Muse Glimmer — NVFP4/MXFP8 (vendor recipe)
Overview
The vendor-recipe quantized checkpoint of Muse Glimmer, packed for
serving with SGLang. This is the model vendor's own fp_awq mixed-precision
recipe (MIXED_PRECISION): NVFP4 (E2M1, group 16) on most projections,
MXFP8 (E4M3, group 32) on the higher-sensitivity set (v_proj,
down_proj, lm_head), norms in bf16 — 18.3 GiB versus 52 GiB for the bf16
export. Grab it and serve it directly by repo id; no conversion step needed.
Model Specifications
- Base model:
meta-models/Muse-Glimmer-30B(Muse Glimmer bf16 export) - Format: Safetensors, 5 shards, 18.34 GiB
- Recipe: vendor
fp_awqhand-off packed with the SGLang fork'sconvert_fp_awq_to_hf.py(default flags — the recipe as shipped); quantization map inhf_quant_config.json(quant_algo: MIXED_PRECISION) - Language model: 52 layers, hidden 6656, 32 Q / 2 KV heads, vocab 202,048
- Text-only serving is recommended (
--language-model-only); the vision tower ships in the base export, not in this quantized checkpoint
Evaluation Results
Measured on 1× DGX Spark (GB10), SGLang, 1024 in / 1024 out, greedy:
| BS | output tok/s | vs bf16 target | + DFlash draft (sim acc=5) |
|---|---|---|---|
| 1 | 12.1 | 2.7× | 36.4 |
| 4 | 47.7 | 2.7× | 156.2 |
| 8 | 92.2 | 2.6× | 300.7 |
Notes: decode is memory-bound, so the speedup tracks the 52→18 GiB weight reduction. On GB10, quantized prefill is slower than bf16 at batch ≥ 4 (~650 vs ~1500 input tok/s) — decode-heavy workloads win, prefill-heavy workloads should measure. Correctness smoke-verified (greedy arithmetic and generation with correct stop tokens); full accuracy suite pending.
Serving with SGLang
Requires the SGLang fork with Muse Glimmer support (sgl-project/sglang#34262, model support PR #3) until merged upstream.
sglang serve \
--model-path RadixArk/Muse-Glimmer-NVFP4 \
--reasoning-parser muse \
--tool-call-parser muse \
--language-model-only \
--tp-size 1 \
--mem-fraction-static 0.85 \
--host 0.0.0.0 --port 30000
With speculative decoding (pairs with meta-models/Muse-Glimmer-30B-assistant):
sglang serve \
--model-path RadixArk/Muse-Glimmer-NVFP4 \
--reasoning-parser muse \
--tool-call-parser muse \
--language-model-only \
--speculative-algorithm DFLASH \
--speculative-draft-model-path meta-models/Muse-Glimmer-30B-assistant \
--speculative-dflash-block-size 5 \
--tp-size 1 \
--mem-fraction-static 0.85 \
--host 0.0.0.0 --port 30000
Hardware notes:
- Fits a 32 GB RTX 5090 (the bf16 export does not).
- DGX Spark / GB10 (unified memory): use
--mem-fraction-static 0.40(0.38 with DFlash). Higher fractions let the KV pool consume the shared CPU/GPU pool and can OOM the machine during load. - Sampling: temperature 0.95, top_p 1.0 (greedy only for determinism checks).
Do not add
<|eom|>(200007) as an EOS token — it breaks parallel tool calling;generation_config.jsonalready carries the correct stops.