riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-02 11:11:05 +06:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

Llama: always convert the causal mask in the SDPA code path (#29663)

Browse Source 
* always convert the mask

* rebase and fix copies
This commit is contained in:
Joao Gante 2024-03-21 16:30:18 +00:00 committed by GitHub
parent 5ffef2a978
commit ee38fc31fb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 12 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/transformers/models/cohere/modeling_cohere.py
View File

@ -1005,17 +1005,10 @@ class CohereModel(CoherePreTrainedModel):
and attention_mask is not None
and attention_mask.device.type == "cuda"
):
# TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
is_tracing = (
torch.jit.is_tracing()
or isinstance(input_tensor, torch.fx.Proxy)
or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
)
if not is_tracing and torch.any(attention_mask != 1):
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
return causal_mask

15
src/transformers/models/gemma/modeling_gemma.py
View File

@ -1011,17 +1011,10 @@ class GemmaModel(GemmaPreTrainedModel):
and attention_mask is not None
and attention_mask.device.type == "cuda"
):
# TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
is_tracing = (
torch.jit.is_tracing()
or isinstance(input_tensor, torch.fx.Proxy)
or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
)
if not is_tracing and torch.any(attention_mask != 1):
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
return causal_mask

15
src/transformers/models/llama/modeling_llama.py
View File

@ -1100,17 +1100,10 @@ class LlamaModel(LlamaPreTrainedModel):
and attention_mask is not None
and attention_mask.device.type == "cuda"
):
# TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
is_tracing = (
torch.jit.is_tracing()
or isinstance(input_tensor, torch.fx.Proxy)
or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
)
if not is_tracing and torch.any(attention_mask != 1):
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
return causal_mask