b70-optimization-lab

20260623T0900 Parallel / Cache Follow-Ups

Goal

Start from the validated optimized-SYCL + FLASH_ATTN=off base and test single-session server settings:

Results

Label Change Canary p512/o512 tok/s after TTFT Wall tok/s Decision
gemma4-q8-gpu0-syclopt0-faoff-parallel1-20260623T0800 --parallel 1, default threads 128/128 41.9229 35.8948 smoke win; default deep run invalidated by script edit mid-run
gemma4-q8-gpu1-syclopt0-faoff-parallel1-poll100-20260623T0815 --parallel 1, POLL=100 128/128 41.8095 35.4492 no win
gemma4-q8-gpu2-syclopt0-faoff-parallel1-t4-20260623T0815 --parallel 1, THREADS=4 128/128 41.9974 36.0963 smoke win over prior approved
gemma4-q8-gpu3-syclopt0-faoff-parallel1-t16-20260623T0815 --parallel 1, THREADS=16 128/128 42.0394 35.3365 deep-validated below
gemma4-q8-gpu3-syclopt0-faoff-parallel1-t16-deep-20260623T0830 --parallel 1, THREADS=16 384/384 41.9213 36.1123 validated, but superseded by cache0
gemma4-q8-gpu2-syclopt0-faoff-parallel1-cache0-20260623T0900 --parallel 1 --cache-ram 0, THREADS=16 128/128 42.0300 35.6997 smoke win
gemma4-q8-gpu2-syclopt0-faoff-parallel1-cache0-deep-20260623T0915 --parallel 1 --cache-ram 0, THREADS=16 384/384 42.1539 36.4063 current promoted best
gemma4-q8-gpu0-syclopt0-faoff-parallel1-cache0-poll100-20260623T0915 cache0 + POLL=100 128/128 41.9924 36.1634 no win
gemma4-q8-gpu1-syclopt0-faoff-parallel1-cache0-poll25-20260623T0915 cache0 + POLL=25 128/128 41.8549 35.7819 no win
gemma4-q8-gpu3-syclopt0-faoff-parallel1-cache0-ub128-20260623T0915 cache0 + UBATCH_SIZE=128 128/128 41.9971 35.8185 no win

32K Viability

gemma4-q8-gpu1-syclopt0-faoff-parallel1-ctx32768-smoke-20260623T0845 loaded with Q8 weights, f16 KV, CTX_SIZE=32768, and --parallel 1, then passed 64/64 chat canary rows. The short p512/o512 smoke measured 41.9509 tok/s after TTFT, but it is not a record because the canary depth was intentionally lighter and the purpose was context viability.

Decision

Promote gemma4-q8-gpu2-syclopt0-faoff-parallel1-cache0-deep-20260623T0915 as the new Q8 B70 single-GPU best and submit a LocalMaxxing update.