Skip to content

fix overlap pagecount #6984

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

Merged
hnyls2002 merged 2 commits into from
Jun 21, 2025
Merged

fix overlap pagecount #6984

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
daf7085
fix overlap pagecount
pansicheng Jun 9, 2025
39b9922
Merge branch 'main' into fix-overlap-pagecnt
xiezhq-hermann Jun 21, 2025
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
6 changes: 5 additions & 1 deletion python/sglang/srt/managers/schedule_batch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1359,7 +1359,11 @@ def new_page_count_next_decode(self):
return len(self.reqs)
# In the decoding phase, the length of a request's KV cache should be
# the total length of the request minus 1
return sum(1 for req in self.reqs if (req.seqlen - 1) % page_size == 0)
return (
sum(1 for req in self.reqs if req.seqlen % page_size == 0)
if self.enable_overlap
else sum(1 for req in self.reqs if (req.seqlen - 1) % page_size == 0)
)
Copy link
Collaborator

@xiezhq-hermann xiezhq-hermann Jun 12, 2025

Choose a reason for hiding this comment

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

Does this fix prevents OOM or simply reduce the chance of retracting decode? If it is the later case, let's keep it simple and conservative.

Copy link
Collaborator Author

@pansicheng pansicheng Jun 13, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

The -1 adjustment here is primarily to align the calculation of new pages in non-overlap mode with the actual allocation code logic (where req.seqlen == seqlens in non-overlap scenarios). Without this -1 adjustment, there would be a discrepancy between the estimated new pages during scheduling and the actual allocation, which could ultimately lead to OOM

The actual allocation code logic: https://github.com/sgl-project/sglang/blob/v0.4.7/python/sglang/srt/mem_cache/paged_allocator.py#L133

image


def check_decode_mem(self, buf_multiplier=1):
tokens_required = (
Expand Down
Loading