gpt-oss-20b · vLLM · MXFP4 · conc 4

status: done

Configuration

Modelopenai/gpt-oss-20b
CompanyOpenAI
Familygpt-oss
Parameters21B / 3.6B (MoE)
EnginevLLM
Quant / precisionMXFP4
Why this quantconc-4 base (non-spec) point of gpt-oss-20b — matched no-spec baseline for the EAGLE3 conc-4 row (EXPERIMENTS.md
Downloadopenai/gpt-oss-20b
Context window65536
Input modalitiestext

Measured results

Prefill tok/s179.74
Decode tok/s127.26
Peak memory (GB)108.4 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-07-01 15:22 +0800

Full run command

# conc-4 base (no spec), cu130-nightly. Harmony vocab via VOCAB_DIR override.
VLLM_IMAGE=vllm/vllm-openai:cu130-nightly VOCAB_DIR=$HOME/tiktoken_encodings \
  scripts/bench-vllm-serving.sh openai/gpt-oss-20b 65536 4 1000 600 256
# 313/1000 prompts (hit 600 s cap), 22 harmony errors.

gpt-oss-20b OpenAI gpt-oss MXFP4 16-40B Spark recipe conc-4

Notes

conc-4 base (no-spec) point of gpt-oss-20b MXFP4 — matched baseline for the EAGLE3 conc-4 row (EXPERIMENTS.md #15). Same cu130-nightly recipe as the published conc-32 base.

  • Result (conc 4): prefill 179.74 / decode 127.26 tok/s aggregate; 313/1000 prompts (hit the 600 s cap), 22 harmony errors (rising with batch — 5 at c2 → 22 at c4, mid-reasoning 256-tok truncation); peak mem 108.4 GB.
  • Anchors the base line for the EAGLE3 artifact test at conc-4. Base line so far: 83.4 (c2) → 127.3 (c4).
  • TTFT/TPOT are buffered-reasoning artifacts — aggregate decode tok/s is the valid metric.
  • Base siblings: -c2 · -c16 · c32 (main). EAGLE3 counterpart: -eagle3-c4.