Exp-2026-08 — int4 MoE BCHAIN A/B knobs (ORK_MOE_MPAD, ORK_I4_M1_NPU): flat/superseded, REMOVED

Repo: ggml-ork (ggml-ork.cpp, NPU int4 BCHAIN expert path). Status: results banked; knobs REMOVED from mainline to cut sprawl. Preserved: branch wip/moe-saga-2026-08-18 (the guard lines are

Patch (in this zip): parent-experiments-stashed.patch (mpad + m1npu guards in ggml-ork.cpp) — git apply onto the banked tree to restore this experiment; the git-native copy is the snapshot branch above.

recoverable there). Knobs: ORK_MOE_MPAD (3603), ORK_I4_M1_NPU (3516).

ORK_MOE_MPAD — M-pad for the BCHAIN NG count

Hypothesis: shrinking the M-pad (default 32) halves NG for the common Mr~16 → fewer/bigger programs → faster. Result: FLAT in practice (routed Mr mostly rounds to 32 anyway; gate/up NG isn't the dominant cost; isolated single-matmul M=16 gain didn't translate). Default kept at 32 = baseline. No follow-up.

ORK_I4_M1_NPU — force the old M=1 per-row int4 path

The M=1 (decode) per-row int4 path is superseded by routing M=1 to the M≥2 BCHAIN case; ORK_I4_M1_NPU only forced the OLD per-row path for A/B. The new default is correct and faster; the escape hatch is no longer needed. No follow-up.

Conclusion

Both are inert A/B escape-hatches on the already-committed int4 BCHAIN path (default = baseline behavior), independent of the CPU NF4 prefill win. Their conclusions are recorded (memory moe-w4a4-prefill-wall); the knobs are removed to reduce surface. Recoverable from wip/moe-saga-2026-08-18 if ever needed for another A/B. No follow-up required.