b70-optimization-lab

20260623T0925 N7 P-Min And P-Split Sweeps

Goal: tune the n=7 filled-long frontier after the 87.878 tok/s p-min 0.15 record. This batch tested p-min variants, n=8, and p-split interaction.

Common successful-run identity:

0925 Sweep

GPU Label Change Canary tok/s after TTFT tok/s wall Decision
0 gemma4-q8-gpu0-mtp-n7-aot-nmin2-pmin015-repeat-filled-long-deep-20260623T092524Z repeat n=7, p-min=0.15 no summary n/a n/a Startup hang before readiness; preserved as negative artifact.
1 gemma4-q8-gpu1-mtp-n7-aot-nmin2-pmin010-filled-long-deep-20260623T092524Z n=7, p-min=0.10 384/384 88.345 80.553 New valid record; submitted to LocalMaxxing.
2 gemma4-q8-gpu2-mtp-n7-aot-nmin2-pmin020-filled-long-deep-20260623T092524Z n=7, p-min=0.20 no summary n/a n/a Startup hang before readiness; preserved as negative artifact.
3 gemma4-q8-gpu3-mtp-n8-aot-nmin2-pmin015-filled-long-deep-20260623T092524Z n=8, p-min=0.15 384/384 60.782 56.980 Valid but large regression; n=8 has too much draft overhead on this path.

Submitted record:

0931 Opportunistic Sweep

After the n=8 and p-min 0.10 runs completed, GPU1 and GPU3 were reused while the two hung 0925 startup attempts were still being investigated.

GPU Label Change Canary tok/s after TTFT tok/s wall Decision
1 gemma4-q8-gpu1-mtp-n7-aot-nmin2-pmin005-filled-long-deep-20260623T093120Z n=7, p-min=0.05 384/384 87.376 79.641 Valid but lower than p-min 0.10.
3 gemma4-q8-gpu3-mtp-n7-aot-nmin2-pmin010-psplit020-filled-long-deep-20260623T093120Z n=7, p-min=0.10, p-split=0.20 384/384 88.313 80.657 Valid, close to record but slightly below after-TTFT.

Takeaways

Follow-Up

The next sweep tests:

The alternate Q8_0 GGUF finished downloading at /mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf/gemma-4-26B-A4B-it-Q8_0.gguf (26,859,859,744 bytes). Test it only after the current Q8_K_XL MTP frontier is no longer moving, because Q8_K_XL is already producing validated records.