Skip to content

[bugfix] fix wrong dcp_local_seq_lens calc #27518

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
merged 1 commit into from
Nov 5, 2025
+1 −1
Merged

[bugfix] fix wrong dcp_local_seq_lens calc #27518

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion vllm/v1/attention/backends/mla/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -799,7 +799,7 @@ def build(
dcp_local_seq_lens[:num_decodes] = seq_lens[
:num_decodes
] // self.dcp_world_size + (
self.dcp_rank <= (seq_lens[:num_decodes] - 1) % self.dcp_world_size
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Oct 26, 2025

Choose a reason for hiding this comment

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

critical

The previous logic for distributing remainder tokens, self.dcp_rank <= (seq_lens[:num_decodes] - 1) % self.dcp_world_size, was flawed. When a sequence length L is perfectly divisible by the world size W, (L - 1) % W evaluates to W - 1. This caused the condition to be true for all ranks, incorrectly incrementing each rank's local sequence length by one.

The new logic, self.dcp_rank < seq_lens[:num_decodes] % self.dcp_world_size, correctly handles this. When L % W == 0, no rank gets an extra token. When there is a remainder, it is correctly distributed among the first L % W ranks. This change fixes the bug and ensures correct behavior for all cases.

Suggested change
self.dcp_rank <= (seq_lens[:num_decodes] - 1) % self.dcp_world_size
self.dcp_rank < seq_lens[:num_decodes] % self.dcp_world_size

self.dcp_rank < seq_lens[:num_decodes] % self.dcp_world_size
)

assert num_decodes + num_prefills == num_reqs
Expand Down