Qwen3.6-35B-A3B · SGLang · NVFP4 + MTP

status: blocked

Configuration

ModelQwen/Qwen3.6-35B-A3B
CompanyAlibaba
FamilyQwen
Parameters35B / 3B (MoE)
EngineSGLang + MTP (NEXTN) (speculative decoding)
Quant / precisionNVFP4
Why this quantNVIDIA's official NVFP4 (nvidia/Qwen3.6-35B-A3B-NVFP4, ModelOpt v0.44.0) + its in-repo MTP module via SGLang's NEXTN path. Preferred over unsloth per policy. NOTE — nvidia documents only a vLLM path, so SGLang support for the ModelOpt NVFP4 format must be verified at run time.
Downloadnvidia/Qwen3.6-35B-A3B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s
Decode tok/s
Peak memory (GB)
Completed2026-06-23 13:08 +08

Full run command

# BLOCKED — same GatedDeltaNet block-FP8 shape wall as the base sibling; load crashes before MTP/NEXTN
# is reached, so spec-decode is moot.
SGLANG_IMAGE=lmsysorg/sglang:nightly-dev-cu13-20260623-ba9d5aed \
  scripts/bench-sglang-serving.sh nvidia/Qwen3.6-35B-A3B-NVFP4 65536 32 1000 900 256 \
  --trust-remote-code --speculative-algo NEXTN --speculative-num-steps 3 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
#   ValueError: Weight output_partition_size = 32 is not divisible by weight quantization block_n = 128
#   (qwen3_5.py create_ba_proj -> fp8.py validate_block_quant_shapes). See the base sibling for the full trace.

qwen3.6-35b-a3b Alibaba Qwen NVFP4 16-40B conc-32

Notes

BLOCKED — same root cause as the base SGLang sibling. Load crashes inside SGLang’s qwen3_5 GatedDeltaNet (create_ba_proj, output_partition_size = 32 not divisible by FP8 block_n = 128) before MTP/NEXTN is ever set up, so the spec-decode path can’t be exercised. The {vLLM, SGLang} × {base, MTP} grid for this model is therefore {vLLM only}.

  • Failure: identical ValueError: Weight output_partition_size = 32 is not divisible by weight quantization block_n = 128 — full trace and analysis in the base config qwen3-6-35b-a3b-nvfp4-sglang and in notes/INCOMPATIBILITIES.md. No trusted alternative NVFP4 packing exists for the 35B-A3B.
  • Use the vLLM MTP path insteadqwen3-6-35b-a3b-nvfp4-vllm-mtp (done): decode 541 tok/s agg (conc-32), accept-len ~3.0, with conc-8 / conc-1 siblings also done. That is the supported NVFP4+MTP measurement for this model.