[Bugfix] Pin EPLB and MLA host-to-device transfer buffers

[khluu]

Main CI build 87842 fails both Qwen3 Sync EPLB accuracy jobs and the TP1/PCP4 evaluation when strict GPU synchronization checks detect nonblocking H2D copies from pageable CPU memory.

Pin Gloo receive staging buffers and allocate sparse MLA context lengths directly in pinned CPU memory with torch.subtract(out=...), matching the dense MLA caller. Preserve the strict checks. Add focused CUDA regressions that exercise the actual transfers and verify the copied values.

Duplicate check: searched open PRs for EPLB/pinned, unpinned, context_lens_cpu, and the failure signatures. No existing PR fixes these two transfers. PCP still separately needs #55879 for autotuning memory and #55499 for FlashInfer ragged-prefill synchronization; this PR does not duplicate them.

Current revision (a9f8512c0c), addressing Lucas Wilkinson's review:

• Moved pinning to the sparse MLA source and removed the helper-level fallback.
• Updated the CUDA regression to exercise the sparse caller with a decode followed by two prefills under strict GPU sync checking.
• Pre-commit checks passed for all three revised files (including mypy). Local runtime tests could not run: this checkout's available local environment has no torch/CUDA.
• Exact-revision attention CI passed: https://buildkite.com/vllm/ci/builds/88262 — image build and both H100 V1 Attention shards passed at a9f8512.

Earlier-revision validation (does not validate the review update above):

• pre-commit run --files vllm/distributed/eplb/eplb_communicator.py vllm/model_executor/layers/attention/mla_attention.py tests/distributed/test_eplb_execute.py tests/v1/attention/test_mla_context_chunks.py: passed.
• H100 and B200 Qwen3 Sync EPLB accuracy passed in https://buildkite.com/vllm/ci/builds/87891 . H100 high-throughput/low-latency: 0.925/0.895; B200: 0.895/0.920 (all require 0.8). The tested production patch is 4a92099; later commits only finalize regression-test setup.
• Finalized regressions passed in combined build https://buildkite.com/vllm/ci/builds/87893 at a2c24ba: pytest -v -s v1/attention/test_mla_context_chunks.py — 13 passed; pytest -v -s distributed/test_eplb_execute.py — 24 passed, 7 skipped. All three test_eplb_spec_decode.py cases skipped under their existing conditions.
• Full PCP model evaluations passed in the same combined build: TP2/PCP2 GSM8K 0.9060 and TP1/PCP4 0.9386 (minimum accepted 0.82). This validates this patch together with [Bugfix][Attention] Stabilize sparse-MLA DCP for GLM PCP evals #55879 and [Perf] Fix TRTLLM ragged prefill perf regression #55499, not this patch alone. The combined branch adds those two existing fixes and an MLA test command; it is not proposed as another PR.
• Standalone build 87891 is not green overall: its old regression-test setup failed before the latest test-only correction, and its PCP job lacks [Bugfix][Attention] Stabilize sparse-MLA DCP for GLM PCP evals #55879/[Perf] Fix TRTLLM ragged prefill perf regression #55499. Both standalone accuracy jobs and all finalized combined validation passed. No failures were suppressed and strict synchronization checking remained enabled.

AI assistance: implemented and validated with OpenAI Codex. The human submitter confirmed review of these changes and authorized PR submission. Test results above identify the automated validation performed.

[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.

[LucasWilkinson]

Overall LGTM but left a comment that I think could be a worthwhile improvement

[LucasWilkinson]

Could we pin this at the sparse MLA source instead?

context_lens_cpu = torch.empty(
    num_prefills, dtype=seq_lens_cpu.dtype, pin_memory=PIN_MEMORY
)
torch.subtract(
    seq_lens_cpu[num_decodes : num_decodes + num_prefills],
    prefill_query_lens_cpu,
    out=context_lens_cpu,
)

Then this helper-level fallback should not be necessary for production callers.

[khluu]

Addressed in a9f8512. Sparse MLA now allocates context_lens_cpu with pin_memory=PIN_MEMORY and computes directly into it with torch.subtract(out=...), matching dense MLA. Removed the helper fallback. Updated the CUDA regression to exercise the sparse caller with mixed decode/prefill inputs under strict sync checking and verify the transferred lengths. Pre-commit (including mypy) passes; exact-revision attention CI is running at https://buildkite.com/vllm/ci/builds/88262. The PR description distinguishes this pending run from the earlier GPU results.

[khluu]

/ci run

[github-actions]

✅ Triggered Buildkite CI #88264 for commit a9f8512c0cb6.