Muse Glimmer-30B NVFP4 (r0b0tlab)

NVFP4 (W4A4, block-16) quantization of meta-models/Muse-Glimmer-30B, built with NVIDIA ModelOpt 0.45.0 and qualified on a single NVIDIA DGX Spark (GB10 / SM121, 128 GB unified memory) with native FP4 tensor-core execution.

	BF16	NVFP4 (this repo)
Weights size	59.6 GB	25.4 GB (2.34x)
Single-stream decode (GB10)	~4.2 tok/s	~10.3 tok/s (2.5x)
Aggregate @ c16 (GB10)	—	52.5 tok/s
Median TPOT (GB10)	~245 ms	~95 ms
Max context (live NIAH)	—	131,072 (3/3 depths incl. 117,734)

Quantization details

Method: ModelOpt NVFP4_DEFAULT_CFG (dynamic input quantizer, block-16), *gate_proj excluded (classic NVFP4 quality protection).
Calibration: cnn_dailymail 3.0.0 train, 512 samples x 2,048 tokens, batch 1 (the ModelOpt canonical convention). Raw fetch SHA and windowing are recorded in modelopt_provenance.json.
Protected (stored BF16): vision tower, adapter, embeddings, lm_head, norms.
364 quantized weights, 1,072 protected tensors; structural audits recorded in quantization.complete.json.

Serving

Requires unmerged vLLM Muse support: PR vllm-project/vllm#51655 (pinned head 1f7f0715848c9acc56ea40faa21c13a02bdc8357). Build the image with --build-arg torch_cuda_arch_list=12.0 — arch-accelerated 12.0a/12.1a cubins do not run on SM121.

vllm serve /path/to/this/repo \
  --served-model-name muse-glimmer-30b-nvfp4 \
  --dtype bfloat16 --max-model-len 131072 \
  --gpu-memory-utilization 0.70 --max-num-seqs 8 \
  --enforce-eager \
  --attention-backend FLASHINFER --mm-encoder-attn-backend FLASHINFER \
  --kv-cache-dtype fp8 \
  --quantization modelopt_fp4 --linear-backend flashinfer_cutlass \
  --reasoning-parser muse_glimmer --tool-call-parser muse_glimmer \
  --enable-auto-tool-choice --limit-mm-per-prompt '{"image":1}'

Full recipe and verification notes: r0b0tlab/muse-glimmer-30b-nvfp4-vllm.

Evaluation (r0b0bench core-subset, official scorers, reasoning_strength=low)

Lane	Result
GSM8K-200	91.0%
ARC-Easy-400	95.75%
IFEval-200	82.0%
HumanEval-164	85.4% pass@1
BFCL-MT-200 (tool calling)	52.0%
NIAH @131K (3 depths)	3/3 retrieved

Caveats (read before citing)

The base model is nondeterministic at temperature 0 (near-tie greedy tokens; verified empirically on repeated identical prompts). Matched-case output comparisons between dtypes are therefore noise-dominated — quality is reported on official scorers, not exact-string deltas.
Muse emits channel-scoped reasoning before the answer channel: short max_tokens budgets (<= ~64) can leave the answer channel empty.
Speculative decoding (DFlash, 3B assistant): correctness-verified and ~2.7x on BF16, but losslessness is unprovable given the temp-0 nondeterminism above. NVFP4+DFlash requires upstream quant-routing fixes not yet merged.

License

Base model: meta-models/Muse-Glimmer-30B under Apache 2.0 (see LICENSE; see USAGE_POLICY.md). This quantization adds no new license terms.

r0b0tlab/Muse-Glimmer-30B-NVFP4