Qwen3.6-35B-A3B · vLLM-ultimate (AEON) · NVFP4 + DFlash · conc 4

status: done

Configuration

Modelnvidia/Qwen3.6-35B-A3B-NVFP4
CompanyAlibaba
FamilyQwen
Parameters35B / 3B (MoE, hybrid GDN+full-attn) + DFlash external drafter
EnginevLLM (aeon-vllm-ultimate custom container, v0.23.0+aeon.sm121a.dflash) + DFlash (z-lab/Qwen3.6-35B-A3B-DFlash @31977fbe small-page rev, num_speculative_tokens 11) (speculative decoding)
Quant / precisionNVFP4 (modelopt_mixed — W4A16_NVFP4 experts + FP8 GDN gates)
Why this quantconc-4 fine-grained point of the DFlash money-chart line, tracing the shallow-loss plateau between the 1→2 crossover and the conc-16/32 collapse. REVISED 2026-07-02 — re-run at ctx 65536 (was 40960) as part of the full ctx-matched six-point re-sweep. Same one-boot protocol (official checkpoint, small-page drafter, max-num-seqs 64). SAFETY — untrusted third-party image; NO credentials, weights + drafter READ-ONLY, port loopback-only.
Downloadnvidia/Qwen3.6-35B-A3B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s193.44
Decode tok/s210.73
Peak memory (GB)111.3 (system MemAvailable delta (idle baseline 118 GiB → 6.7 GiB available at load) over the one-boot conc-1/2/4/8/16/32 sweep — vLLM static KV (util 0.85) + DFlash drafter)
Completed2026-07-02 07:21 +0800

Full run command

# UNTRUSTED image — NO creds; weights + drafter READ-ONLY; loopback port. ONE boot (ctx 65536,
# max-num-seqs 64, small-page drafter @31977fbe) sweeping client conc 1/2/4/8/16/32 — see the main
# (c1) page for the full docker run command. Only the client --concurrency differs.
python3 scripts/bench-serving.py --base-url http://127.0.0.1:8000 --model official \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json --concurrency 4 --num-prompts 1000 --max-seconds 600 --max-tokens 256
# 499/1000 prompts (hit 600 s cap), 0 errors.

qwen3.6-35b-a3b Alibaba Qwen NVFP4 16-40B Spark recipe conc-4

Notes

conc-4 point of the Qwen3.6-35B-A3B NVFP4 + DFlash line — REVISED 2026-07-02, now ctx-matched to MTP/base (65536). Official nvidia/Qwen3.6-35B-A3B-NVFP4 on the AEON image, DFlash n=11 via the small-page drafter, one-boot sweep. 0 errors.

  • Result (conc 4): prefill 193.44 / decode 210.73 tok/s aggregate; 499/1000 prompts (hit the 600 s cap), 0 errors. vs MTP conc-4 (232.43): −9.3%.
  • The shallow-loss plateau, slightly steeper once ctx-matched. Across the ctx-matched sweep DFlash sits ~−6 to −12% behind MTP through conc 2/4/8 (−6.1% / −9.3% / −12.0%) — a modest but now monotonically widening loss, not the flatter ~−7-8% plateau the old ctx-40960 numbers showed — before the batch-fill collapse at conc-16 (−21.5%) and conc-32 (−24.8%). So the DFlash penalty is closer to smooth-then-cliff than flat-then-cliff once context is controlled.
  • Why — wasted draft compute. DFlash drafts 11 at ~26.5% acceptance / accept-len ~3.92 (~7 wasted forward passes/step); MTP drafts 3 at ~67% / accept-len 3.0. On this compute-bound MoE the extra draft work already dominates by conc-2 and worsens as streams divide the GPU. Acceptance flat vs conc (workload-driven) — matches c1 (26.8%) and c2 (26.7%) within noise.
  • One server lifetime for the whole six-point sweep → mem is the single ~111.3 GB reservation. TPOT 0.0 = qwen3 reasoning-parser client artifact.
  • Series: c1 (main) · c2 · c8 · c16 · c32. Matched MTP: -mtp-c4.