[bugfix] fix wrong dcp_local_seq_lens calc

[pisceskkk]

When the seq_lens is exactly divisible by the dcp_world_size, it causes the dcp_local_seq_lens on all ranks to be incremented by one. Fix this calculation logic.

CC @youzhedian @minosfuture @youkaichao

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors.

You ask your reviewers to trigger select CI tests on top of fastcheck CI.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[gemini-code-assist]

Code Review

This pull request addresses a critical bug in the calculation of local sequence lengths for decode requests under distributed context parallelism (dcp_local_seq_lens). The previous logic failed when the sequence length was perfectly divisible by the world size, incorrectly assigning an extra token to every rank. The fix implements the standard and correct method for distributing remainder tokens, ensuring the calculation is accurate in all scenarios. This change is essential for the correctness of distributed computations.

[gemini-code-assist]

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

[minosfuture]

Thanks for the fix!