[Perf]:Optimize qwen2-vl to reduce cudaMemcpyAsync #14377
Merged
Conversation
This file contains hidden or 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
Isotr0py
approved these changes
Mar 7, 2025
cynthieye
changed the title
cynthieye
force-pushed
the
main
branch
3 times, most recently
from
ae09649 to
1fbb69c
Compare
github-actions
bot
added
the
ready
auto-merge was automatically disabled
March 10, 2025 13:26
Head branch was pushed to by a user without write access
Signed-off-by: cynthieye <987073381@qq.com>
1 task
Signed-off-by: cynthieye <987073381@qq.com>
|
@cynthieye Thank you for making this PR! Can you update this branch with our main branch? I think thr CI error should be fixed on main a while ago. |
ywang96
approved these changes
Mar 11, 2025
| @@ -259,6 +259,8 @@ def forward( | |||
| x: torch.Tensor, | |||
| cu_seqlens: torch.Tensor, | |||
| rotary_pos_emb: torch.Tensor, | |||
| max_seqlen: int = None, | |||
There was a problem hiding this comment.
Shouldn't max_seqlen be also Optional[int]?
Comment on lines
372
to
373
| max_seqlen: int, | ||
| seqlens: list[int], |
There was a problem hiding this comment.
Please modify the typing accordingly
Comment on lines
310
to
311
| max_seqlen: int = None, | ||
| seqlens: Optional[list[int]] = None, |
Comment on lines
417
to
418
| max_seqlen: int, | ||
| seqlens: list[int], |
Comment on lines
372
to
373
| max_seqlen: int, | ||
| seqlens: list[int], |
There was a problem hiding this comment.
I think it's probably a good idea to add a small documentation here to indicate that max_seqlen is only used for FA and seqlens is only used to xformers.
Signed-off-by: cynthieye <987073381@qq.com>
Signed-off-by: cynthieye <987073381@qq.com>
This was referenced Mar 20, 2025
lulmer
pushed a commit
to lulmer/vllm
that referenced
this pull request
Apr 7, 2025
Signed-off-by: cynthieye <987073381@qq.com> Signed-off-by: Louis Ulmer <ulmerlouis@gmail.com>
shreyankg
pushed a commit
to shreyankg/vllm
that referenced
this pull request
May 3, 2025
Signed-off-by: cynthieye <987073381@qq.com>
RichardoMrMu
pushed a commit
to RichardoMrMu/vllm
that referenced
this pull request
May 12, 2025
Signed-off-by: cynthieye <987073381@qq.com> Signed-off-by: Mu Huai <tianbowen.tbw@antgroup.com>
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.
qwen2-vl logic optimization: During each forward propagation, the xformer branch of Qwen2VisionTransformer will execute multiple tensor tolist methods (flash attn branch will execute multiple tensor items) to force the GPU tensor to be copied to the CPU, triggering CUDAMemcpyAsync to increase time consumption. Since the input and output are the same multiple times, it will be executed once, and the remaining will reuse the first result. After optimization, the online environment xformer branch QPS can be improved by 15%, and the flash attn branch QPS can be improved by 7%