chore: retest and rebench on RTX 5070 - #9

Merged

mudler merged 10 commits into

mudler:mainfrom

richiejp:bench-lever-sampled-token-20260727

Jul 30, 2026

Commits on Jul 29, 2026

bench(qwen35): revalidate 4B on current main, and VOID the prior series as contended
richiejp
committed
4357db9
feat(engine): ENG-ASYNC-SCHED W4 discrete device-resident sampled tokens (opt-in), and the attribution that motivated it is wrong
richiejp
committed
deed7c2
fix(build): make the sanitizer lanes actually build, and repair the CPU break they found
richiejp
committed
58e106b
fix(bench): cool down before sampling GPU idleness in the A/B harness
richiejp
committed
9dd3db3
bench(w4): A/B says NEUTRAL, and the profile says the syncs were never the async sampler's
richiejp
committed
f48973f
fix(runner): ask the backend whether memory is unified, not whether the device is CUDA
richiejp
committed
691af16
docs(bench): record the cutlass_80-on-sm_120 observation as an open lead
richiejp
committed
7041665
bench(harness): let the vLLM arm use a venv's own CUDA toolkit
richiejp
committed
add1680
docs(bench): re-baseline the 4B ratio against an oracle at the parity pin
richiejp
committed
629c9aa
bench(qwen35): re-validate the 4B lever across a 139-commit rebase — nothing moved
richiejp
committed
bfd51fb