jaytonde05/kv-cache-quantization-commands

KV cache quantization: run commands

These are the scripts and commands for the post "KV Cache Quantization, Measured: What FP8 and NVFP4 Really Cost". This repo holds commands only. It does not hold model weights, benchmark data, or model responses.

Setup

Item	Value
Models	Qwen/Qwen3.8-27B, google/gemma-4-31B-it (weights in BF16)
GPU	NVIDIA H200 141 GB, one GPU per config. A 2-GPU VM ran two configs at the same time.
vLLM	0.32.0rc1.dev30+g3ad096faf (scripts/bench/setup_vllm.sh)
Python packages	requirements-freeze.txt
Terminal-Bench	Terminal-Bench 2.1 with Harbor and the terminus-2 agent

Configs

Name	--kv-cache-dtype	--attention-backend
BF16	bfloat16	FLASH_ATTN
BF16-FI	bfloat16	FLASHINFER
FP8	fp8_e4m3	FLASH_ATTN
FP4 (NVFP4)	nvfp4	FLASHINFER
BF16-TFA	bfloat16	TRITON_FLASH_ATTN

For RULER, DocVQA and MRCR, Gemma FP8 used TRITON_FLASH_ATTN, and Gemma BF16-FI and FP4 used --language-model-only (text only), because FlashInfer has no Gemma image attention on H100 or H200. All Gemma Terminal-Bench runs used --language-model-only and the backend in the table above.

Serve command used by scripts/bench/exp2.sh:

vllm serve $MODEL --served-model-name $NAME --port $PORT --dtype bfloat16 \
  --max-model-len $MAXLEN --gpu-memory-utilization 0.90 --reasoning-parser $RP --seed 0 \
  --kv-cache-dtype $KV --attention-backend $BE

Order of work

1. Install and download: setup_vllm.sh, download_model.sh.
2. Prepare data: prep_data.py, make_6t.py, mrcr.py prep, prep_gemma.sh.
3. RULER, DocVQA, MRCR: exp2.sh, driven by exp2_gpu0.sh, exp2_gpu1.sh, after_tb.sh, gemma_rest.sh. Score with score_exp2.py and mrcr.py.
4. Terminal-Bench: tb_queue.sh, tb_server.sh, tb_rerun_timeouts.sh, gemma_tb.sh, gemma_env.sh. Score with score_tb.py and score_nolim.py.
5. Nightly backend check for Gemma images, 9 Nov 2026: scripts/nightly_check/.

COMMANDS.md has every vllm serve and vllm bench serve command written out in full.

top_level_commands.txt has the launch commands from the VM shell history. exp0* scripts, kl_probe.py, and the smoke scripts were exploratory, and the post does not use their results.

Run times

Time is the wall clock for the whole benchmark, including interruptions. It is not a speed test. Each config used the same concurrency, but response lengths differ.

Config	RULER 4K + 32K + 128K	DocVQA	MRCR	Terminal-Bench k=3
Qwen BF16	55 min	7.2 min	78 min	37.3 h
Qwen BF16-FI	62 min	7.3 min	89 min	39.9 h
Qwen FP8	62 min	7.3 min	90 min	24.0 h
Qwen NVFP4	93 min	7.5 min	131 min	26.1 h
Gemma BF16	207 min	10.9 min	415 min	34.5 h
Gemma BF16-FI	190 min	not supported	252 min	29.2 h
Gemma FP8	181 min	11.7 min	337 min	28.4 h
Gemma NVFP4	316 min	not supported	443 min	27.1 h

Qwen no-limit reruns of timed-out Terminal-Bench tries took about 42 to 45 hours per config. Gemma reruns are not finished, so Gemma Terminal-Bench scores are raw.