Replace FlashAttention with xformers #70
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
WoosukKwon
commented
May 4, 2023
| pip install sentencepiece # Required for LlamaTokenizer. | ||
| pip install ninja # To parallelize the compilation of flash-attn. | ||
| pip install flash-attn # This may take up to 10 mins. | ||
| pip install ninja psutil numpy sentencepiece ray torch transformers xformers |
There was a problem hiding this comment.
TODO (in the next PR): specify the exact dependencies in setup.py.
|
is the memory footprint same with flashattention? |
|
I did a test myself and found the memory saving is almost the same. |
|
It seems the memory usage is comparable to FlashAttention's. @zhuohan123 Please review the PR. |
zhuohan123
approved these changes
May 5, 2023
| @@ -213,7 +213,7 @@ def add_server_arguments(parser: argparse.ArgumentParser): | |||
| parser.add_argument('--use-np-cache', action='store_true', | |||
| help='save a numpy copy of model weights for faster loading') | |||
| parser.add_argument('--use-dummy-weights', action='store_true', help='use dummy values for model weights') | |||
| # NOTE(woosuk): FlashAttention does not support float32. | |||
| # TODO(woosuk): Support FP32 for debugging. | |||
There was a problem hiding this comment.
Yes, it does. It is our attention kernel that does not support FP32. More precisely, our attention kernel currently does not support some block sizes when FP32 is used. I will fix this in the future.
Closed
hongxiayang
pushed a commit
to hongxiayang/vllm
that referenced
this pull request
Feb 13, 2024
yukavio
pushed a commit
to yukavio/vllm
that referenced
this pull request
Jul 3, 2024
SUMMARY: for Apache 4(b) -- "You must cause any modified files to carry prominent notices stating that You changed the files" https://www.apache.org/licenses/LICENSE-2.0 TEST PLAN: GHA
dllehr-amd
pushed a commit
to dllehr-amd/vllm
that referenced
this pull request
Jul 22, 2024
* Enabling some basic tests for ROCm 6.2 Use strict xfail for ROCm 6.2 test repairs * Use lenient xfail instead --------- Co-authored-by: Alexei V. Ivanov <[email protected]>
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.
This PR replaces FlashAttention with xformers.
Pros:
Cons: