gpt-oss-120b · vLLM · MXFP4 + EAGLE3 (LMSYS draft) · conc 8

# vLLM + LMSYS draft, conc-8. cu130-nightly, VOCAB_DIR override.
VLLM_IMAGE=vllm/vllm-openai:cu130-nightly VOCAB_DIR=$HOME/tiktoken_encodings \
  scripts/bench-vllm-serving.sh openai/gpt-oss-120b 65536 8 1000 900 256 \
  --speculative-config '{"model":"lmsys/EAGLE3-gpt-oss-120b-bf16","method":"eagle3","num_speculative_tokens":3}'
# 628/1000 prompts (hit 900 s cap), 32 harmony errors. ready after 486 s.

conc-8 — the LMSYS draft converts BEST at moderate batch (acceptance ~48%, near SGLang’s ~55%). gpt-oss-120b MXFP4 + lmsys/EAGLE3-gpt-oss-120b-bf16 on vLLM.

• Result (conc 8): prefill 201.76 / decode 175.31 tok/s aggregate; 628/1000 prompts (hit the 900 s cap), 32 harmony errors; peak mem 108.4 GB.
• Acceptance peaks here, ~48% (mean accept-len ~2.45) — higher than both conc-1 (~20%) and conc-32 (~29%), and close to the same draft’s ~55% on SGLang. So the vLLM+LMSYS acceptance curve is peak-in-the-middle: ~20% (c1) → ~48% (c8) → ~29% (c32), not the monotone rise the gpt-oss-20b showed (~5% ≤c8 → ~44% ≥c16).
• What generalizes vs what doesn’t. Common to both models: acceptance is depressed at the lowest concurrency (120b c1 ~20%, 20b c2 ~5%) — the low-batch EAGLE3 pathology. What differs is the high end: the 20b stays high at c16/c32, the 120b declines by c32 (the classic “draft loses value as the batch saturates”). So the robust, cross-model claim is the low-batch depression; the high-concurrency shape is model-dependent, and shouldn’t be stated as a universal “acceptance rises with concurrency” rule.
• No matched 120b base at conc-8 exists (base is conc-32 only), so this row is characterized by acceptance
  • absolute decode, not a spec-vs-base delta. The decisive draft-vs-engine delta is on the -lmsys-c32 page.
• TTFT/TPOT are buffered-reasoning artifacts — aggregate decode tok/s is the valid metric.
• Sweep: -lmsys-c1 · -lmsys-c32 · SGLang+LMSYS sglang-eagle3-c1.