[yangzhuxinyzx]

Purpose

Use real-weight, four-pattern accuracy-first QPN8 schedules for the existing SM70 PP2 x TP4 route. Remove shared-expert gate/up from QPN8 admission because its fused activation has low element agreement with TurboMind.

Test Plan

• Run focused CPU route and dispatch suites.
• Run real DeepSeek V4 layer-0 weights at M=1 and M=9 against FP32 and TurboMind.
• Verify changing-input CUDA Graph stability, grouped WO-A layout, excluded gate/up and indexer roles, and bounded workspace reuse.
• Before promotion, run clean eight-GPU matched pure decode plus pinned GSM8K quality gates and collect a new trace.

Test Result

• 29 focused CPU tests passed.
• Latest-main one-V100 source gate passed all five admitted projections and both exclusion checks.
• Maximum QPN8-vs-TurboMind relative L2 was 1.4264e-5; one 16 MiB workspace was reused.
• Eight-GPU endpoint and dataset gates are pending, so this PR remains Draft.

[yangzhuxinyzx]

Superseded by #313, which rebuilt this audit on the latest main, retained the numerically bounded speed-winner schedules, fell back only the fused shared gate/up outlier, passed focused tests and whole-repository pre-commit, and is now merged as f5fc157. The optimization goal is merged rather than rejected.

[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]

Audit repair #367 preserves this PR history while syncing its final source tree exactly to current main (6cf99a61...). The validated QPN8 functionality is already on main through #313, so no stale policy/test code is replayed. Waiting for locked-head CI before merging #367 and then #302.

[yangzhuxinyzx]

源码审计完成：该 PR 的 PP2/TP4 QPN8 调度修复已由 #313 的更新实现吸收。通过 #367 同步后，当前 head 的 Git tree 与最新 main 完全一致（6cf99a61fb720978b84134b09ddf58ffd5eb96ca），全量 pre-commit 通过；以零源码差异合并，保留原 PR 历史。