peterhdd/beam-rtx-pro-6000-runs

Beam-1.0 on RTX PRO 6000: every benchmark run

Raw outputs and scripts behind the write-up Beam on RTX PRO 6000. Three models, measured on 8× NVIDIA RTX PRO 6000 Blackwell (96 GB) with vllm/vllm-openai:reflection-beam:

Folder	Model	GPUs
results/official-nvfp4	reflection/Beam-1.0-501B-A23B-NVFP4	4
results/official-nvfp4-dflash2	same + DFlash2 (speed only)	4
results/reap-bf16	mrfakename/Beam-1.0-REAP-255B-A23B	8
results/reap-nvfp4	peterhdd/Beam-1.0-REAP-255B-A23B-NVFP4A16	2

Summary

	MMLU-Pro (420)	IFEval prompt strict (541)	HumanEval (164)	1 user, real prompts	64 users
Official NVFP4	0.843 ± 0.017	0.939 ± 0.010	158	100.5 tok/s (167.7 with DFlash2)	1,173.8 tok/s
REAP BF16	0.721 ± 0.021	0.913 ± 0.012	156	–	–
REAP NVFP4A16	0.702 ± 0.022	0.896 ± 0.013	156	81.9 tok/s	988.8 tok/s

All numbers in machine-readable form: summary.json.

What's in each model folder

• mmlu_pro/, ifeval/: lm-eval results and per-question samples (--log_samples), incl. full answers. mmlu_pro.txt, ifeval.txt: console tables.
• humaneval_chat/: samples.jsonl (every answer, extracted code, pass/fail) and summary.json.
• speed_c{1,16,64}.txt: vllm bench serve output (1024 in / 512 out, random prompts).
• real_speed.json: 20 real prompts sent one at a time.
• kernels.txt, gpu_memory.txt: server log excerpts (backends chosen) and nvidia-smi memory.
• reap-bf16/humaneval/: the broken lm-eval humaneval_instruct run (all empty answers: the task's answer prefill doesn't work with Beam's thinking template). Kept as evidence; use humaneval_chat instead.

Settings

temperature 0.7, top_p 0.99 (Reflection's eval settings), max 32,000 generated tokens incl. thinking, seed 42, one run each. MMLU-Pro: mmlu_pro_llama, --limit 30 per subject. Server: --max-model-len 65536 --kv-cache-dtype fp8 --moe-backend marlin --reasoning-parser beam --tool-call-parser beam.

Scripts (scripts/)

Script	What it does
setup.sh, download.sh	VM setup, pull image + weights
serve.sh	start vLLM in the Beam Docker image on chosen GPUs/port
quick_eval.sh	MMLU-Pro (420) + IFEval via lm-eval against the server
humaneval_chat.py	HumanEval through the chat API, runs tests in a sandboxed subprocess
speed.sh, real_speed.py	throughput at 1/16/64 users; single-user speed on real prompts
quantize_reap.py	REAP BF16 → NVFP4A16 with llm-compressor model_free_ptq
summarize.py, paired_test.py	rebuild summary.json; exact McNemar tests between models

Limits

One sampled run per benchmark; MMLU-Pro is a subset. Good for comparing these models with each other, not for leaderboard claims. Not tested: long context, tool calling at scale, non-English, DFlash2 under load.