b70-optimization-lab

Gemma 4 26B A4B Q8 B70 Experiments

Active experiment lane for one Gemma 4 26B A4B Q8-quality replica per Intel B70.

The promoted/reference packet is ../../results/gemma4-26b-a4b-q8-b70/. Keep raw attempts, failed knobs, and patch notes here until a result is worth promoting.

Before launching new sweeps, read the current research plan and model/runtime options.

Layout

Suggested subfolders:

experiments/gemma4-26b-a4b-q8-b70/
  logs/          # short copied logs or checksums of external logs
  patches/       # source/config patches tested in this lane
  sweeps/        # small markdown/json summaries of parameter sweeps

Large logs should live under /mnt/fast-ai/bench-results/gemma4-26b-a4b-q8/ with paths and summaries recorded here or in ../../data/.

First Sweep Matrix

After the first replica serves:

GPU	Purpose	First knobs
0	Control	-fa 1, f16 KV, default graph/DNN
1	ubatch	-ub 64/128/256/512
2	graph/DNN	GGML_SYCL_DISABLE_GRAPH=0/1, GGML_SYCL_DISABLE_DNN=0/1
3	alternative runtime	vLLM int8-per-channel or candidate patch

Every sweep entry should state quality status, not only speed.

This site is open source. Improve this page.