gpt-oss-20b · vLLM · MXFP4 + EAGLE3 · conc 2

status: done

Configuration

Modelopenai/gpt-oss-20b
CompanyOpenAI
Familygpt-oss
Parameters21B / 3.6B (MoE)
EnginevLLM + EAGLE3 (speculative decoding)
Quant / precisionMXFP4
Why this quantconc-2 point of the EAGLE3 fine-grained sweep (EXPERIMENTS.md
Downloadopenai/gpt-oss-20b
Context window65536
Input modalitiestext

Measured results

Prefill tok/s67.51
Decode tok/s58.98
Peak memory (GB)108.43 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + EAGLE3 head)
Completed2026-07-01 16:02 +0800

Full run command

# conc-2 EAGLE3, cu130-nightly + RedHatAI speculator. 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 2 1000 600 256 \
  --speculative-config '{"model":"RedHatAI/gpt-oss-20b-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
# 141/1000 prompts (hit 600 s cap), 11 harmony errors. ready after 195 s.

gpt-oss-20b OpenAI gpt-oss MXFP4 16-40B Spark recipe conc-2

Notes

conc-2 EAGLE3 — spec-decode is a NET LOSS here, confirming the “+28%@c32 is an artifact” thesis. gpt-oss-20b MXFP4 + RedHatAI EAGLE3 speculator on vLLM, conc 2.

  • Result (conc 2): prefill 67.51 / decode 58.98 tok/s aggregate; 141/1000 prompts (hit the 600 s cap), 11 harmony errors; peak mem 108.4 GB.
  • EAGLE3 vs base at conc-2: −29.3% (58.98 vs base 83.37). The draft doesn’t just fail to help — it actively hurts at low batch, because the wasted draft/verify work isn’t offset by any acceptance. This is the decisive evidence for #15: if the conc-32 “+28%” came from acceptance, spec would help here too; instead it’s strongly negative, so the c32 win is a scheduling/prefill effect at that one batch size, not a real draft win.
  • Acceptance is dismal (as at c1/c8): mean accept-len ~1.1–1.45 (≈1 extra token per 3 drafted), avg draft acceptance ~3–15% (centered ~6%), per-position collapsing to near-zero by slot 3. Far below EAGLE3’s ~3.0 / ~70% expectation — ShareGPT general chat + gpt-oss’s harmony reasoning channel is off-distribution for this draft (see notes/INCOMPATIBILITIES.md).
  • TTFT/TPOT are buffered-reasoning artifacts — aggregate decode tok/s is the valid metric.
  • Sweep: -c1 · -c4 · -c8 · -c16 · c32 (main). Base: -c2.