Sentence Similarity
Transformers
Safetensors
English
llama
feature-extraction
text-embedding
embeddings
information-retrieval
beir
text-classification
language-model
text-clustering
text-semantic-similarity
text-evaluation
text-reranking
Sentence Similarity
natural_questions
ms_marco
fever
hotpot_qa
mteb
custom_code
text-generation-inference
Inference Endpoints
File size: 11,004 Bytes
899a4e8 48b3d3b 899a4e8 48b3d3b 899a4e8 48b3d3b 899a4e8 48b3d3b 899a4e8 48b3d3b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 |
from typing import List, Optional, Tuple, Union
import torch
from packaging import version
import importlib.metadata
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.utils.import_utils import _is_package_available
def is_transformers_attn_greater_or_equal_4_39():
if not _is_package_available("transformers"):
return False
return version.parse(importlib.metadata.version("transformers")) >= version.parse(
"4.39.0"
)
def _prepare_4d_attention_mask_for_sdpa(
attention_mask: Optional[torch.Tensor],
input_shape: Union[torch.Size, Tuple, List],
inputs_embeds: torch.Tensor,
past_key_values_length: int,
sliding_window: Optional[int] = None,
):
attn_mask_converter = AttentionMaskConverter(is_causal=False, sliding_window=sliding_window)
key_value_length = input_shape[-1] + past_key_values_length
batch_size, query_length = input_shape
# torch.jit.trace and torchdynamo with fullgraph=True are unable to capture the controlflow `is_causal=attention_mask is None and q_len > 1`
# used as an SDPA argument. We keep compatibility with these tracing tools by always using SDPA's `attn_mask` argument in case we are tracing.
# TODO: Fix this as well when using torchdynamo with fullgraph=True.
is_tracing = torch.jit.is_tracing()
if attention_mask is not None:
if torch.all(attention_mask == 1):
if is_tracing:
pass
elif query_length == 1:
# For query_length == 1, causal attention and bi-directional attention are the same.
attention_mask = None
elif key_value_length == query_length:
attention_mask = None
else:
# Unfortunately, for query_length > 1 and key_value_length != query_length, we cannot generally ignore the attention mask, as SDPA causal mask generation
# may be wrong. We will set `is_causal=False` in SDPA and rely on Transformers attention_mask instead, hence not setting it to None here.
# Reference: https://github.com/pytorch/pytorch/issues/108108
pass
elif query_length > 1 and key_value_length != query_length:
# See the comment above (https://github.com/pytorch/pytorch/issues/108108).
# Ugly: we set it to True here to dispatch in the following controlflow to `to_causal_4d`.
attention_mask = True
elif is_tracing:
raise ValueError(
'Attention using SDPA can not be traced with torch.jit.trace when no attention_mask is provided. To solve this issue, please either load your model with the argument `attn_implementation="eager"` or pass an attention_mask input when tracing the model.'
)
if attention_mask is None:
expanded_4d_mask = None
elif attention_mask is True:
expanded_4d_mask = attn_mask_converter.to_causal_4d(
input_shape[0], input_shape[-1], key_value_length, dtype=inputs_embeds.dtype, device=inputs_embeds.device
)
else:
expanded_4d_mask = attn_mask_converter.to_4d(
attention_mask,
input_shape[-1],
dtype=inputs_embeds.dtype,
key_value_length=key_value_length,
)
# From PyTorch 2.1 onwards, F.scaled_dot_product_attention with the memory-efficient attention backend
# produces nans if sequences are completely unattended in the attention mask. Details: https://github.com/pytorch/pytorch/issues/110213
if query_length > 1:
if is_transformers_attn_greater_or_equal_4_39():
expanded_4d_mask = AttentionMaskConverter._unmask_unattended(
expanded_4d_mask, min_dtype=torch.finfo(inputs_embeds.dtype).min
)
else:
expanded_4d_mask = AttentionMaskConverter._unmask_unattended(
expanded_4d_mask, attention_mask, unmasked_value=0.0
)
return expanded_4d_mask
def _prepare_4d_attention_mask(
attention_mask: Optional[torch.Tensor],
input_shape: Union[torch.Size, Tuple, List],
inputs_embeds: torch.Tensor,
past_key_values_length: int,
sliding_window: Optional[int] = None,
):
attn_mask_converter = AttentionMaskConverter(is_causal=False, sliding_window=sliding_window)
key_value_length = input_shape[-1] + past_key_values_length
# 4d mask is passed through the layers
if attention_mask is not None:
attention_mask = attn_mask_converter.to_4d(
attention_mask, input_shape[-1], key_value_length=key_value_length, dtype=inputs_embeds.dtype
)
else:
attention_mask = attn_mask_converter.to_causal_4d(
input_shape[0], input_shape[-1], key_value_length, dtype=inputs_embeds.dtype, device=inputs_embeds.device
)
return attention_mask
def _prepare_4d_causal_attention_mask(
attention_mask: Optional[torch.Tensor],
input_shape: Union[torch.Size, Tuple, List],
inputs_embeds: torch.Tensor,
past_key_values_length: int,
sliding_window: Optional[int] = None,
):
attn_mask_converter = AttentionMaskConverter(is_causal=False, sliding_window=sliding_window)
key_value_length = input_shape[-1] + past_key_values_length
# 4d mask is passed through the layers
if attention_mask is not None:
attention_mask = attn_mask_converter.to_4d(
attention_mask, input_shape[-1], key_value_length=key_value_length, dtype=inputs_embeds.dtype
)
else:
attention_mask = attn_mask_converter.to_causal_4d(
input_shape[0], input_shape[-1], key_value_length, dtype=inputs_embeds.dtype, device=inputs_embeds.device
)
return attention_mask
def _prepare_4d_causal_attention_mask_for_sdpa(
attention_mask: Optional[torch.Tensor],
input_shape: Union[torch.Size, Tuple, List],
inputs_embeds: torch.Tensor,
past_key_values_length: int,
sliding_window: Optional[int] = None,
):
"""
Prepares the correct `attn_mask` argument to be used by `torch.nn.functional.scaled_dot_product_attention`.
In case no token is masked in the `attention_mask` argument, we simply set it to `None` for the cases `query_length == 1` and
`key_value_length == query_length`, and rely instead on SDPA `is_causal` argument to use causal/non-causal masks,
allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is passed).
"""
attn_mask_converter = AttentionMaskConverter(is_causal=False, sliding_window=sliding_window)
key_value_length = input_shape[-1] + past_key_values_length
batch_size, query_length = input_shape
# torch.jit.trace, symbolic_trace and torchdynamo with fullgraph=True are unable to capture the controlflow `is_causal=attention_mask is None and q_len > 1`
# used as an SDPA argument. We keep compatibility with these tracing tools by always using SDPA's `attn_mask` argument in case we are tracing.
# TODO: Fix this as well when using torchdynamo with fullgraph=True.
is_tracing = torch.jit.is_tracing() or isinstance(inputs_embeds, torch.fx.Proxy)
if attention_mask is not None:
# 4d mask is passed through
if len(attention_mask.shape) == 4:
expected_shape = (input_shape[0], 1, input_shape[1], key_value_length)
if tuple(attention_mask.shape) != expected_shape:
raise ValueError(
f"Incorrect 4D attention_mask shape: {tuple(attention_mask.shape)}; expected: {expected_shape}."
)
else:
# if the 4D mask has correct shape - invert it and fill with negative infinity
inverted_mask = 1.0 - attention_mask.to(inputs_embeds.dtype)
attention_mask = inverted_mask.masked_fill(
inverted_mask.to(torch.bool), torch.finfo(inputs_embeds.dtype).min
)
return attention_mask
elif not is_tracing and torch.all(attention_mask == 1):
if query_length == 1:
# For query_length == 1, causal attention and bi-directional attention are the same.
attention_mask = None
elif key_value_length == query_length:
attention_mask = None
else:
# Unfortunately, for query_length > 1 and key_value_length != query_length, we cannot generally ignore the attention mask, as SDPA causal mask generation
# may be wrong. We will set `is_causal=False` in SDPA and rely on Transformers attention_mask instead, hence not setting it to None here.
# Reference: https://github.com/pytorch/pytorch/issues/108108
pass
elif query_length > 1 and key_value_length != query_length:
# See the comment above (https://github.com/pytorch/pytorch/issues/108108).
# Ugly: we set it to True here to dispatch in the following controlflow to `to_causal_4d`.
attention_mask = True
elif is_tracing:
raise ValueError(
'Attention using SDPA can not be traced with torch.jit.trace when no attention_mask is provided. To solve this issue, please either load your model with the argument `attn_implementation="eager"` or pass an attention_mask input when tracing the model.'
)
if attention_mask is None:
expanded_4d_mask = None
elif attention_mask is True:
expanded_4d_mask = attn_mask_converter.to_causal_4d(
input_shape[0], input_shape[-1], key_value_length, dtype=inputs_embeds.dtype, device=inputs_embeds.device
)
else:
expanded_4d_mask = attn_mask_converter.to_4d(
attention_mask,
input_shape[-1],
dtype=inputs_embeds.dtype,
key_value_length=key_value_length,
)
# From PyTorch 2.1 onwards, F.scaled_dot_product_attention with the memory-efficient attention backend
# produces nans if sequences are completely unattended in the attention mask. Details: https://github.com/pytorch/pytorch/issues/110213
#
# This fix is not applied in case we are tracing with torch.jit.trace or symbolic_trace, as _unmask_unattended has a data-dependent
# controlflow that can not be captured properly.
# TODO: _unmask_unattended does not work either with torch.compile when using fullgraph=True. We should find a way to detect this case.
if query_length > 1 and not is_tracing:
if is_transformers_attn_greater_or_equal_4_39():
expanded_4d_mask = AttentionMaskConverter._unmask_unattended(
expanded_4d_mask, min_dtype=torch.finfo(inputs_embeds.dtype).min
)
else:
expanded_4d_mask = AttentionMaskConverter._unmask_unattended(
expanded_4d_mask, attention_mask, unmasked_value=0.0
)
return expanded_4d_mask
|