Qwen3.8-27B-DFlash2 — FP8 (vLLM)

FP8-quantized DFlash 2 draft model for Qwen/Qwen3.8-27B, packaged for speculative decoding in vLLM.

This is not a standalone language model. It runs inside a speculative decoding server and drafts token blocks for the target model to verify. It is an FP8 requantization of the DFlash 2 drafter published by Inco AI / incoai/Qwen3.8-27B-DFlash2.

What this is

DFlash 2 is a block-diffusion drafter for speculative decoding: it predicts a whole block of tokens in a single pass, keeps the top candidates at every position, and a lightweight selector traces one coherent path through them. Two-tap dynamic convolutions in the backbone keep the draft from decaying toward the end of the block.

This repository quantizes the transformer projection weights (q/k/v/o_proj, gate/up/down_proj) to FP8 (E4M3) using compressed-tensors:

• Weights: FP8, per-channel static scales, symmetric.
• Activations: FP8, dynamic per-token.
• Kept in higher precision (not quantized): the DFlash convolution taps (attention_conv, mlp_conv), the candidate selector, the fc head, all norms, and the embeddings.

The FP8 weights shrink the drafter to ~2.1 GB. Because the target model verifies every drafted token, speculative decoding stays distribution-preserving regardless of drafter precision — quantizing the drafter changes acceptance length (speed), not the correctness of the emitted tokens.

Usage (vLLM)

Requires a vLLM build with DFlash speculative-decoding support (vLLM PR #53122). Pass this repo as the speculative model and declare compressed-tensors so the FP8 drafter weights are loaded correctly:

vllm serve Qwen/Qwen3.8-27B \
  --speculative-config '{
    "method": "dflash",
    "model": "TechPrototyper/Qwen3.8-27B-DFlash2-fp8-vllm",
    "quantization": "compressed-tensors",
    "num_speculative_tokens": 7
  }'

The drafter uses a block size of 8 (7 draft tokens per verification step). Use Qwen3.8's recommended sampling parameters (temperature 1.0, top_p 0.95, top_k 20).

Attribution

• Drafter architecture and weights: DFlash 2 by Inco AI — incoai/Qwen3.8-27B-DFlash2, built on the original DFlash work from z-lab/dflash.
• vLLM integration path: DFlash speculative decoding, vLLM PR #53122.
• FP8 requantization: this repository (compressed-tensors, per-channel).

Citation

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = {August},
  url    = {https://inco.ai/blog/dflash2/}
}

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}