Add llama-bench-style benchmark: llama-tornado --bench (pp/tg/pg matrix, avg±stddev t/s, md/csv/json)#133

Open

mikepapadim wants to merge 5 commits into

beehive-lab:feat/mma_cudafrom

mikepapadim:feat/gpullama3-bench

Commits on Jul 14, 2026

Fix Qwen3 FP16 garbage output on CUDA: RMS-norm reduction had a cross-workgroup data race (workgroup 0 combined other groups' partials with no synchronization); NVIDIA path now uses a race-free sin…
mikepapadim
committed
Add llama-bench-style benchmark (llama-tornado --bench): pp/tg/pg test matrix over multiple models, warmup + repetitions, avg±stddev tok/s, markdown/CSV/JSON output; forward-pass-only timing (no to…
mikepapadim
committed
Bench: add remaining applicable llama-bench features - context depth (-d, untimed KV prefill, tests named ppN@dK/tgN@dK), jsonl + sql output, -oe second-format-to-stderr, --delay between tests
mikepapadim
committed
Bench: add -b prompt-processing batch size — pp tokens run through the batched-prefill tensor-core (MMA) path (compute-bound, ~47x pp on Llama-1B b128), mirroring llama-bench -b; generation stays s…
mikepapadim
committed
Bench launcher: when --bench-args supplies -m, don't also prepend --model (avoids the primary model appearing twice in the results)
mikepapadim
committed