Skip to content

Fix the default chunked prefill size #2268

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

Jump to bottom
Merged
merged 1 commit into from
Nov 30, 2024
Merged

Fix the default chunked prefill size #2268

merged 1 commit into from
Nov 30, 2024

Conversation

merrymercy
Copy link
Contributor

@merrymercy merrymercy commented Nov 29, 2024 •
edited
Loading

This is a follow up for the chunked prefill size adjustment in #2225

  • Only do chunk size adjustment when the chunked prefill size is not specified by the arguments, so we do not silently change users' arguments.
  • Use a smaller cuda graph max bs for small memory GPUs because cuda graph typically does not bring speedup on these GPUs. A smaller cuda graph max bs saves memory and prevents OOM.

@merrymercy merrymercy changed the title Fix chunked prefill size Fix the default chunked prefill size Nov 29, 2024
@merrymercy
Copy link
Contributor Author

cc @BBuf

@merrymercy merrymercy merged commit 94e167e into main Nov 30, 2024
9 of 15 checks passed
@merrymercy merrymercy deleted the pr-fix-cuda-graph branch November 30, 2024 00:03
@BBuf
Copy link
Collaborator

BBuf commented Nov 30, 2024

cc @BBuf

Thanks. Actually, in my nsight-system profiling of Qwen2.5 inference on HuggingFace, I observed that using cuda-graph made no difference because the kernel launch time was already at nanosecond level. I'd like to know the reason for this. In contrast, on A800, without using cuda graph, the kernel launch time during decoding phase is even longer than the inference time itself. With cuda graph enabled, the time for a complete decoding iteration can be reduced by half.

@merrymercy
Copy link
Contributor Author

merrymercy commented Dec 1, 2024 •
edited
Loading

@BBuf I do not fully understand that as well. Based on my e2e test. On A100/H100, cuda graph is very useful. On low-end GPUs (3090, A10G, L40), it almost has no effect when tp=1.

I do not have time to dig into that, but you can play with it more.

@BBuf BBuf mentioned this pull request Dec 5, 2024
Merged
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
@merrymercy
Fix the default chunked prefill size (sgl-project#2268)
537389a
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@Ying1123 Ying1123 Awaiting requested review from Ying1123 Ying1123 is a code owner

@hnyls2002 hnyls2002 Awaiting requested review from hnyls2002 hnyls2002 is a code owner

@zhyncs zhyncs Awaiting requested review from zhyncs zhyncs is a code owner

@ispobock ispobock Awaiting requested review from ispobock ispobock is a code owner

@ByronHsu ByronHsu Awaiting requested review from ByronHsu

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@merrymercy @BBuf