[Kernel][SM70] Prescale DeepSeek V4 FP8 decode

[yangzhuxinyzx]

Purpose

Remove the redundant E4M3 exponent-bias FP16 multiply from the exact SM70 DeepSeek V4 M=1 fused-WQA/WKV path by absorbing its power-of-two factor into UE8M0 scales at load time. The route is opt-in and restricted to the measured PP2 x TP4 no-spec single-request contract.

Test Plan

• Build the SM70 extension for compute capability 7.0.
• Run real-weight CUDA Graph A/B/B/A on five V4 FP8 shapes with changing inputs.
• Run the focused Python routing/admission tests.
• Run matched full-model GSM8K-64 and three pure-decode repeats before promotion.

Test Result

• Real layer-0 TP4-rank-0 operator screen: 320/320 changing-input comparisons bitwise equal.
• Fused WQA/WKV: 63.330 us to 20.577 us (3.078x).
• Five-shape weighted projection: -1.867 ms/token warm, -1.821 ms/token after cache scrub.
• pytest -q tests/quantization/test_sm70_fp8_prefill_exact_dense.py -k prescaled: 3 passed.
• Ruff and git diff --check: passed.
• Full-model endpoint and pinned dataset gates: pending; keep Draft.

[yangzhuxinyzx]

Audited against latest main. The original stacked branch was not replayed wholesale because it was opt-in, reused an older prefill-only operator as its capability signal, and prescaled FP16 values without an actual finite/reversibility gate.

The performance work was repaired in #316: generic SM70 M=1 tensor/runtime admission with no model or checkpoint identity, default-on behavior, explicit rollback, safe fallback for older extensions and unsupported scale ranges, a distinct M=1 operator contract, targeted tests, and full local/remote pre-commit. The recorded 320/320 changing-input comparisons remain bitwise equal and the fused projection remains 3.078x faster.

#316 is merged into main at d9d0da2. Closing #309 only because its audited and repaired functionality is now present on main.

[yangzhuxinyzx]

公开开发已恢复：本 PR 已取消归档、解锁并重新打开。为避免把旧基线上的问题直接带入最新 main，现统一作为 Draft 审计项；恢复可见性不代表已经批准合并。 该组涉及 SM70/V100 性能或数值路径，需先与当前主线及 #344–#351 去重，再用同条件的聚焦质量和性能证据决定修复后合并。

[yangzhuxinyzx]

审计结论：安全的 FP8 prescale 实现已由后续 PR 吸收；该分支已通过修复 PR #370 同步到最新 main，最新完整 pre-commit 通过，且锁定提交的源码树与 main 完全一致。按零差异历史归档合并，不改变主线源码。