Modelpublic

mgaruccio/Muse-Glimmer-30B-assistant-GPTQ-Int4-sym-G128

sourceHugging Faceapache-2.0updated 21d agoView on Hugging Face

0likes82downloads

Model Card

Muse Glimmer DFlash assistant GPTQ Int4 G128

GPTQ 4-bit / group-128 / symmetric of Muse Glimmer's 5-layer DFlash drafter. Apache 2.0, derived from Meta's DFlash assistant (not a standalone chat model).

Use as the speculative draft for vLLM-XPU on one Arc Pro B70 with `Muse-Glimmer-30B-GPTQ-Int4-sym-G128`. Embeddings and lm_head are shared with the target at runtime.

—Quantizer: GPTQModel 7.3.2
—35 decoder linears quantized; encoder / norms BF16
—vLLM fused module names in quantize_config.json: qkv_proj, o_proj, gate_up_proj, down_proj

Experimental B70 concurrency sweep artifact

The C8–C128 short-burst sweep used this assistant with an experimental frozen 32,768-token vocabulary shortlist and DFlash K3. Download the public-safe runtime input from `artifacts/glimmer-b70-k3-shortlist-32768.json`. It contains only the immutable original-vocabulary IDs and contract metadata—no calibration prompts, votes, raw generations, or model weights.

This is not the default recipe or a production capacity claim. Its highest observed short-burst median was 840.8 aggregate tok/s at C96; C48 was the workload-specific latency/throughput knee. The measured workload used repeated 83-token prompts and capped 256-token reasoning-only outputs. See the presentation report for methodology, caveats, and the exact source patch at `b70-inference@faf4ba9`.

bash

hf download mgaruccio/Muse-Glimmer-30B-assistant-GPTQ-Int4-sym-G128 --local-dir ./models/draft