gpt-oss-20b · vLLM · MXFP4 · conc 2

# conc-2 base (no spec), cu130-nightly to match the gpt-oss-20b series. Harmony vocab via VOCAB_DIR override
# (the historical ~/models/tiktoken_cache was empty+root-owned — see notes/INCOMPATIBILITIES.md).
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
# 209/1000 prompts (hit 600 s cap), 5 harmony errors. TTFT/TPOT are buffered-reasoning artifacts.

conc-2 base (no-spec) point of gpt-oss-20b MXFP4 — the matched baseline the EAGLE3 artifact test needs (EXPERIMENTS.md #15). Same cu130-nightly recipe as the published conc-32 base; only --max-num-seqs changes.

• Result (conc 2): prefill 128.27 / decode 83.37 tok/s aggregate; 209/1000 prompts (hit the 600 s cap), 5 harmony errors (mid-reasoning truncation at 256 tok — same class as the conc-32 base’s 108 errors, fewer at low batch); peak mem 108.5 GB.
• Purpose: anchors the base line so the EAGLE3 conc-2 point (pending) can be expressed as a real decode-speedup. The post’s claim is that EAGLE3’s conc-32 “+28%” is a scheduling/prefill effect, not acceptance — if so, EAGLE3-vs-base should be flat or negative at conc 2/4/16 and only spike at conc-32. This base row is half of that measurement at conc-2.
• TTFT/TPOT are buffered-reasoning chat-path artifacts (vLLM emits the harmony reasoning channel in one burst) — the aggregate decode tok/s is the valid metric.
• Base siblings: -c4 · -c16 · c32 (main). EAGLE3 counterpart: -eagle3-c2.