gpt-oss-20b · vLLM · MXFP4 · conc 1

status: done

Configuration

Model	`openai/gpt-oss-20b`
Company	OpenAI
Family	gpt-oss
Parameters	21B / 3.6B (MoE)
Engine	vLLM
Quant / precision	MXFP4
Why this quant	conc-1 base (non-spec) point of gpt-oss-20b — completes the matched no-spec baseline for the EAGLE3 conc-1 row alongside the existing conc-2/4/16 base points. Same cu130-nightly recipe as the published conc-32 base; only --max-num-seqs differs.
Download	openai/gpt-oss-20b
Context window	65536
Input modalities	text

Measured results

Prefill tok/s	64.57
Decode tok/s	45.56
Peak memory (GB)	108.06 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed	2026-07-02 07:51 +0800

Full run command

# conc-1 base (no spec), cu130-nightly to match the gpt-oss-20b series. 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 1 1000 600 256
# 115/1000 prompts (hit 600 s cap), 2 harmony errors. ready after 151 s.

gpt-oss-20b OpenAI gpt-oss MXFP4 16-40B Spark recipe conc-1

Notes

conc-1 base (no-spec) point of gpt-oss-20b MXFP4 — completes the base line down to single-stream, pairing with the EAGLE3 conc-1 row for the low-batch loss comparison.

Result (conc 1): prefill 64.57 / decode 45.56 tok/s single-stream aggregate; 115/1000 prompts (hit the 600 s cap), 2 harmony errors (fewest in the series — errors scale with batch: 2→5→22→35→89→108 at c1→c2→c4→c8→c16→c32); peak mem 108.06 GB; TTFT median 5507.8 ms.
Full base line now complete c1→c32: decode 45.56 (c1) → 83.4 (c2) → 127.3 (c4) → 212.5 (c8) → 340.7 (c16) → 535.3 (c32) — clean monotone scaling, no dip.
The EAGLE3-vs-base “inverted” pathology now has a gap-free curve. Against this base, EAGLE3 decode (-eagle3-c1 at 38.55) is −15.4% at conc-1 — a real loss even single-stream, not the neutral/break-even point one might expect from a bandwidth-bound regime. Full curve: −15.4% (c1) → −29.3% (c2) → −32.9% (c4) → −40.6% (c8) → +26.8% (c16) → +28.2% (c32) — EAGLE3 is a loss at every batch size up to c8, and only turns positive at c16/c32, tracking the acceptance jump documented in notes/INCOMPATIBILITIES.md (~5% accept at conc≤8 vs ~44% at conc≥16).
TTFT/TPOT are buffered-reasoning artifacts — aggregate decode tok/s is the valid metric.
Base siblings: -c2 · -c4 · -c8 · -c16 · c32 (main). EAGLE3 counterpart: -eagle3-c1.