[Kernel][SM70] Repair #378 on latest main

[yangzhuxinyzx]

Purpose

Repair #378 on the latest public main while preserving its default-on SM70 QSA page4 XQA acceleration and contributor ancestry.

Changes

• Preserve original [Kernel][SM70] Accelerate Qwen3.8 QSA paged prefill #378 head ba759b85836fbca326567af8e06f0d434476fba7 as an ancestor.
• Convert selected four-token QSA blocks into locality-sorted virtual Flash-V100 pages for the exact FP16 Hq6/Hkv1/D256 prefill shape at 4096 or more rows.
• Keep the accelerated route default-on, with VLLM_SM70_QSA_XQA_PAGE4=0 as an operational escape hatch.
• Clamp padded/stale positions to the live request length and admit a partial tail only when the expanded QSA indices contain that exact token.
• Add route-threshold, metadata-forwarding, invalid-request, stale-tail, and physical-tail mapping tests.
• Record the locked V100 quality/performance, fallback, CUDA Graph, and existing-XQA compatibility evidence.

Test Result

• Focused CPU contracts before final main sync: 9 passed.
• Combined QSA route, metadata, and [Bugfix][QSA] Keep compressed block tables in physical-page units #381 physical-page contracts on V100:
  15 passed.
• Full SM70 extension build from the locked original C++ source: passed with CUDA 12.0 / Torch cu128 for sm_70.
• Changed-file pre-commit: passed, including Ruff, clang-format, mypy, Markdown, and the accelerator API gate.
• Python compile: passed.
• 4095-row boundary: 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; Graph replay is
  bitwise equal to eager output.
• Page-16 context sweep at 4096 rows remains positive from short to saturated
  selections: 64 tokens 29.225 -> 1.010 ms (28.95x), 512 tokens
  27.063 -> 3.204 ms (8.45x), and 2048 tokens 27.835 -> 11.203 ms
  (2.48x); every arm passes relative L2 below 3.84e-4 and cosine at least
  0.99999994.
• CUDA Graph: capture passed; two replays are bitwise equal to eager output, hash 9b4c76f8420d6e349dc7d552c72d6f0a861332e7e8e8f62459a1c48f0faf278f.
• Existing FP16 page-16/page-784 XQA smokes: relative L2 3.03e-5 / 4.10e-5, cosine approximately 1.0.
• Shared-memory audit: the two-block page4 and pipeline variants remain below V100's 96 KiB limit, so the enlarged page-ID table does not reduce declared occupancy.

Locked candidate

• latest-main base including [Bugfix][QSA] Keep compressed block tables in physical-page units #381: 7f02cebc49398e0310e0514f04c10a2f5e3dd17e
• repair head: 8cd30655d13d4d91d83b329924850180a0b72461
• repair tree: a6d9af1056de0db77294f2209f86c39e24eb5def

Draft repair for #378. Merge only after GitHub reports this exact head and the final full pre-commit succeeds.

[yangzhuxinyzx]

审计完成并锁定 8cd30655d13d4d91d83b329924850180a0b72461（tree a6d9af1056de0db77294f2209f86c39e24eb5def）。

• 已包含最新 main 与 [Bugfix][QSA] Keep compressed block tables in physical-page units #381 的 QSA 物理页修复。
• 远端完整 pre-commit 通过；组合聚焦测试 15/15 通过。
• V100 page-16/page-784、连续/交错 KV、尾部 0/1/2/3、4095 回退和 CUDA Graph 捕获/重放均通过。
• 4096 行 page-16 上，64/512/2048 token 选择分别约加速 28.95×/8.45×/2.48×；所有质量臂 relative L2 小于 3.84e-4、cosine 至少 0.99999994。
• 路由保持默认开启，同时保留环境变量关闭开关；修复了过期位置、无效请求与虚假尾页边界。

准备按锁定提交合并回 #378。