Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix for Neuron #30259

Merged
merged 16 commits into from
May 2, 2024
Merged

Fix for Neuron #30259

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
d32aee0
Fix _update_causal_mask for Neuron
michaelbenayoun Apr 15, 2024
9429cd9
Remove the use of Ellipsis
michaelbenayoun Apr 16, 2024
5f6526c
Update warning message
michaelbenayoun Apr 16, 2024
32be4b7
Fixup
michaelbenayoun Apr 17, 2024
9e72932
Fix
michaelbenayoun Apr 17, 2024
3198401
Fix _update_causal_mask for Neuron, it works
michaelbenayoun Apr 18, 2024
446cb62
Fix FX when defining custom leaf module
michaelbenayoun Apr 25, 2024
65ebf45
Partial fix
michaelbenayoun Apr 25, 2024
1483dbb
Fix
michaelbenayoun Apr 26, 2024
206b007
Merge branch 'main' into fix_for_neuron
michaelbenayoun Apr 29, 2024
c991c89
[WIP] use metaclasses instead
michaelbenayoun Apr 30, 2024
c8362c0
Works
michaelbenayoun Apr 30, 2024
2eae3e0
Cleanup
michaelbenayoun Apr 30, 2024
8ed56cb
Fix
michaelbenayoun Apr 30, 2024
0d547f4
Removing comment
michaelbenayoun Apr 30, 2024
2a24b86
nit
michaelbenayoun Apr 30, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 5 additions & 2 deletions src/transformers/models/cohere/modeling_cohere.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1003,8 +1003,11 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position, curr
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
if attention_mask.dim() == 2:
mask_length = attention_mask.shape[-1]
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

you risk overflows here no?
If you have torch_min + torch_min -> produced inf

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1 - we should try to avoid this. I think it's technically OK because of the masked_fill below

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Even if you overflow it's ok because the next line you create a boolean tensor: padding_mask = padding_mask == 0. So in the end you get the same result as long as the elements in the tensor that need to be zero end up having this value.

This code, which overflows in the context of the function:

padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
padding_mask *= 2
padding_mask = padding_mask == 0

produces the same output as this code, which does not overflow:

padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
padding_mask = padding_mask == 0

padding_mask = padding_mask == 0
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
padding_mask, min_dtype
)
elif attention_mask.dim() == 4:
# backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
# cache. In that case, the 4D attention mask attends to the newest tokens only.
Expand Down
7 changes: 5 additions & 2 deletions src/transformers/models/gemma/modeling_gemma.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -989,8 +989,11 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position, curr
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
if attention_mask.dim() == 2:
mask_length = attention_mask.shape[-1]
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
padding_mask = padding_mask == 0
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
padding_mask, min_dtype
)
elif attention_mask.dim() == 4:
# backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
# cache. In that case, the 4D attention mask attends to the newest tokens only.
Expand Down
7 changes: 5 additions & 2 deletions src/transformers/models/llama/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1082,8 +1082,11 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position, curr
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
if attention_mask.dim() == 2:
mask_length = attention_mask.shape[-1]
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
padding_mask = padding_mask == 0
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
padding_mask, min_dtype
)
elif attention_mask.dim() == 4:
# backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
# cache. In that case, the 4D attention mask attends to the newest tokens only.
Expand Down
4 changes: 2 additions & 2 deletions src/transformers/training_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,12 +84,12 @@
if os.environ.get("TORCHELASTIC_RUN_ID"):
if is_optimum_neuron_available():
logger.info(
"Make sure that you are performing the training with the TrainiumTrainer from optimum[neuron], this "
"Make sure that you are performing the training with the NeuronTrainer from optimum[neuron], this "
"will fail otherwise."
)
else:
logger.warning(
"Please use the TrainiumTrainer from optimum[neuron] instead of the Transformers library to perform "
"Please use the NeuronTrainer from optimum[neuron] instead of the Transformers library to perform "
"training on AWS Trainium instances. More information here: "
"https://github.com/huggingface/optimum-neuron"
)
Expand Down
Loading