GUIDELLM__MP_CONTEXT_TYPE=forkserver
guidellm benchmark run --target http://localhost:8003 --model Qwen3-4B --processor /home/ckl/projects/arle/infer/models/Qwen3-4B --profile concurrent --data prompt_tokens=512\,prompt_tokens_min=512\,prompt_tokens_max=512\,output_tokens=2048\,output_tokens_min=2048\,output_tokens_max=2048 --max-seconds 180 --random-seed 20260416 --output-dir /home/ckl/projects/arle/bench-output/2026-05-09-baseline-B3-bf16-c4-decode-r1 --backend openai_http --backend-kwargs \{\"validate_backend\":\ \"/v1/models\"\,\ \"request_format\":\ \"/v1/completions\"\} --disable-console-interactive --outputs json --outputs csv --outputs html --rate 4 --warmup 10
