[Kernel][SM70] Accelerate Qwen3.8 QSA paged prefill

[yangzhuxinyzx]

Purpose

Accelerate Qwen3.8 selected sparse prefill on SM70/V100 by exposing each four-token QSA selection as a zero-copy Flash-V100 XQA page while preserving task-relevant numerical quality.

Changes

• Convert the exact FP16 Hq6/Hkv1/D256, 2051-token QSA selection layout into locality-sorted virtual page-4 KV.
• Route prefill batches with at least 4096 real query rows by default; retain VLLM_SM70_QSA_XQA_PAGE4=0 as an operational escape hatch.
• Support both contiguous and interleaved K/V cache strides without copying cache data.
• Keep the causal partial page last and validate it against the expanded QSA indices.
• Clamp padded/stale positions to live sequence lengths and make invalid requests inert.
• Expand the Flash-V100 per-tile page-ID capacity without reducing declared V100 occupancy.
• Forward QSA positions/lengths, add focused route and metadata tests, and record the audit in the migration control document.
• Include [Bugfix][QSA] Keep compressed block tables in physical-page units #381 so the compressed QSA scorer consumes scheduler-level physical page IDs before this attention fast path runs.

Test Result

• GitHub full pre-commit on the repaired latest-main branch: passed.
• Combined route, metadata, and [Bugfix][QSA] Keep compressed block tables in physical-page units #381 physical-page tests on V100: 15 passed.
• Full SM70 extension build: passed with CUDA 12.0 / Torch cu128 for sm_70.
• 4095 rows: exact Triton fallback, identical with or without forwarded metadata.
• 4096-row interleaved-KV A/B: 27.8303 -> 8.84736 ms including table/sort (3.1456x); max abs 6.104e-5, relative L2 2.845e-4, cosine 0.99999994, all finite.
• Nonmonotonic contiguous-page A/B: max abs 3.815e-6, relative L2 3.645e-4, cosine 1.0.
• Tail counts 1/2/3: relative L2 at most 3.64e-4, cosine at least 0.99999988.
• Hybrid 784/16 geometry with a nonmonotonic 128-entry page-16 table: max abs 3.815e-6, relative L2 3.631e-4, cosine 0.99999994; CUDA Graph replay is bitwise equal to eager output.
• Page-16 context sweep at 4096 rows:
  • 64 selected tokens: 29.225 -> 1.010 ms (28.95x)
  • 512 selected tokens: 27.063 -> 3.204 ms (8.45x)
  • 2048 selected tokens: 27.835 -> 11.203 ms (2.48x)
  • every arm: relative L2 below 3.84e-4, cosine at least 0.99999994
• Prewarmed CUDA Graph capture and two replays are bitwise equal to eager page4 output, hash 9b4c76f8420d6e349dc7d552c72d6f0a861332e7e8e8f62459a1c48f0faf278f.
• Existing FP16 page-16/page-784 XQA-to-scalar smokes pass at relative L2 3.03e-5 / 4.10e-5.

Decision

The route is materially faster across short, medium, and saturated selected contexts, and its differences are small accumulation/order effects rather than quality regressions. It remains default-on under the project policy.

Repair chain: #382. Final candidate head: 7c4c795f5db1542ef0e496d3215afaef04b00d9f; tree: a6d9af1056de0db77294f2209f86c39e24eb5def.

[yangzhuxinyzx]

Repair #382 is now merged into this branch. The final candidate is 7c4c795 (tree a6d9af1056de0db77294f2209f86c39e24eb5def), includes #381, and keeps the audited performance route default-on. Final full pre-commit is running on this exact head.