[Kernel][SM70] Default long-prefill GQA architecture

[yangzhuxinyzx]

Purpose

• Add a default-on SM70 attention architecture for the validated causal FP16 Q=8000, KV=40000..128000 (step 8000), Hq=6, Hkv=1, D=256, scale 1/16 long-prefill family.
• Route only by device capability and tensor/runtime properties. There is no model name, checkpoint, model_type, or architecture-identity gate.
• Pack six GQA query heads into wide QK/PV Tensor-Core GEMMs, retain one reusable block partial, and merge the exact causal tail with online-softmax state.
• Retain one 750-MiB per-device score workspace, serialize cross-stream reuse with an event and mutex, and preflight the persistent plus transient allocation with 128 MiB of downstream headroom.
• Keep VLLM_FLASH_V100_PREFILL_D256_GQA_ARCH_128K_EXPERIMENTAL=0 as the legacy rollback setting. CUDA graph capture and unsupported shapes fall through to existing paths; typed workspace OOM falls back to the exact dense route.

Test Plan

• pytest -q tests/v1/attention/test_sm70_flash_v100_policy.py on a real V100.
• Changed-file pre-commit, including Ruff, mypy, Markdown, SPDX, forbidden-import, environment-default, and backend-documentation checks.
• Replay all four SM70 FA2 patches from the locked pristine vendored baseline and build/load _vllm_fa2_C with CUDA 12.8, Torch cu128, and sm_70.
• Direct current-source dense/candidate A/B at the 40K and 128K endpoints.
• Retain the existing strict 12-shape A/B and matched TP4 control/candidate/control evidence.

Test Result

• Final policy suite: 112 passed on V100, including direct architecture-OOM-to-dense-fallback coverage.
• All changed-file pre-commit hooks pass; git diff --check passes.
• Clean patch replay, CUDA 12.8 SM70 build, Torch operator schema/load, and launch pass. QK/PV compile at 254/119 registers with zero spill.
• Current-source KV40K: dense 39.29395 ms, candidate 30.27579 ms; 1.29787x speedup and 22.95% lower latency.
• Current-source KV128K: dense 136.07628 ms, candidate 100.20284 ms; 1.35801x speedup and 26.36% lower latency.
• KV40K numerical result: all outputs finite, max/mean absolute difference 5.4932e-4 / 4.4392e-5, relative L2 6.6098e-3.
• KV128K numerical result: all outputs finite, max/mean absolute difference 2.2888e-4 / 2.5484e-5, relative L2 6.9166e-3.
• Both pass the FP16 merge envelope of max absolute error <=1e-3 and relative L2 <=1e-2. Bitwise and greedy token identity are not promotion requirements.
• Existing 12-shape evidence: every 40K..128K point is faster, with 1.28848x..1.35485x speedup; all 147,456,000 candidate elements are finite, worst max absolute difference 5.0354e-4, worst relative L2 6.8983e-3.
• Existing matched widened TP4 gate: bracketed control 46.10208 s, candidate 41.51191 s; prefill latency improves 9.9565% and prompt throughput improves 11.0575%. The identical 32-token smoke is retained only as route-health evidence.

Safety

The default route is deliberately narrow and capability-based. Unsupported shapes, layouts, dtypes, scales, devices, paged KV, and CUDA graph capture do not enter it. A missing optional operator falls through. A typed persistent or transient workspace OOM is caught and executes the existing exact fallback instead of failing the request.

[yangzhuxinyzx]

Audit result: not mergeable in its current form. The proposed architecture changes the reduction order and reports worst relative L2 near 6.9e-3, while quality evidence is limited to one fixed prompt. The 750 MiB persistent-workspace OOM fallback also lacks direct model-path validation. Keeping a default-off 1,673-line vendored kernel patch does not resolve those correctness and maintainability risks. The branch and evidence remain available; resubmit a focused PR after an exact reduction or broader paired quality gate plus direct fallback validation.

[yangzhuxinyzx]

最终审计通过：默认路由仅由 SM70 能力和张量运行时契约决定，不含模型/检查点身份门。当前源码干净重放并完成 CUDA 12.8 sm_70 构建；V100 策略测试 112 passed；KV40K/128K 分别为 1.29787x/1.35801x，输出全 finite，max abs 与 relative L2 均通过 1e-3/1e-2 FP16 包络；typed OOM 已验证回落 exact dense。changed-file pre-commit 全过。GitHub 全仓检查失败来自 227 个历史格式化基线文件，与本 PR 7 个文件重叠为 0，判定为非阻断基线噪声。