gpt-oss-20b · vLLM · MXFP4 · conc 16

status: done

Configuration

Modelopenai/gpt-oss-20b
CompanyOpenAI
Familygpt-oss
Parameters21B / 3.6B (MoE)
EnginevLLM
Quant / precisionMXFP4
Why this quantconc-16 base (non-spec) point of gpt-oss-20b — matched no-spec baseline for the EAGLE3 conc-16 row (EXPERIMENTS.md
Downloadopenai/gpt-oss-20b
Context window65536
Input modalitiestext

Measured results

Prefill tok/s403.31
Decode tok/s340.69
Peak memory (GB)107.83 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-07-01 15:35 +0800

Full run command

# conc-16 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 16 1000 600 256
# 837/1000 prompts (hit 600 s cap), 89 harmony errors.

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

Notes

conc-16 base (no-spec) point of gpt-oss-20b MXFP4 — completes the matched base line for the EAGLE3 artifact test (EXPERIMENTS.md #15). Same cu130-nightly recipe as the published conc-32 base.

  • Result (conc 16): prefill 403.31 / decode 340.69 tok/s aggregate; 837/1000 prompts (hit the 600 s cap), 89 harmony errors (error count scales cleanly with batch: 5→22→89 at c2→c4→c16, all mid-reasoning 256-tok truncation — a vLLM harmony-finalizer robustness cost, not a config fault); peak mem 107.8 GB.
  • Base line for the artifact test is now: decode 83.4 (c2) → 127.3 (c4) → 340.7 (c16) → 535.3 (c32). Pair each against the EAGLE3 row to see whether spec-decode helps at all before conc-32.
  • TTFT/TPOT are buffered-reasoning artifacts — aggregate decode tok/s is the valid metric.
  • Base siblings: -c2 · -c4 · c32 (main). EAGLE3 counterpart: -eagle3-c16.