Qwen3.8-27B-DSpark-Gauss-NVFP4-W4A4

A Static NVFP4 W4A4, Gaussian DSpark drafter derived from RedHatAI/Qwen3.8-27B-speculator.dspark at revision 7f33c272e5da240978e0d55767abab8193d74b95. Pair it with Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. This repository contains a drafter component, not a standalone chat model.

Quantization

Static NVFP4 W4A4 using seeded Gaussian calibration, 1,892 aligned calibration records, and sequence cap 2,048.

This arm used seeded Gaussian calibration values rather than prompts. The quantization manifest records 1,892 calibration records and a sequence cap of 2,048. Quantization commands, manifests, source files, patches, and the exported checkpoint checksum are in provenance/quantization/.

Example serving command

vllm serve Qwen/Qwen3.8-27B \
  --spec-model inference-optimization/Qwen3.8-27B-DSpark-Gauss-NVFP4-W4A4 \
  --spec-method dspark \
  --spec-tokens 8 \
  --kernel-config '{"linear_backend":"emulation"}'

For NVFP4, this example selects vLLM's emulation backend; it makes no claim of native H100 NVFP4 support.

Evaluation status

Evaluation is pending. No completed acceptance, speed, or quality results are included with this publication. The checkpoint has quantization provenance only; serving/runtime validation and the planned evaluation matrix have not completed.

Reproducibility

provenance/quantization/ includes the training and quantization commands, quantization manifest, calibration metadata where applicable, source scripts and patches, and the SHA-256 digest of the published drafter weights. Calibration prompt data is not redistributed.