[Kernel][SM70] Optimize DeepSeek V4 PP2 TP4 decode hot paths

[yangzhuxinyzx]

Purpose

Stack the exact PP2 x TP4 DeepSeek V4 Flash decode candidates after the QPN8 audit branch:

• use 16-head sparse MLA grouping on TP4;
• fuse Q normalization/RoPE with KV insertion;
• broadcast the physical B1 MXFP4 W13 input;
• keep B1 top-6 routes in original order to remove redundant sort/inverse-permutation work;
• specialize the exact top-k-6 weighted reduction with half2.

All new runtime behavior remains opt-in while matched endpoint and aggregate-quality gates are pending.

Test Plan

• compile the SM70 CUDA translation unit with CUDA 12.8 and link an SM70-only extension;
• run the weighted-reduce exactness regression on V100;
• compare generic, current direct-top6, and direct-order MXFP4 pipelines under CUDA Graph, including route IDs changed after capture;
• run matched eight-V100 PP2 x TP4 performance and aggregate-quality gates when the shared node is free.

Test Result

• ruff check passed for changed Python files;
• CUDA 12.8 source build passed; extension SHA256 3cbf943f2a417c8632368d0890c80bf5e5c446a31a9f0f18c12b9555b49f6fad, SM70 cubins only;
• test_sm70_single_token_weighted_reduce_matches_moe_unpermute: 1 passed (including new top-k=6, H=4096 case);
• MXFP4 CUDA Graph pipeline: current direct-top6 54.414 us/layer, direct-order 46.519 us/layer, projected 0.339 ms/token service reduction across 43 layers;
• W13, clamped SwiGLU, W2, and final output are bitwise equal initially and after dynamic route change;
• artifact: /data/models/v100-dsv4-0731-pp2tp4-mxfp4-direct-order-screen-20260826-r1/results/full_pipeline.json.

Endpoint and dataset results are intentionally pending.

[yangzhuxinyzx]

Superseded by merged PR #314. The four hot paths were rebuilt on latest main, source-audited independently, repaired to align Python/C++ default semantics and token bounds, defaulted with explicit rollbacks, and passed focused tests plus whole-repository pre-commit. The optimization stack is merged as 1b1c9fa rather than rejected.

[yangzhuxinyzx]

Post-endpoint quality update: the first direct-order run reached 69.949 token/s but produced incoherent text and is rejected. Root cause is cross-shape state reuse: M2-M8 active-expert compaction overwrites compact_expert_offsets before B1 graph capture, while the direct-order branch assumed fixed one-row offsets. Commit 052312c switches B1 direct-order to immutable slot_expert_offsets and adds a poisoned-workspace regression; the focused CPU tests pass (2 passed). Full-model and GSM8K-64 revalidation remain pending, so this PR stays Draft.

[yangzhuxinyzx]

Superseded in full by current-main repairs: #314 merged the four audited/defaulted hot paths, and #326 now merged the post-endpoint direct-order offset isolation as b06f54ce. The bad 69.949 token/s run remains recorded as rejected quality evidence; main now uses immutable slot offsets so M2-M8 warmup cannot poison B1 graph capture. This closes the stale stacked branch only after both the optimization and its newly reported bugfix landed.

[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：原始实验路径已由后续安全实现吸收；该分支已通过修复 PR #369 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。