
+ MIN_BLOCK_SIZE = 16
881
+ assert BLOCK_N >= MIN_BLOCK_SIZE
882
+ BLOCK_M = 16 if q.shape[2] <= 16 else BLOCK_N # BLOCK_M has to be power of 2
883
+
884
+ # this following code only works for causal attention
885
+ K_BLOCKS = triton.cdiv(k.shape[2], kernel_block_size)
886
+ # Q_START_BLOCKS = K_BLOCKS - 1 if q.shape[2] == 1 else 0
887
+ Q_START_BLOCKS = K_BLOCKS - triton.cdiv(q.shape[2], BLOCK_N)
888
+ # print(Q_START_BLOCKS, K_BLOCKS)
889
+
890
+ layout_crow_indices = grand_layout_crow_indices[..., Q_START_BLOCKS:K_BLOCKS+1]
891
+ layout_col_indices = grand_layout_col_indices
892
+ # print(BLOCK_M, BLOCK_N, Q_START_BLOCKS, K_BLOCKS+1, layout_crow_indices, layout_col_indices)
893
+
894
+ return _forward(ctx, q, k, v, layout_crow_indices, layout_col_indices, sm_scale, BLOCK_M, BLOCK_N,
895
+ **kwargs
896
+ )
897
+ @staticmethod
898
+ def backward(ctx, do):
899
+ q, k = ctx.saved_tensors[:2]
900
+ assert q.shape[2] == k.shape[2], '> currently backward can only be done if q, k have same length. Contact @EricLin if you need it.'
901
+ # assume q, k have same length
902
+ block_len = triton.cdiv(do.shape[2], kernel_block_size)
903
+ backward_layout = get_backward_layout_by_block_len(block_len)
904
+ return _backward(ctx, do, *backward_layout)[:4]
905
+
906
+
907
+ def _q_k_v_sparse_attention_fn(*args):
908
+ return _q_k_v_sparse_attention.apply(*args)
909
+
910
+ _q_k_v_sparse_attention_fn.sparse_pattern = sparse_pattern
911
+ _q_k_v_sparse_attention_fn.grand_layout_crow_indices = grand_layout_crow_indices
912
+ _q_k_v_sparse_attention_fn.grand_layout_col_indices = grand_layout_col_indices
913
+ _q_k_v_sparse_attention_fn.grand_layout_ccol_indices = grand_layout_ccol_indices
914
+ _q_k_v_sparse_attention_fn.grand_layout_row_indices = grand_layout_row_indices
915
+
916
+ return _q_k_v_sparse_attention_fn
917
+
918
+ ###########################################################
919
+ ###########################################################
920
+
921
+ ###########################################################
922
+ ################ Inference Kernels ########################
923
+ ###########################################################
924
+
925
+ def blocksparse_flash_attn_padded_fwd(
926
+ q, k, v, # (batch, tokens, n_heads, head_size)
927
+ sm_scale,
928
+ sparse_layout,
929
+ *,
930
+ left_paddings = None,
931
+ seqlens = None,
932
+ block_size = 64,
933
+ max_seqlen = None
934
+ ):
935
+ '''
936
+ q, k, v: (batch, tokens, n_heads/n_kv_heads, head_size)
937
+ left_paddings: (batch, ), number of left paddings for each sample.
938
+ seqlens: can be used to specify right padding. No need to specify if left_paddings is used.
939
+ '''
940
+ batches, q_len, n_heads, head_size = q.shape
941
+ _, k_len, n_kv_heads, _ = k.shape
942
+
943
+
944
+ assert q.dim() == k.dim() == v.dim() == 4
945
+ assert q.size(2) % k.size(2) == 0
946
+ assert q.size(0) == k.size(0) and q.size(3) == k.size(3)
947
+ assert k.shape == v.shape # TODO: allow diff head_size for k, v
948
+ assert q_len == 1 or q_len == k_len, \
949
+ f'q length can only 1 for decoding for same as k length for prefilling.'
950
+
951
+ q_k_ratio = q.size(2) // k.size(2)
952
+
953
+ if max_seqlen:
954
+ assert k.size(1) <= max_seqlen, f'k has seqlen {k.size(1)} while max sequence length is set to {max_seqlen}.'
955
+
956
+ # paddings always has zero output, a little slower than using empty
957
+ out = q.new_zeros(q.shape)
958
+
959
+ layout_crow_indices, layout_col_indices = sparse_layout
960
+ block_d = triton.next_power_of_2(head_size)
961
+
962
+ if left_paddings is not None:
963
+ assert left_paddings.shape == (batches,)
964
+ k_batch_starts = left_paddings.to(q.device, dtype=torch.int32).contiguous()
965
+ else:
966
+ k_batch_starts = torch.zeros((batches,), dtype=torch.int32, device=q.device)
967
+
968
+ if seqlens is not None:
969
+ k_batch_ends = k_batch_starts + seqlens.type_as(k_batch_starts)
970
+ assert k_batch_ends.max() <= k_len, f'seqlens (+left_paddings if any) exceeds seqlen.'
971
+ else:
972
+ k_batch_ends = torch.zeros_like(k_batch_starts) + k_len
973
+
974
+ if q_len == 1:
975
+ q_batch_starts = torch.zeros_like(k_batch_starts)
976
+ q_batch_ends = q_batch_starts + 1
977
+ else:
978
+ q_batch_starts = k_batch_starts
979
+ q_batch_ends = k_batch_ends
980
+
981
+ # switch to use cpu to avoid too many kernel lauch when iterate over
982
+ q_lens = (q_batch_ends - q_batch_starts).cpu()
983
+ n_blocks = (q_lens + block_size - 1) // block_size
984
+
985
+ q_batch_ids = torch.tensor([i for i, n in enumerate(n_blocks) for _ in range(n)],
986
+ dtype=q_batch_starts.dtype,
987
+ device=q_batch_starts.device)
988
+ q_start_sids = torch.tensor([i * block_size for n in n_blocks for i in range(n)],
989
+ dtype=q_batch_starts.dtype,
990
+ device=q_batch_starts.device)
991
+
992
+ grid = (len(q_start_sids), n_heads)
993
+
994
+ _fwd_kernel_batch_inference[grid](
995
+ q, k, v, out,
996
+ sm_scale,
997
+ q_batch_starts,
998
+ q_batch_ends,
999
+ k_batch_starts,
1000
+ k_batch_ends,
1001
+ q_batch_ids,
1002
+ q_start_sids,
1003
+
1004
+ *q.stride(),
1005
+ *k.stride(),
1006
+ *v.stride(),
1007
+ *out.stride(),
1008
+
1009
+ layout_crow_indices,
1010
+ layout_col_indices,
1011
+ *layout_crow_indices.stride(),
1012
+ *layout_col_indices.stride(),
1013
+
1014
+ q_k_ratio,
1015
+ HAS_BATCH_DIM = True,
1016
+ D_HEAD = head_size,
1017
+ BLOCK_M = block_size,
1018
+ BLOCK_N = block_size,
1019
+ BLOCK_D = block_d,
1020
+ BLOCK_M_LOADING = 16 if q_len == 1 else block_size, # smaller for decoding
1021
+ EVEN_D = block_d == head_size,
1022
+ num_warps = 1 if q_len == 1 else 4,
1023
+ num_stages = 3
1024
+ )
1025
+
1026
+ return out
1027
+
1028
+
1029
+ def blocksparse_flash_attn_varlen_fwd(
1030
+ q, k, v, # (#tokens, n_heads, head_size)
1031
+ cu_seqlens_k,
1032
+ cu_seqlens_q,
1033
+ sm_scale,
1034
+ sparse_layout,
1035
+ *,
1036
+ block_size=64,
1037
+ max_seqlen = None
1038
+ ):
1039
+ # split q to blocks
1040
+ _, n_heads, head_size = q.shape
1041
+ batch_size = cu_seqlens_k.size(0) - 1
1042
+
1043
+
1044
+ # print(f'> {q.shape=}, {k.shape=}')
1045
+ assert q.dim() == k.dim() == v.dim() == 3
1046
+ assert q.size(1) % k.size(1) == 0
1047
+ assert q.size(2) == k.size(2)
1048
+ assert k.shape == v.shape # TODO: allow diff head_size for k, v
1049
+ assert cu_seqlens_k.dim() == 1
1050
+
1051
+ q_k_ratio = q.size(1) // k.size(1)
1052
+
1053
+ if cu_seqlens_q is None:
1054
+ if q.size(0) == batch_size: # decoding only
1055
+ cu_seqlens_q = torch.arange(0, batch_size + 1,
1056
+ dtype=cu_seqlens_k.dtype,
1057
+ device=cu_seqlens_k.device)
1058
+ elif q.size(0) == k.size(0):
1059
+ cu_seqlens_q = cu_seqlens_k
1060
+ else:
1061
+ raise ValueError('cu_seqlens_q must be specified if it is mix of prefilling and decoding.')
1062
+ else:
1063
+ assert cu_seqlens_k.size(0) == cu_seqlens_q.size(0)
1064
+
1065
+ # switch to use cpu to avoid too many kernel lauch when iterate over
1066
+ q_lens = (cu_seqlens_q[1:] - cu_seqlens_q[:-1]).cpu()
1067
+ k_lens = (cu_seqlens_k[1:] - cu_seqlens_k[:-1]).cpu()
1068
+
1069
+ assert torch.logical_or(q_lens == 1, k_lens == q_lens).all(), \
1070
+ 'length of q should either be 1 (decoding) or same as k (prefilling).'
1071
+
1072
+ if max_seqlen:
1073
+ assert k_lens.max() <= max_seqlen
1074
+
1075
+ n_blocks = (q_lens + block_size - 1) // block_size
1076
+
1077
+ q_batch_ids = torch.tensor([i for i, n in enumerate(n_blocks) for _ in range(n)],
1078
+ dtype=cu_seqlens_q.dtype,
1079
+ device=cu_seqlens_q.device)
1080
+ q_start_sids = torch.tensor([i * block_size for n in n_blocks for i in range(n)],
1081
+ dtype=cu_seqlens_q.dtype,
1082
+ device=cu_seqlens_q.device)
1083
+
1084
+
1085
+ out = q.new_empty(q.shape)
1086
+ cu_seqlens_q = cu_seqlens_q.contiguous()
1087
+ cu_seqlens_k = cu_seqlens_k.contiguous()
1088
+
1089
+ layout_crow_indices, layout_col_indices = sparse_layout
1090
+ block_d = triton.next_power_of_2(head_size)
1091
+
1092
+ decoding_only = (q_lens == 1).all()
1093
+
1094
+ grid = (len(q_start_sids), n_heads)
1095
+
1096
+ _fwd_kernel_batch_inference[grid](
1097
+ q, k, v, out,
1098
+ sm_scale,
1099
+ cu_seqlens_q[:-1],
1100
+ cu_seqlens_q[1:],
1101
+ cu_seqlens_k[:-1],
1102
+ cu_seqlens_k[1:],
1103
+ q_batch_ids,
1104
+ q_start_sids,
1105
+
1106
+ 0, *q.stride(),
1107
+ 0, *k.stride(),
1108
+ 0, *v.stride(),
1109
+ 0, *out.stride(),
1110
+
1111
+ layout_crow_indices,
1112
+ layout_col_indices,
1113
+ *layout_crow_indices.stride(),
1114
+ *layout_col_indices.stride(),
1115
+
1116
+ q_k_ratio,
1117
+ HAS_BATCH_DIM = False,
1118
+ D_HEAD = head_size,
1119
+ BLOCK_M = block_size,
1120
+ BLOCK_N = block_size,
1121
+ BLOCK_D = block_d,
1122
+ BLOCK_M_LOADING = 16 if decoding_only else block_size, # smaller for decoding
1123
+ EVEN_D = block_d == head_size,
1124
+ num_warps = 1 if decoding_only else 4,
1125
+ num_stages = 3
1126
+ )
1127
+
1128
+ return out
1129
+
1130
+
1131
+ @triton.jit
1132
+ def _fwd_kernel_inner(
1133
+ acc, l_i, m_i,
1134
+ q, Q,
1135
+ k_block_col_idx,
1136
+ layout_col_ptr,
1137
+ layout_col_stride_h, layout_col_stride_m,
1138
+ k_ptrs,
1139
+ v_ptrs,
1140
+ off_h, offs_m, offs_n, offs_d,
1141
+ stride_kt, stride_vt,
1142
+ sm_scale,
1143
+ k_seqlen,
1144
+ past_len,
1145
+ LAST_K_BLOCK: tl.constexpr,
1146
+ BLOCK_M_LOADING: tl.constexpr,
1147
+ BLOCK_N: tl.constexpr,
1148
+ D_HEAD: tl.constexpr,
1149
+ EVEN_D: tl.constexpr,
1150
+ M_LT_N: tl.constexpr
1151
+ ):
1152
+ k_block_id = tl.load(layout_col_ptr + off_h * layout_col_stride_h + k_block_col_idx * layout_col_stride_m).to(tl.int32)
1153
+ start_n = k_block_id * BLOCK_N
1154
+ # -- compute qk ----
1155
+ if LAST_K_BLOCK:
1156
+ if EVEN_D:
1157
+ k = tl.load(k_ptrs + start_n * stride_kt,
1158
+ mask=offs_n[None, :] + start_n < k_seqlen)
1159
+ else:
1160
+ # mask = mask & (offs_d[:, ])
1161
+ k = tl.load(k_ptrs + start_n * stride_kt,
1162
+ mask=(offs_n[None, :] + start_n < k_seqlen) & (offs_d[:, None] < D_HEAD))
1163
+ else:
1164
+ if EVEN_D:
1165
+ k = tl.load(k_ptrs + start_n * stride_kt)
1166
+ else:
1167
+ k = tl.load(k_ptrs + start_n * stride_kt,
1168
+ mask=offs_d[:, None] < D_HEAD)
1169
+
1170
+
1171
+ qk = tl.zeros([BLOCK_M_LOADING, BLOCK_N], dtype=tl.float32)
1172
+ qk += tl.dot(q, k)
1173
+
1174
+ qk *= sm_scale
1175
+
1176
+ # the following is needed only when LAST_K_BLOCK or BLOCK_M < BLOCK_N
1177
+ if LAST_K_BLOCK | M_LT_N:
1178
+ qk += tl.where(offs_m[:, None] + past_len >= (start_n + offs_n[None, :]), 0, float('-inf'))
1179
+
1180
+ # -- compute m_ij, p, l_ij
1181
+ m_ij = tl.max(qk, 1)
1182
+ p = tl.exp(qk - m_ij[:, None])
1183
+
1184
+ l_ij = tl.sum(p, 1)
1185
+ # -- update m_i and l_i
1186
+ m_i_new = tl.maximum(m_i, m_ij)
1187
+ alpha = tl.exp(m_i - m_i_new)
1188
+ beta = tl.exp(m_ij - m_i_new)
1189
+ l_i_new = alpha * l_i + beta * l_ij
1190
+ # -- update output accumulator --
1191
+ # scale p
1192
+ p_scale = beta / l_i_new
1193
+ p = p * p_scale[:, None]
1194
+ # scale acc
1195
+ acc_scale = l_i / l_i_new * alpha
1196
+ acc = acc * acc_scale[:, None]
1197
+
1198
+ p = p.to(Q.dtype.element_ty)
1199
+ # update acc
1200
+ if LAST_K_BLOCK:
1201
+ if EVEN_D:
1202
+ v = tl.load(v_ptrs + start_n * stride_vt,
1203
+ mask=offs_n[:, None] + start_n < k_seqlen)
1204
+ else:
1205
+ v = tl.load(v_ptrs + start_n * stride_vt,
1206
+ mask=(offs_n[:, None] + start_n < k_seqlen) & (offs_d[None, :] < D_HEAD))
1207
+ else:
1208
+ if EVEN_D:
1209
+ v = tl.load(v_ptrs + start_n * stride_vt)
1210
+ else:
1211
+ v = tl.load(v_ptrs + start_n * stride_vt,
1212
+ mask=offs_d[None, :] < D_HEAD)
1213
+
1214
+ acc += tl.dot(p, v)
1215
+ # update m_i and l_i
1216
+ l_i = l_i_new
1217
+ m_i = m_i_new
1218
+ return acc, l_i, m_i
1219
+
1220
+
1221
+ @triton.heuristics(
1222
+ {
1223
+ 'M_LT_N': lambda kwargs: kwargs['BLOCK_M'] < kwargs['BLOCK_N'],
1224
+ }
1225
+ )
1226
+ @triton.jit
1227
+ def _fwd_kernel_batch_inference(
1228
+ Q, K, V, Out,
1229
+
1230
+ sm_scale,
1231
+ q_batch_starts,
1232
+ q_batch_ends,
1233
+ k_batch_starts,
1234
+ k_batch_ends,
1235
+ q_batch_ids,
1236
+ q_start_sids,
1237
+
1238
+ stride_qb, stride_qt, stride_qh, stride_qd,
1239
+ stride_kb, stride_kt, stride_kh, stride_kd,
1240
+ stride_vb, stride_vt, stride_vh, stride_vd,
1241
+ stride_ob, stride_ot, stride_oh, stride_od,
1242
+
1243
+ layout_crow_ptr,
1244
+ layout_col_ptr,
1245
+ layout_crow_stride_h, layout_crow_stride_m,
1246
+ layout_col_stride_h, layout_col_stride_m,
1247
+
1248
+ q_k_ratio,
1249
+
1250
+ HAS_BATCH_DIM: tl.constexpr,
1251
+ D_HEAD: tl.constexpr,
1252
+ BLOCK_M: tl.constexpr,
1253
+ BLOCK_N: tl.constexpr,
1254
+ BLOCK_D: tl.constexpr,
1255
+ BLOCK_M_LOADING: tl.constexpr,
1256
+ EVEN_D: tl.constexpr,
1257
+ M_LT_N: tl.constexpr
1258
+ ):
1259
+ '''
1260
+ NOTATION:
1261
+ pid: position id
1262
+ sid: storage id
1263
+ sbid: storage block id
1264
+ pbid: position block id
1265
+ offs_m, offs_n: storage offsets of m-dim(q, row) and n-dim(k, col)
1266
+
1267
+ q and blocks in KV needs to be contiguous
1268
+
1269
+ Arguments:
1270
+ kv_seq_lens: for compute past_len
1271
+ kv_storage_offsets: similar to block_tables in vllm, except it is dynamic.
1272
+ TODO: fix this
1273
+
1274
+ TODO:
1275
+ Optimize grouped-attn
1276
+
1277
+ CUDA graph support issue
1278
+ 1. grid is dynamic: vllm set up multiple cuda graph in decoding phase, with diff max token size (16, 32, ...)
1279
+ since we mix prompt and decoing phase here, it can be more complex.
1280
+ need to set up diff cuda-graph for diff (off_zm, off_z)
1281
+
1282
+ # indeed, q_batch_ids can be padded to maximum number of grid[0], i.e., assume all decoding
1283
+ therefore, cu_seqlens_q, kv_seq_lens
1284
+
1285
+ '''
1286
+ off_zm = tl.program_id(0)
1287
+ off_h = tl.program_id(1)
1288
+
1289
+ off_h_for_kv = off_h // q_k_ratio
1290
+ off_z = tl.load(q_batch_ids + off_zm).to(tl.int32) # [0, 0, 0, 1]
1291
+ q_start_sid = tl.load(q_start_sids + off_zm)
1292
+ start_m = q_start_sid // BLOCK_M
1293
+
1294
+ if HAS_BATCH_DIM:
1295
+ Q += off_z * stride_qb
1296
+ K += off_z * stride_kb
1297
+ V += off_z * stride_vb
1298
+ Out += off_z * stride_ob
1299
+
1300
+ offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M_LOADING)
1301
+ offs_n = tl.arange(0, BLOCK_N)
1302
+ offs_d = tl.arange(0, BLOCK_D)
1303
+
1304
+ q_cu_start = tl.load(q_batch_starts + off_z).to(tl.int32)
1305
+ q_seqlen = tl.load(q_batch_ends + off_z).to(tl.int32) - q_cu_start
1306
+
1307
+ k_cu_start = tl.load(k_batch_starts + off_z).to(tl.int32)
1308
+ k_seqlen = tl.load(k_batch_ends + off_z).to(tl.int32) - k_cu_start
1309
+
1310
+ past_len = k_seqlen - q_seqlen
1311
+
1312
+ Q += q_cu_start * stride_qt + off_h * stride_qh
1313
+ K += k_cu_start * stride_kt + off_h_for_kv * stride_kh
1314
+ V += k_cu_start * stride_vt + off_h_for_kv * stride_vh
1315
+ Out += q_cu_start * stride_ot + off_h * stride_oh
1316
+
1317
+ q_pbid = (past_len + q_start_sid) // BLOCK_M
1318
+
1319
+ if EVEN_D:
1320
+ q = tl.load(Q + offs_m[:, None] * stride_qt + offs_d[None, :] * stride_qd,
1321
+ mask=offs_m[:, None] < q_seqlen)
1322
+ else:
1323
+ q = tl.load(Q + offs_m[:, None] * stride_qt + offs_d[None, :] * stride_qd,
1324
+ mask=(offs_m[:, None] < q_seqlen) & (offs_d[None, :] < D_HEAD),
1325
+ other=0)
1326
+
1327
+ sparse_crow_ptr = layout_crow_ptr + off_h * layout_crow_stride_h + q_pbid * layout_crow_stride_m
1328
+
1329
+ # TODO: load at once, supported in new Triton
1330
+ k_block_start = tl.load(sparse_crow_ptr).to(tl.int32)
1331
+ k_block_end = tl.load(sparse_crow_ptr + 1).to(tl.int32)
1332
+
1333
+ m_i = tl.zeros([BLOCK_M_LOADING], dtype=tl.float32) - float('inf')
1334
+ l_i = tl.zeros([BLOCK_M_LOADING], dtype=tl.float32)
1335
+ acc = tl.zeros([BLOCK_M_LOADING, BLOCK_D], dtype=tl.float32)
1336
+
1337
+ k_ptrs = K + offs_n[None, :] * stride_kt + offs_d[:, None] * stride_kd
1338
+ v_ptrs = V + offs_n[:, None] * stride_vt + offs_d[None, :] * stride_vd
1339
+
1340
+ for k_block_col_idx in range(k_block_start, k_block_end - 1):
1341
+ acc, l_i, m_i = _fwd_kernel_inner(
1342
+ acc, l_i, m_i,
1343
+ q, Q,
1344
+ k_block_col_idx,
1345
+ layout_col_ptr,
1346
+ layout_col_stride_h, layout_col_stride_m,
1347
+ k_ptrs,
1348
+ v_ptrs,
1349
+ off_h, offs_m, offs_n, offs_d,
1350
+ stride_kt, stride_vt,
1351
+ sm_scale,
1352
+ k_seqlen,
1353
+ past_len,
1354
+ False,
1355
+ BLOCK_M_LOADING,
1356
+ BLOCK_N,
1357
+ D_HEAD,
1358
+ EVEN_D,
1359
+ M_LT_N
1360
+ )
1361
+
1362
+ acc, l_i, m_i = _fwd_kernel_inner(
1363
+ acc, l_i, m_i,
1364
+ q, Q,
1365
+ k_block_end - 1,
1366
+ layout_col_ptr,
1367
+ layout_col_stride_h, layout_col_stride_m,
1368
+ k_ptrs,
1369
+ v_ptrs,
1370
+ off_h, offs_m, offs_n, offs_d,
1371
+ stride_kt, stride_vt,
1372
+ sm_scale,
1373
+ k_seqlen,
1374
+ past_len,
1375
+ True,
1376
+ BLOCK_M_LOADING,
1377
+ BLOCK_N,
1378
+ D_HEAD,
1379
+ EVEN_D,
1380
+ M_LT_N
1381
+ )
1382
+
1383
+ # write output
1384
+ if EVEN_D:
1385
+ tl.store(Out + offs_m[:, None] * stride_ot + offs_d[None, :] * stride_od, acc,
1386
+ mask=offs_m[:, None] < q_seqlen)
1387
+ else:
1388
+ tl.store(Out + offs_m[:, None] * stride_ot + offs_d[None, :] * stride_od, acc,
1389
+ mask=(offs_m[:, None] < q_seqlen) & (offs_d[None, :] < D_HEAD))
1390
+
1391
+
1392
+ ###########################################################
1393
+ ###########################################################
1394
+
1395
+ ###########################################################
1396
+ ################## Testing Utilities ######################
1397
+ ###########################################################
1398
+
1399
+
1400
+ def torch_attention(q, k, v, attn_mask=None, sm_scale=None, block_attn_mask=None, block_size=128, do=None):
1401
+ '''
1402
+ q, k, v: shape=(batch, n_heads, seq, dim)
1403
+ '''
1404
+ # for verification
1405
+ if sm_scale is None:
1406
+ sm_scale = math.sqrt(float(q.size(-1)))
1407
+
1408
+ if block_attn_mask is not None:
1409
+ assert attn_mask is None
1410
+ outs = []
1411
+ for s in range(0, q.size(2), block_size):
1412
+ e = min(s + block_size, q.size(2))
1413
+ q_block = q[:, :, s:e]
1414
+ attn = torch.einsum('bhmd,bhnd->bhmn', q_block, k[:, :, :e]).float() * sm_scale
1415
+ mask = block_attn_mask[..., s // block_size, : (s // block_size + 1)]
1416
+ mask = torch.kron(mask, torch.ones(block_size, block_size, device=mask.device))
1417
+ mask[..., :, s:].masked_fill_(torch.arange(0, block_size)[:, None] <= torch.arange(0, block_size)[None, :], 0)
1418
+ attn = attn.masked_fill((1 - mask).bool(), float('-inf'))
1419
+ attn = attn.softmax(-1)
1420
+ out = torch.einsum('bhmn,bhnd->bhmd', attn.type_as(v), v[:, :, :e])
1421
+ outs.append(out)
1422
+ torch_output = torch.cat(outs, dim=2)
1423
+ else:
1424
+ attn = torch.einsum('bhmd,bhnd->bhmn', q, k).float() * sm_scale
1425
+ # import ipdb; ipdb.set_trace()
1426
+ if attn_mask is not None:
1427
+ attn = attn.masked_fill((1 - attn_mask).bool(), float('-inf'))
1428
+ # print(f'> torch attn: {attn.exp().sum(-1)=}')
1429
+
1430
+ attn = attn.softmax(-1)
1431
+ if do is not None:
1432
+ dv = torch.einsum('bhqk,bhqd->bhkd', attn.type_as(do), do)
1433
+ print(f'> torch_attn computed dv: {dv=}')
1434
+ torch_output = torch.einsum('bhmn,bhnd->bhmd', attn.type_as(v), v)
1435
+ return torch_output
1436
+
1437
+ ###########################################################
1438
+ ###########################################################
1439
+
1440
+ ###########################################################
1441
+ #################### Unit Tests ###########################
1442
+ ###########################################################
1443
+
1444
+
1445
+ @pytest.mark.parametrize('Z, H, N_CTX, D_HEAD', [(2, 8, 2048, 128), (1, 4, 4096, 64)])
1446
+ def test_op(Z, H, N_CTX, D_HEAD, Q_LEN=None, dtype=torch.bfloat16, homo_head=True, kernel_block_size=None, sparse_block_size=128, backward=True,
1447
+ sparse_attention_fn=None, local_blocks=4, vert_stride=4, sm_scale=None, max_length=None):
1448
+ Q_LEN = Q_LEN or N_CTX
1449
+ torch.manual_seed(20)
1450
+ q = torch.empty((Z, H, Q_LEN, D_HEAD), dtype=dtype, device='cuda').normal_(mean=0, std=.5) # .requires_grad_()
1451
+ k = torch.empty((Z, H, N_CTX, D_HEAD), dtype=dtype, device='cuda').normal_(mean=0, std=.5) # .requires_grad_()
1452
+ v = torch.empty((Z, H, N_CTX, D_HEAD), dtype=dtype, device='cuda').normal_(mean=0, std=.5) # .requires_grad_()
1453
+
1454
+ if sm_scale is None:
1455
+ sm_scale = 1. / math.sqrt(D_HEAD)
1456
+
1457
+ # for debugging
1458
+ # print(f'>> {q.shape=}, {k.shape=}, {v.shape=}, {homo_head=}, {kernel_block_size=}, {sparse_block_size=}, {local_blocks=}, {vert_stride=}')
1459
+ sm_scale = 0.0078125
1460
+ if backward:
1461
+ q.requires_grad_(), k.requires_grad_(), v.requires_grad_()
1462
+
1463
+ # qkv = torch.empty((Z, N_CTX, 3*H*D_HEAD), dtype=dtype, device='cuda').normal_(mean=0, std=.5)
1464
+ # q = qkv[..., :H*D_HEAD]
1465
+ # k = qkv[..., H*D_HEAD:2*H*D_HEAD]
1466
+ # v = qkv[..., 2*H*D_HEAD:]
1467
+ # q = q.view(Z, N_CTX, H, -1).permute(0, 2, 1, 3)
1468
+ # k = k.view(Z, N_CTX, H, -1).permute(0, 2, 1, 3)
1469
+ # v = v.view(Z, N_CTX, H, -1).permute(0, 2, 1, 3)
1470
+
1471
+ # if Q_LEN and Q_LEN < N_CTX:
1472
+ # q = q[:, :, -Q_LEN:] # .contiguous()
1473
+
1474
+ # q = q.requires_grad_()
1475
+ # k = k.requires_grad_()
1476
+ # v = v.requires_grad_()
1477
+
1478
+ dout = torch.randn_like(q).contiguous()
1479
+
1480
+ # dout = torch.eye(N_CTX)[:, :D_HEAD][None, None].expand_as(q).type_as(q).contiguous()
1481
+ # print(dout)
1482
+
1483
+ mask_csr, _, mask_dense = get_sparse_attn_mask(q, N_CTX, BLOCK=sparse_block_size,
1484
+ local_blocks=local_blocks, vert_stride=vert_stride, homo_head=homo_head, return_dense=True)
1485
+
1486
+ if sparse_attention_fn is None:
1487
+ sparse_attention_fn = get_local_strided_sparse_attention_op(H, N_CTX,
1488
+ sparse_block_size=sparse_block_size,
1489
+ local_blocks=local_blocks,
1490
+ vert_stride=vert_stride,
1491
+ homo_head=homo_head,
1492
+ device=q.device,
1493
+ dtype=q.dtype,
1494
+ kernel_block_size=kernel_block_size)
1495
+ # reference implementation
1496
+ ref_out = torch_attention(q, k, v, mask_dense, sm_scale)
1497
+
1498
+ # lengths = torch.full((Z,), fill_value=N_CTX, device='cuda')
1499
+ # cu_seqlens = torch.zeros((Z + 1,), device='cuda', dtype=torch.int32)
1500
+ # cu_seqlens[1:] = lengths.cumsum(0)
1501
+ # # qkv = torch.randn((Z * N_CTX, 3, H, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1502
+
1503
+ # qkv_list = list(map(lambda x: x.permute(0, 2, 1, 3).contiguous().view(Z * N_CTX, 1, H, D_HEAD), [q, k, v]))
1504
+ # qkv = torch.cat(qkv_list, dim=1)
1505
+ # ref_out0 = flash_attn_func(qkv, cu_seqlens, dropout_p=0, max_s=N_CTX, softmax_scale=sm_scale, causal=True)
1506
+ # ref_out = ref_out0.view(Z, N_CTX, H, D_HEAD).permute(0, 2, 1, 3).contiguous()
1507
+
1508
+
1509
+ if backward:
1510
+ ref_out.backward(dout)
1511
+ ref_dv, v.grad = v.grad.clone(), None
1512
+ ref_dk, k.grad = k.grad.clone(), None
1513
+ ref_dq, q.grad = q.grad.clone(), None
1514
+
1515
+ tri_out = sparse_attention_fn(q, k, v, sm_scale)
1516
+
1517
+ decimal = 1 if dtype == torch.bfloat16 else 2
1518
+ assert torch.allclose(ref_out.cpu(), tri_out.cpu(), atol=1e-2, rtol=0), f'>> {ref_out[0, 0, :, 0].tolist()=}\n\n{tri_out[0, 0, :, 0].tolist()=}'
1519
+
1520
+ if backward:
1521
+ tri_out.backward(dout)
1522
+ tri_dv, v.grad = v.grad.clone(), None
1523
+ tri_dk, k.grad = k.grad.clone(), None
1524
+ tri_dq, q.grad = q.grad.clone(), None
1525
+
1526
+ if backward:
1527
+ assert torch.allclose(ref_dv, tri_dv, atol=1e-2, rtol=1e-2)
1528
+ assert torch.allclose(ref_dk, tri_dk, atol=1e-2, rtol=0)
1529
+ assert torch.allclose(ref_dq, tri_dq, atol=1e-2, rtol=0)
1530
+
1531
+ print(f'> test passed: {Z=}, {H=}, {N_CTX=}, {D_HEAD=}, {Q_LEN=}, {dtype=}, {homo_head=}, {sparse_block_size=}')
1532
+
1533
+ ###########################################################
1534
+
1535
+ if __name__ == '__main__':
1536
+
1537
+ GPU_TYPE = os.popen('nvidia-smi --query-gpu=name --format=csv | tail -n 1').read().strip()
1538
+ # print(GPU_TYPE)
1539
+ support_backward = True # 'A100' in GPU_TYPE. Wasn't supportted in consumer A1000.
1540
+
1541
+ ###############
1542
+ # benchmarking
1543
+
1544
+ HAS_DENSE_TRITON_FLASH = False
1545
+ # try:
1546
+ # from triton.ops.flash_attention import attention as triton_attention
1547
+ # HAS_DENSE_TRITON_FLASH = True
1548
+ # except:
1549
+ # HAS_DENSE_TRITON_FLASH = False
1550
+ # print('> cannot import Trition flash attn')
1551
+
1552
+ try:
1553
+ from flash_attn.flash_attn_interface import flash_attn_func, flash_attn_unpadded_func
1554
+ HAS_FLASH = True
1555
+ except BaseException:
1556
+ HAS_FLASH = False
1557
+ print('> cannot import flash_attn')
1558
+
1559
+
1560
+ # BATCH, N_HEADS, N_CTX, D_HEAD = 4, 48, 4096, 64
1561
+ BATCH, N_HEADS, N_CTX, D_HEAD = 4, 32, 4096, 128 # 6.7B model, with 4k len
1562
+ # BATCH, N_HEADS, N_CTX, D_HEAD = 4, 16, 4096, 128 # 204m model
1563
+
1564
+ BLOCK_SIZE = 64
1565
+ LOCAl_BLOCKS = 8 # 4
1566
+ VERT_STRIDE = 1 # 16 # 8
1567
+ HOMO_HEAD = False
1568
+ sparse_type = 'home' if HOMO_HEAD else 'hetero'
1569
+ dtype = torch.bfloat16
1570
+
1571
+
1572
+ modes = ['fwd', 'bwd'] if support_backward else ['fwd']
1573
+
1574
+ configs = [triton.testing.Benchmark(
1575
+ x_names=['SEQ_LEN'],
1576
+ x_vals=[2**i for i in range(8, 16)],
1577
+ line_arg='provider',
1578
+ line_vals=(['triton'] if HAS_DENSE_TRITON_FLASH else []) + (['flash'] if HAS_FLASH else []) + ['triton_sparse'],
1579
+ line_names=(['Triton-Dense'] if HAS_DENSE_TRITON_FLASH else []) + (['Flash-Dense'] if HAS_FLASH else []) + ['Triton-Sparse'],
1580
+ styles=[('red', '-'), ('blue', '-'), ('green', '-')],
1581
+ ylabel='ms',
1582
+ plot_name=f'fused-attention-batch{BATCH}-head{N_HEADS}-d{D_HEAD}-sparse-local{LOCAl_BLOCKS}-vert{VERT_STRIDE}-{sparse_type}-{dtype}-{mode}',
1583
+ args={'H': N_HEADS, 'BATCH': BATCH, 'D_HEAD': D_HEAD, 'dtype': dtype, 'mode': mode}
1584
+ ) for mode in modes]
1585
+
1586
+
1587
+ @triton.testing.perf_report(configs)
1588
+ def bench_flash_attention(BATCH, H, SEQ_LEN, D_HEAD, mode, provider, dtype=torch.bfloat16, device='cuda', sparse_attention_fn=None):
1589
+ assert mode in ['fwd', 'bwd']
1590
+ warmup = 25
1591
+ rep = 100
1592
+ N_CTX = SEQ_LEN
1593
+ if provider == 'triton':
1594
+ q = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1595
+ k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1596
+ v = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1597
+ sm_scale = 1.3
1598
+ fn = lambda: triton_attention(q, k, v, sm_scale)
1599
+ if mode == 'bwd':
1600
+ o = fn()
1601
+ do = torch.randn_like(o)
1602
+ fn = lambda: o.backward(do, retain_graph=True)
1603
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1604
+ return ms
1605
+ if provider == 'triton_sparse':
1606
+ q = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1607
+ k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1608
+ v = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1609
+ sm_scale = 1.3
1610
+ # q_pos = torch.arange(N_CTX // BLOCK, device='cuda')[:, None]
1611
+ # k_pos = torch.arange(N_CTX // BLOCK, device='cuda')[None]
1612
+ # local_blocks = 4 # num_block per attn, block_size is tied to BLOCK
1613
+ # vert_stride =N_CTX + 1 # 4
1614
+ # mask_vert_strided = torch.arange(N_CTX // BLOCK, device='cuda') % vert_stride == vert_stride - 1
1615
+ # mask_dense = ((q_pos >= k_pos) & ((q_pos - k_pos < local_blocks) | mask_vert_strided)).type_as(q)
1616
+ # mask = mask_dense.to_sparse_csr()
1617
+ # mask_csr, _ = get_sparse_attn_mask(q, N_CTX, BLOCK=BLOCK, local_blocks=LOCAl_BLOCKS, vert_stride=VERT_STRIDE, homo_head=HOMO_HEAD)
1618
+
1619
+ if sparse_attention_fn is None:
1620
+ # sparse_attention_fn = sparse_attention
1621
+ sparse_attention_fn = get_local_strided_sparse_attention_op(H, SEQ_LEN,
1622
+ local_blocks=LOCAl_BLOCKS,
1623
+ vert_stride=VERT_STRIDE,
1624
+ homo_head=HOMO_HEAD,
1625
+ sparse_block_size=BLOCK_SIZE,
1626
+ kernel_block_size=BLOCK_SIZE,
1627
+ device=q.device)
1628
+ # sparse_attention_fn = sparse_attention_factory(128, 128, num_warps=8)
1629
+
1630
+ # fn = lambda: sparse_attention_fn(q, k, v, mask_csr[0], mask_csr[1], sm_scale)
1631
+ fn = lambda: sparse_attention_fn(q, k, v, sm_scale)
1632
+ if mode == 'bwd':
1633
+ o = fn()
1634
+ do = torch.randn_like(o)
1635
+ fn = lambda: o.backward(do, retain_graph=True)
1636
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1637
+ return ms
1638
+ if provider == 'flash':
1639
+ lengths = torch.full((BATCH,), fill_value=N_CTX, device=device)
1640
+ cu_seqlens = torch.zeros((BATCH + 1,), device=device, dtype=torch.int32)
1641
+ cu_seqlens[1:] = lengths.cumsum(0)
1642
+ qkv = torch.randn((BATCH * N_CTX, 3, H, D_HEAD), dtype=dtype, device=device, requires_grad=True)
1643
+ fn = lambda: flash_attn_func(qkv, cu_seqlens, 0., N_CTX, causal=True)
1644
+ if mode == 'bwd':
1645
+ o = fn()
1646
+ do = torch.randn_like(o)
1647
+ fn = lambda: o.backward(do, retain_graph=True)
1648
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1649
+ return ms
1650
+
1651
+ # if provider == 'torch':
1652
+ # q = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1653
+ # k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1654
+ # v = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=True)
1655
+ # sm_scale = 1.3
1656
+ # causal_mask = torch.tril(torch.ones(N_CTX, N_CTX)).type_as(q)
1657
+ # fn = lambda: torch_attention(q, k, v, causal_mask, sm_scale)
1658
+ # ms = triton.testing.do_bench(fn, percentiles=None, warmup=warmup, rep=rep)
1659
+ # return ms
1660
+
1661
+
1662
+ BATCH, N_HEADS, N_CTX, D_HEAD, Q_LEN = 4, 32, 4096, 128, 1 # 6.7B model, with 4k len
1663
+
1664
+ BLOCK_SIZE = 64
1665
+ LOCAl_BLOCKS = 8 # 4
1666
+ VERT_STRIDE = 16 # 8
1667
+ HOMO_HEAD = False
1668
+ sparse_type = 'home' if HOMO_HEAD else 'hetero'
1669
+ dtype = torch.bfloat16
1670
+ MAX_N_CTX = 8192
1671
+
1672
+ configs = [triton.testing.Benchmark(
1673
+ x_names=['PAST_LEN'],
1674
+ x_vals=[2**i - 1 for i in range(8, 14)],
1675
+ line_arg='provider',
1676
+ line_vals=['torch'] + (['flash'] if HAS_FLASH else []) + ['triton_sparse', 'triton_dense'],
1677
+ line_names=['Torch'] + (['Flash-Dense'] if HAS_FLASH else []) + ['Triton-Sparse', 'Triton-Dense'],
1678
+ styles=[('red', '-'), ('blue', '-'), ('green', '-'), ('cyan', '-')],
1679
+ ylabel='ms',
1680
+ plot_name=f'fused-attention-inference-batch{BATCH}-head{N_HEADS}-d{D_HEAD}-sparse-local{LOCAl_BLOCKS}-vert{VERT_STRIDE}-{sparse_type}',
1681
+ args={'H': N_HEADS, 'BATCH': BATCH, 'D_HEAD': D_HEAD, 'Q_LEN': Q_LEN, 'dtype': torch.float16, 'mode': mode}
1682
+ ) for mode in ['fwd']]
1683
+ @triton.testing.perf_report(configs)
1684
+ def bench_flash_attention_inference(BATCH, H, PAST_LEN, D_HEAD, Q_LEN, mode, provider, dtype=torch.bfloat16, device='cuda'):
1685
+ assert mode in ['fwd']
1686
+ warmup = 25
1687
+ rep = 100
1688
+ N_CTX = PAST_LEN + Q_LEN
1689
+ if provider == 'torch':
1690
+ q = torch.randn((BATCH, H, Q_LEN, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1691
+ k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1692
+ v = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1693
+ sm_scale = 1.3
1694
+ mask_csr, _, mask_dense = get_sparse_attn_mask(q, N_CTX, BLOCK=BLOCK_SIZE,
1695
+ local_blocks=LOCAl_BLOCKS, vert_stride=VERT_STRIDE, homo_head=VERT_STRIDE, return_dense=True)
1696
+
1697
+ fn = lambda: torch_attention(q, k, v, mask_dense, sm_scale=sm_scale, block_size=2048)
1698
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1699
+ return ms
1700
+ if provider == 'triton_sparse':
1701
+ q = torch.randn((BATCH, H, Q_LEN, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1702
+ k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1703
+ v = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1704
+ sm_scale = 1.3
1705
+ sparse_attention_fn = get_local_strided_sparse_attention_op(H, MAX_N_CTX,
1706
+ local_blocks=LOCAl_BLOCKS,
1707
+ vert_stride=VERT_STRIDE,
1708
+ homo_head=HOMO_HEAD,
1709
+ sparse_block_size=BLOCK_SIZE,
1710
+ kernel_block_size=BLOCK_SIZE,
1711
+ device=q.device,
1712
+ inference=True)
1713
+
1714
+ fn = lambda: sparse_attention_fn(q, k, v, sm_scale)
1715
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1716
+ return ms
1717
+ if provider == 'triton_dense':
1718
+ q = torch.randn((BATCH, H, Q_LEN, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1719
+ k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1720
+ v = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1721
+ sm_scale = 1.3
1722
+ sparse_attention_fn = get_local_strided_sparse_attention_op(H, MAX_N_CTX,
1723
+ local_blocks=1,
1724
+ vert_stride=1,
1725
+ homo_head=True,
1726
+ sparse_block_size=BLOCK_SIZE,
1727
+ kernel_block_size=BLOCK_SIZE,
1728
+ device=q.device,
1729
+ inference=True)
1730
+
1731
+ fn = lambda: sparse_attention_fn(q, k, v, sm_scale)
1732
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1733
+ return ms
1734
+ if provider == 'flash':
1735
+ assert Q_LEN == 1
1736
+ lengths = torch.full((BATCH,), fill_value=N_CTX, device=device)
1737
+ cu_seqlens = torch.zeros((BATCH + 1,), device=device, dtype=torch.int32)
1738
+ cu_seqlens[1:] = lengths.cumsum(0)
1739
+ cu_seqlens_q = torch.arange(BATCH + 1, device=device, dtype=torch.int32)
1740
+
1741
+ # (total_q, nheads, headdim),
1742
+ q = torch.randn((BATCH, H, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1743
+ k = torch.randn((BATCH*N_CTX, H, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1744
+ v = torch.randn((BATCH*N_CTX, H, D_HEAD), dtype=dtype, device='cuda', requires_grad=False)
1745
+
1746
+ fn = lambda: flash_attn_unpadded_func(q, k, v, cu_seqlens_q, cu_seqlens, 1, N_CTX, dropout_p=0, softmax_scale=1.3, causal=False)
1747
+ ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
1748
+ return ms
1749
+
1750
+
1751
+ test_op(1, 4, 512, 128, dtype=torch.float16, homo_head=False, backward=support_backward)
1752
+ # bench_flash_attention.run(save_path='.', print_data=True)
1753
+
1754
+ bench_flash_attention_inference.run(save_path='.', print_data=True)
1755
+ exit()
1756
+ # head_dim=64
1757
+ test_op(1, 2, 1024, 64, kernel_block_size=64, sparse_block_size=64,
1758
+ dtype=torch.bfloat16, homo_head=False, backward=support_backward)
1759
+ # uneven length, bf16
1760
+ test_op(1, 16, 224, 128, dtype=torch.bfloat16, homo_head=False, backward=False, sparse_block_size=128,
1761
+ kernel_block_size=64, local_blocks=8, vert_stride=8)
1762
+ test_op(3, 2, 2047, 128, homo_head=False, backward=False)
1763
+
1764
+ # diff kernel/sparse block size
1765
+ test_op(1, 16, 224, 128, dtype=torch.bfloat16, homo_head=False, backward=False, kernel_block_size=64)
1766
+ # inference
1767
+ # test_op(1, 4, 512 + 256, 128, Q_LEN=1, dtype=torch.bfloat16, homo_head=False, backward=support_backward)
1768
+
1769
+ # dense flash attn
1770
+ test_op(1, 2, 1024, 128, kernel_block_size=128, sparse_block_size=128, dtype=torch.bfloat16, homo_head=False,
1771
+ backward=support_backward, local_blocks=1, vert_stride=1)
1772
+
1773
+ # fp16
1774
+ test_op(1, 4, 512 + 256, 128, dtype=torch.float16, homo_head=False, backward=support_backward)
1775
+
1776
+ # longer sequence
1777
+ test_op(2, 4, 8192, 64, homo_head=False, backward=support_backward)
1778
+ test_op(2, 4, 8192, 128, dtype=torch.bfloat16, homo_head=False, backward=support_backward)
1779
+
1780
+ # homo head
1781
+ test_op(3, 2, 2048, 64, homo_head=True, dtype=torch.bfloat16, backward=False)
1782
+ test_op(3, 2, 2048, 64, homo_head=True, backward=support_backward)
1783
+
1784
+ # sparse_attention_fn = sparse_attention_factory(16, 128, num_warps=1, INFERENCE=True)
1785
+ # test_op(8, 1, 2047, 128, 1, backward=False, sparse_attention_fn=None)
1786
+ # test_op_inference(3, 2, 2048, 128, 2048)
1787
+ # test_op_inference(3, 2, 2047, 64, 2047)
1788
+ # test_op_inference(3, 2, 256, 64, 128)
1789
+ # test_op_inference(3, 2, 2048, 64, 1)
1790
+
1791
+ bench_flash_attention.run(save_path='.', print_data=True)
1792
+ # bench_flash_attention_inference.run(save_path='.', print_data=True)
1793
+
1794
+ # ========================
1795
+ # Some Benchmark Results #
1796
+ # ========================
1797
+
1798
+ # fused-attention-batch4-head48-d64-sparse-local4-vert4-hetero-fwd
1799
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1800
+ # 0 256.0 0.057184 0.069646 0.052567
1801
+ # 1 512.0 0.131688 0.187658 0.110212
1802
+ # 2 1024.0 0.391844 0.524990 0.247875
1803
+ # 3 2048.0 1.305190 1.456685 0.596506
1804
+ # 4 4096.0 4.623019 4.968653 1.600277
1805
+ # 5 8192.0 17.513062 18.332262 4.802458
1806
+ # 6 16384.0 68.453377 70.337540 16.052908
1807
+ # 7 32768.0 270.655487 276.020233 57.938946
1808
+ # fused-attention-batch4-head48-d64-sparse-local4-vert4-hetero-bwd (num_warp=8):
1809
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1810
+ # 0 256.0 0.190120 0.150313 0.181451
1811
+ # 1 512.0 0.406348 0.391767 0.391177
1812
+ # 2 1024.0 1.029704 1.182967 0.885741
1813
+ # 3 2048.0 2.985456 3.843399 2.040469
1814
+ # 4 4096.0 9.808897 13.073701 5.069609
1815
+ # 5 8192.0 34.995201 47.863808 13.948782
1816
+ # 6 16384.0 132.740097 182.579193 42.816513
1817
+ # 7 32768.0 542.223389 714.820618 147.053574
1818
+ # fused-attention-inference-batch4-head32-d128-sparse-local4-vert4-hetero:
1819
+ # PAST_LEN Torch-Dense Flash-Dense Triton-Sparse
1820
+ # 0 256.0 0.050949 0.032357 0.107513
1821
+ # 1 512.0 0.073624 0.050651 0.199086
1822
+ # 2 1024.0 0.107472 0.080379 0.245445
1823
+ # 3 2048.0 0.178423 0.129448 0.338259
1824
+ # 4 4096.0 0.327647 0.223106 0.517048
1825
+ # 5 8192.0 0.588423 0.411263 0.884606
1826
+ # 6 16384.0 1.098898 0.798941 1.611809
1827
+ # 7 32768.0 2.094537 1.594726 3.044160
1828
+
1829
+
1830
+ # 6.7B
1831
+ # fused-attention-batch4-head32-d128-sparse-local4-vert4-hetero-fwd:
1832
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1833
+ # 0 256.0 0.069208 0.082156 0.065097
1834
+ # 1 512.0 0.138271 0.201393 0.144467
1835
+ # 2 1024.0 0.391521 0.624614 0.322382
1836
+ # 3 2048.0 1.268443 2.406325 0.784367
1837
+ # 4 4096.0 4.455703 9.139097 2.100856
1838
+ # 5 8192.0 16.764315 35.289600 6.328320
1839
+ # 6 16384.0 65.221634 138.401794 21.069057
1840
+ # 7 32768.0 257.251343 548.085754 76.111870
1841
+ # fused-attention-batch4-head32-d128-sparse-local4-vert4-hetero-bwd:
1842
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1843
+ # 0 256.0 0.297118 0.266469 0.255255
1844
+ # 1 512.0 0.672826 0.613685 0.552954
1845
+ # 2 1024.0 1.718434 1.705066 1.251953
1846
+ # 3 2048.0 4.936755 5.403875 2.927895
1847
+ # 4 4096.0 15.911594 18.959362 7.436288
1848
+ # 5 8192.0 55.357441 70.808578 21.140224
1849
+ # 6 16384.0 208.188416 273.617920 68.018173
1850
+ # 7 32768.0 806.037476 1081.453613 218.720261
1851
+ # fused-attention-inference-batch4-head32-d128-sparse-local4-vert4-hetero:
1852
+ # PAST_LEN Torch-Dense Flash-Dense Triton-Sparse
1853
+ # 0 256.0 0.050151 0.032337 0.107593
1854
+ # 1 512.0 0.073409 0.051737 0.200200
1855
+ # 2 1024.0 0.107533 0.082099 0.247067
1856
+ # 3 2048.0 0.177259 0.128891 0.338510
1857
+ # 4 4096.0 0.325866 0.223621 0.524842
1858
+ # 5 8192.0 0.586926 0.408913 0.885490
1859
+ # 6 16384.0 1.100834 0.793277 1.612271
1860
+ # 7 32768.0 2.098851 1.595831 3.064544
1861
+
1862
+ # fused-attention-batch4-head32-d128-sparse-local4-vert8-hetero-fwd:
1863
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1864
+ # 0 256.0 0.066673 0.082037 0.065085
1865
+ # 1 512.0 0.137379 0.201880 0.143473
1866
+ # 2 1024.0 0.390675 0.624234 0.312046
1867
+ # 3 2048.0 1.267739 2.406950 0.696045
1868
+ # 4 4096.0 4.445138 9.136333 1.665788
1869
+ # 5 8192.0 16.768614 35.265533 4.380486
1870
+ # 6 16384.0 65.235970 138.393600 12.997633
1871
+ # 7 32768.0 257.317902 550.442993 42.821121
1872
+ # fused-attention-batch4-head32-d128-sparse-local4-vert8-hetero-bwd:
1873
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1874
+ # 0 256.0 0.296461 0.266581 0.254022
1875
+ # 1 512.0 0.671427 0.613643 0.551283
1876
+ # 2 1024.0 1.719918 1.704295 1.229982
1877
+ # 3 2048.0 4.945305 5.403364 2.721906
1878
+ # 4 4096.0 15.934293 18.960999 6.259371
1879
+ # 5 8192.0 55.406593 70.832130 15.676929
1880
+ # 6 16384.0 208.750595 275.004425 44.837891
1881
+ # 7 32768.0 808.057861 1080.647705 141.856766
1882
+ # fused-attention-inference-batch4-head32-d128-sparse-local4-vert8-hetero:
1883
+ # PAST_LEN Torch-Dense Flash-Dense Triton-Sparse
1884
+ # 0 256.0 0.050739 0.032886 0.107837
1885
+ # 1 512.0 0.073507 0.051996 0.200293
1886
+ # 2 1024.0 0.106394 0.080679 0.240610
1887
+ # 3 2048.0 0.177659 0.127660 0.287625
1888
+ # 4 4096.0 0.326326 0.226971 0.377500
1889
+ # 5 8192.0 0.586339 0.407367 0.559266
1890
+ # 6 16384.0 1.102279 0.786221 0.920976
1891
+ # 7 32768.0 2.097370 1.545090 1.644288
1892
+
1893
+
1894
+ ################
1895
+ ##### fp16 #####
1896
+ ################
1897
+
1898
+ # fused-attention-batch4-head16-d64-sparse-local4-vert8-hetero-fwd:
1899
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1900
+ # 0 256.0 0.032518 0.035472 0.029939
1901
+ # 1 512.0 0.054266 0.087841 0.054320
1902
+ # 2 1024.0 0.133447 0.263090 0.102045
1903
+ # 3 2048.0 0.384615 1.023293 0.201763
1904
+ # 4 4096.0 1.300890 4.023936 0.449555
1905
+ # 5 8192.0 4.774144 15.816704 1.150854
1906
+ # 6 16384.0 18.220032 62.771198 3.356001
1907
+ # 7 32768.0 71.405571 250.273788 10.976142
1908
+ # fused-attention-batch4-head16-d64-sparse-local4-vert8-hetero-bwd:
1909
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1910
+ # 0 256.0 0.083342 0.069742 0.079496
1911
+ # 1 512.0 0.159894 0.170995 0.151705
1912
+ # 2 1024.0 0.386071 0.522407 0.331443
1913
+ # 3 2048.0 1.067715 1.737333 0.715248
1914
+ # 4 4096.0 3.382731 6.219520 1.597457
1915
+ # 5 8192.0 11.857793 23.560448 3.879035
1916
+ # 6 16384.0 44.422142 91.251709 10.626843
1917
+ # 7 32768.0 175.011841 359.473145 32.340992
1918
+
1919
+
1920
+ ################
1921
+ ##### bf16 #####
1922
+ ################
1923
+
1924
+ # fused-attention-batch4-head16-d64-sparse-local4-vert8-hetero-fwd:
1925
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1926
+ # 0 256.0 0.037636 0.035902 0.031512
1927
+ # 1 512.0 0.058591 0.087229 0.058125
1928
+ # 2 1024.0 0.143337 0.263919 0.108443
1929
+ # 3 2048.0 0.414458 1.025985 0.214114
1930
+ # 4 4096.0 1.390841 4.020010 0.480550
1931
+ # 5 8192.0 5.067938 15.808171 1.230874
1932
+ # 6 16384.0 19.442280 62.765057 3.597274
1933
+ # 7 32768.0 75.501572 250.443771 11.768959
1934
+ # fused-attention-batch4-head16-d64-sparse-local4-vert8-hetero-bwd:
1935
+ # SEQ_LEN Triton-Dense Flash-Dense Triton-Sparse
1936
+ # 0 256.0 0.084404 0.070663 0.082613
1937
+ # 1 512.0 0.161510 0.172882 0.157661
1938
+ # 2 1024.0 0.388954 0.526047 0.339855
1939
+ # 3 2048.0 1.075814 1.736057 0.732420
1940
+ # 4 4096.0 3.401622 6.221376 1.636039
1941
+ # 5 8192.0 11.915136 23.483391 3.968725
1942
+ # 6 16384.0 44.660225 91.302910 10.857130
1943
+ # 7 32768.0 175.038467 359.048187 32.778240