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:
dec5ca557, SYCL/Level Zero, AOT BMG build
/home/steve/src/llama.cpp/build-sycl-b70-aot-bmg-g31/bin/llama-server;gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf;mtp-gemma-4-26B-A4B-it.gguf;BENCH_PROMPT_MODE=filled-long, GGML_SYCL_DISABLE_OPT=0,
FLASH_ATTN=off, POLL=50, --parallel 1 --cache-ram 0, f16 main and
draft KV;384/384 chat canary before promotion;588 prompt tokens and 512 output tokens.| 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:
cmqqg1r0l015xqo01e6d696mx;data/localmaxxing-gemma4-26b-a4b-q8-b70-mtp-n7-aot-nmin2-pmin010-filledlong512-20260623.queue.json;data/localmaxxing-responses/gemma4-26b-a4b-q8-b70-mtp-n7-aot-nmin2-pmin010-filledlong512-20260623.submit.log;cmqqfv296015sqo0126mym3ko (87.878 tok/s, n=7 p-min 0.15).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. |
p-min=0.10 is the current best confidence gate for n=7.0.05 hurt throughput, and raising n to 8 was a clear
loss despite high acceptance length. The n=8 final stats showed more draft
generation cost than the extra accepted token could pay back.p-split=0.20 nearly tied the record and slightly improved wall throughput,
but did not beat after-TTFT throughput. It is still worth testing
p-split=0.05/0.15 around p-min=0.10.models.json. They were killed and kept as negative artifacts.
Do not over-interpret those as correctness failures; they look like transient
launch/resource hangs.The next sweep tests:
n=7, p-min=0.10;n=7, p-min=0.10, p-split=0.15;n=7, p-min=0.10, p-split=0.05;n=7, p-min=0.10, --no-spec-draft-backend-sampling.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.