Filter sparse MLA top-k indices per DCP rank, and fix decode LSE

[pramodp-dotcom]

Backport onto the v0.28.0 release, which does not carry the upstream fix: #46514 (63ff748, 2026-08-19) landed this on main, but the v0.28.0 release branch forked before it, so the released wheel still has the bug. flashmla_sparse.py is byte-identical from v0.26.0 through v0.28.0, so this applies unchanged across all three.

Under decode context parallelism each rank holds only its own shard of the KV cache, but the sparse FlashMLA path fed the top-k results - which are global token positions - straight into
triton_convert_req_index_to_global_index. That converter assumes every index is local, so it translated positions owned by other ranks through this rank's block table and read unrelated cache slots, corrupting decode output.

Use triton_filter_and_convert_dcp_index when dcp_world_size > 1: it drops the indices this rank does not own before converting the rest. Applied at all three call sites (bf16 KV, fp8 KV separate prefill/decode, fp8 KV mixed batch). Non-DCP runs keep the original converter unchanged. The helper is already in sparse_utils.py at this tag.

Ownership depends on the interleave stride, so plumb cp_kv_cache_interleave_size from parallel_config through the metadata rather than leaving the dataclass default of 1.

Two LSE fixes needed for the cross-rank reduction to be correct:

• flash_mla_sparse_fwd returns (output, max_logits, lse), so read kernel_out[2]. The previous kernel_out[1] fed max_logits into the reduction. Upstream main expresses the same thing as 'output, _, lse = flash_mla_sparse_fwd(...)'.
• flash_mla_with_kvcache returns LSE as (batch, num_heads, seq_len) while callers reshape assuming (batch, seq_len, num_heads); transpose it, and trim the padded head dim.

Also guard the two LSE reshape sites against a None lse.

Claude-Session: https://claude.ai/code/session_01MYhgsjgQ2GQcpkXZfM3dec

Purpose

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[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.

PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment /ci run for upstream CI or /amd-ci run for AMD CI only whenever CI signals are needed.

Once the PR is approved or has the ready label, the PR author can also use the corresponding /ci run, /ci retry, and /ci cancel commands, or their /amd-ci variants. New commits do not start upstream CI automatically.

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

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀