UtevoLux-DFlash-test (fine-tuned DFlash drafter)

NOTICE: This is a derivative work of z-lab/Qwen3.6-27B-DFlash (MIT License, Copyright (c) 2026 Jian Chen, Yesheng Liang, Zhijian Liu / z-lab), via the fine-tuned drafter cont1037/Qwen-3.6-27B-AmberKite-DFlash (MIT). The weights were modified by further fine-tuning. See NOTICE and LICENSE. Not an official z-lab release.

A DFlash block-diffusion draft model for speculative decoding, fine-tuned to draft for the target model Tooony133/UtevoLux-3.6-27B (revision 589a12bc, a Qwen3.6-27B derivative) on image-to-Three.js code generation. It also drafts for further fine-tunes of that target.

Changes made to the original model

• Warm start: cont1037/Qwen-3.6-27B-AmberKite-DFlash @ 285795f0143bcb7a35efc4647a70bff22bb90b47 (z-lab layout: 5 layers, hidden 5120, 32/8 heads, head_dim 128, block size 8; all draft layers sliding-window 2048, full rotary).
• Fine-tuned 1 epoch (443 steps, LR 6e-5, 256 anchors per sequence, max sequence 20,480 tokens) with NeMo AutoModel on 1,769 prompts whose responses were generated by the UtevoLux target model (48 held-out prompts for validation).
• Weights only; no architecture or tokenizer change. Checkpoint epoch_1_step_443.

Result (same target model and system prompt, 64 held-out images)

drafter	vLLM 0.32, 7 draft tokens: accept length	SGLang 0.5.6, 8 draft tokens: accept length
cont1037/UtevoLux-3.6-27B-DFlash-ft	4.07	4.1
this model	4.81	5.03

Usage

SGLang: --speculative-algorithm DFLASH --speculative-draft-model-path <this repo> --speculative-num-draft-tokens 8 vLLM: --speculative-config '{"method":"dflash","model":"<this repo>","num_speculative_tokens":7}'

Citation (original DFlash work)

@article{chen2026dflash,
  title   = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author  = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  journal = {arXiv preprint arXiv:2602.06036},
  year    = {2026}
}