b70-optimization-lab

2026-06-27: Reordered-Q8 Grouped Multi-Token MoE Negative

Status: valid strict realistic-suite loss. Do not promote or submit to LocalMaxxing.

Context

Current strict record:

This experiment kept the same record-family launch identity and added the default-off source flag LLAMA_SYCL_MUL_MAT_ID_MULTI_TOKEN_GROUPED_Q8_0_REORDER=1.

Patch

Saved under:

The patch snapshot is against the local dirty llama.cpp Gemma research stack. It is not a minimal upstream patch. The specific experiment adds a grouped multi-token reordered-Q8 MoE path and activates it only with LLAMA_SYCL_MUL_MAT_ID_MULTI_TOKEN_GROUPED_Q8_0_REORDER=1.

Note: the run happened before the harness identity patch for this new flag, so summary.json does not contain launcher_identity.llama_sycl_mul_mat_id_multi_token_grouped_q8_0_reorder. The command/server log and patch artifact identify the tested flag. The harness now captures it for future runs.

Result

Run:

Validity:

Metrics:

Metric Value
median tok/s, tokens 1-100 after TTFT 83.90758854375754
p10 tok/s, tokens 1-100 after TTFT 78.17221426541892
mean tok/s, tokens 1-100 after TTFT 86.02284187160546
median full-512 after TTFT 82.00778999411864
median wall full-512 79.09301505839233
median TTFT 179.04035048559308 ms

Conclusion

The idea lost despite being valid. The route profile correctly showed duplicate expert hits, but grouping them in the reordered-Q8 path appears to add more branch/scatter/register pressure than it saves at the strict n_max=3 verifier shape. Do not repeat this exact path without a lower-level kernel profile that shows the grouping overhead can be removed.

Next higher-signal work should target verifier-side structural cost that affects fresh responses directly: target LM-head/output selection, draft acceptance quality that raises accepted tokens without history reuse, or a tighter source-level profile of the VDR2 verifier step before writing another MoE kernel variant.