Update LLaMA attention fusions #19200
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
RyanUnderhill
previously approved these changes
Jan 19, 2024
RyanUnderhill
approved these changes
Jan 19, 2024
YUNQIUGUO
pushed a commit
that referenced
this pull request
Jan 23, 2024
### Description This PR updates the LLaMA-2 attention fusions by adding the following. - Loading the PyTorch model from Hugging Face with the `LlamaAttention` class before exporting - Updating the attention mask pattern matching to support another case This PR also fixes [this issue](#19040). ### Motivation and Context Recent changes to Hugging Face's `transformers` library break the existing pattern matching. Since the attention fusions aim to change the graph from `LayerNorm Op --> Set of Attention Nodes --> LayerNorm Op` to `LayerNorm Op --> Attention Op --> LayerNorm Op` per layer, ultimately it does not matter what nodes comprise the `Set of Attention Nodes` because they will all be removed and replaced by the `Attention Op` in the end. Therefore, it does not matter whether the `LlamaAttention` class or a different attention class is used to load the PyTorch model before exporting because the expected graphs after the attention fusions will look identical no matter the attention class chosen. By loading the PyTorch model with the `LlamaAttention` class instead of other attention classes (e.g. `LlamaFlashAttention2` or `LlamaSdpaAttention`) and then exporting it to ONNX, the existing pattern matching will continue to work.
This was referenced Jan 26, 2024
YUNQIUGUO
pushed a commit
that referenced
this pull request
Jan 30, 2024
### Description This PR updates the Whisper export with beam search by adding the following. - Fixes a bug when running `DecoderMaskedMultiHeadAttention` in the Whisper with beam search model - Sets the default PyTorch attention implementation to `eager` to allow existing attention fusions to continue working - Re-uses the cache directory when loading the PyTorch model to reduce memory used on disk - Adds `--disable_auto_mixed_precision` to the example FP16 export command ### Motivation and Context - [This PR](#19112) added the `is_unidirectional` parameter to `CheckInputs`, but it was not provided when checking the inputs in `DecoderMaskedMultiHeadAttention`. - [This PR](#19200) explains the reasoning behind why `eager` is used to load the `WhisperAttention` class. - By re-using the cache directory for loading the PyTorch model, only one copy of the PyTorch model is saved on disk instead of two copies. - By providing this flag, there will be less Cast nodes in the Whisper with beam search model to switch between FP16 and FP32 precision.
YUNQIUGUO
pushed a commit
that referenced
this pull request
Jan 30, 2024
### Description This PR updates the Whisper export with beam search by adding the following. - Fixes a bug when running `DecoderMaskedMultiHeadAttention` in the Whisper with beam search model - Sets the default PyTorch attention implementation to `eager` to allow existing attention fusions to continue working - Re-uses the cache directory when loading the PyTorch model to reduce memory used on disk - Adds `--disable_auto_mixed_precision` to the example FP16 export command ### Motivation and Context - [This PR](#19112) added the `is_unidirectional` parameter to `CheckInputs`, but it was not provided when checking the inputs in `DecoderMaskedMultiHeadAttention`. - [This PR](#19200) explains the reasoning behind why `eager` is used to load the `WhisperAttention` class. - By re-using the cache directory for loading the PyTorch model, only one copy of the PyTorch model is saved on disk instead of two copies. - By providing this flag, there will be less Cast nodes in the Whisper with beam search model to switch between FP16 and FP32 precision.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This PR updates the LLaMA-2 attention fusions by adding the following.
LlamaAttentionclass before exportingThis PR also fixes this issue.
Motivation and Context
Recent changes to Hugging Face's
transformerslibrary break the existing pattern matching. Since the attention fusions aim to change the graph fromLayerNorm Op --> Set of Attention Nodes --> LayerNorm OptoLayerNorm Op --> Attention Op --> LayerNorm Opper layer, ultimately it does not matter what nodes comprise theSet of Attention Nodesbecause they will all be removed and replaced by theAttention Opin the end.Therefore, it does not matter whether the
LlamaAttentionclass or a different attention class is used to load the PyTorch model before exporting because the expected graphs after the attention fusions will look identical no matter the attention class chosen. By loading the PyTorch model with theLlamaAttentionclass instead of other attention classes (e.g.LlamaFlashAttention2orLlamaSdpaAttention) and then exporting it to ONNX, the existing pattern matching will continue to work.