Qwen3.8-27B NVFP4 + MTP (DGX Spark / SM121 build)

Self-quantized Qwen3.8-27B checkpoint for NVIDIA GB10 (DGX Spark, SM121), produced with NVIDIA ModelOpt 0.46.0rc1 using the shipped qwen3_5 family recipe verbatim (w4a16_nvfp4-fp8_attn-kv_fp8_cast), plus the trained MTP head merged back in BF16 for speculative decoding.

Quantization map (census-verified, matches the official family control)

• 193 × W4A16_NVFP4 — all MLP gate/up/down projections + lm_head (block-16, e4m3 scales; weights-only, activations BF16)
• 208 × FP8 — self-attn q/k/v/o + GDN linear-attn projections (static input scales exported)
• 257 × BF16 — remaining tensors (conv1d, gates/routers, small GDN paths)
• 15 × BF16 MTP tensors — trained multi-token-prediction head (shard 4, sha256 47202b11…, mirrors the official NVFP4-family checkpoint contract)

Calibration (NVIDIA-documented path)

get_dataset_dataloader with the registered combo cnn_dailymail + nvidia/Nemotron-Post-Training-Dataset-v2 (stem/chat/math/code splits), 2048 packed rows @ 1024 tokens, left padding, pack=True, max algorithm.

IMPORTANT: FP8 KV flag

This checkpoint ships kv_cache_quant_algo: "FP8" (cast mode, default scales). Serve with --kv-cache-dtype fp8. Do not strip this flag: on vLLM 0.27.x with qwen3_5-family models, a checkpoint without the flag that is served with the runtime fp8 option takes a broken generic path and produces a deterministic arithmetic defect (e.g. 19×23 answered "417"). With the flag, KV routes through ModelOptKVCacheMethod and is exact. Root-caused during this campaign; upstream-report candidate.

Verified results (single DGX Spark, vLLM v0.27.2rc0-sm121)

Files (complete 4-of-4 tree)

This repo must contain all four shards. A tree with only model-00004-of-00004.safetensors (the MTP head) is incomplete and will not reproduce the published numbers.

model-00001-of-00004.safetensors   4208cd3b…   ~9.3 GiB
model-00002-of-00004.safetensors   024111b9…   ~9.3 GiB
model-00003-of-00004.safetensors   927ee343…   ~1.1 GiB
model-00004-of-00004.safetensors   47202b11…   ~0.8 GiB  (BF16 MTP head)

Do not mix these shards with a sibling NVFP4 checkpoint.

Serve

Use the published SM121 image. Stock vllm-nightly typically hits [AutoTuner] No tuned config covers fp4_gemm ... tactic=-1 and lands around 20 tok/s MTP instead of 27.8.

docker run --gpus all -p 8000:8000 \
  -v /path/to/this/checkpoint:/model:ro \
  ghcr.io/r0b0tlab/qwen38-27b-nvfp4-sm121:v0.27.2rc0-sm121 \
  --model /model --served-model-name qwen38-27b \
  --max-model-len 32768 --gpu-memory-utilization 0.70 \
  --kv-cache-dtype fp8 --enforce-eager --no-enable-prefix-caching \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Canary: 19 × 23 must answer 437. 417 means the FP8-KV flag path is broken.

For 262K-context serving: --max-model-len 262144 --gpu-memory-utilization 0.85 --max-num-batched-tokens 8192 (KV capacity ≈ 2.5M tokens with fp8 KV).

DSpark (optional, vLLM only — not SGLang): adapt RadixArk/Qwen3.8-27B-DSpark with scripts/adapt_dspark_draft.py from the repro repo, then --speculative-config '{"method":"dspark","model":"/draft","num_speculative_tokens":7}'. A raw SpecForge draft deadlocks after FlashInfer autotune on this engine.

Provenance

• Source: Qwen3.8-27B (BF16), self-quantized per the NVIDIA ModelOpt shipped recipe; calibration via ModelOpt's documented dataset path.
• Full reproducibility pack (PTQ worker, MTP merge, all gates, raw evidence): https://github.com/r0b0tlab/qwen38-27b-nvfp4-sm121-vllm
• Runtime image: ghcr.io/r0b0tlab/qwen38-27b-nvfp4-sm121:v0.27.2rc0-sm121 (registry digest sha256:5bd3f329…b775f).

Weights follow the upstream Qwen license terms; this repository provides the quantized derivative for reproducibility. No warranties.