[yangzhuxinyzx]

Purpose

Rebuild and repair PR #302 on the latest main at 1d64f84.

The original PR changed every admitted PP2 x TP4 QPN8 projection to its lowest-error schedule. Source audit shows that this pays a 15--36% isolated operator slowdown for errors that are already bounded around 1e-5. The actual numerical outlier is only fused shared-expert gate/up.

This repair therefore:

• removes shared-expert gate/up from the default QPN8 operator contract so it falls back to TurboMind;
• retains the existing measured speed-winner schedules for WQA/WKV, WQ-B, WO-B, grouped WO-A, and shared down;
• keeps the replicated indexer excluded;
• preserves the existing PP2 x TP4 B1 no-spec hardware/topology/tensor admission and rollback controls;
• never reads model name, checkpoint path, model_type, or architecture identity.

This PR supersedes #302 rather than pushing into the agent-owned branch.

Numerical and performance audit

Four-pattern real-weight comparison against same-process TurboMind:

• retained fast schedules: exact FP16 fraction at least 99.707%, relative L2 at most 1.641e-5;
• fused shared gate/up: exact fraction 53.516%, relative L2 5.989e-4;
• accuracy-first alternatives for the retained roles are about 15--36% slower in the isolated operator matrix.

Matrix artifact:
/data/models/v100-dsv4-0731-pp2tp4-qpn8-accuracy-matrix-20260826-r1/

The conservative all-accuracy source gate passes M=1/M=9, changing-input CUDA Graph replay, FP32/TurboMind references, grouped WO-A layout, both excluded roles, and one reused 16 MiB workspace:
/data/models/v100-dsv4-0731-pp2tp4-qpn8-accuracy-source-gate-20260826-r2-main/

Matched PP2 x TP4 B1 no-spec evidence for that slower all-accuracy variant:

• QPN8 off median: 63.6434 tok/s, 15.71254 ms/token;
• all-accuracy QPN8 median: 63.9120 tok/s, 15.64652 ms/token;
• difference: about +0.42%, -0.066 ms/token;
• pinned GSM8K64: 62/64 versus 61/64, zero invalid outputs, one directional difference at item 37.

The single GSM8K difference is recorded rather than treated as a greedy-identity failure. It also provides no evidence that selecting the lowest already-small operator error improves aggregate task quality. The accepted source repair is consequently the speed-preserving hybrid: retain every bounded speed winner and fall back only the demonstrated outlier.

Full artifacts:
/data/models/v100-dsv4-0731-pp2tp4-qpn8-accuracy-fullmodel-control-20260826-r1/
/data/models/v100-dsv4-0731-pp2tp4-qpn8-accuracy-fullmodel-candidate-20260826-r2/

Test plan and result

• 9 focused PP2 x TP4 QPN8 route, rollback, workspace, and grouped-dispatch tests: passed under Python 3.12, Torch 2.10, and the matching source-built extension.
• pre-commit run -a on latest main: all hooks passed.
• git diff --check: passed.
• no new end-to-end run: the user-requested audit is source/operator focused, and the matched artifacts above already close the conservative variant.