This is our trained DFlash baseline drafter for the Qwen3-8B target and a 16-token draft block. It is a draft model for speculative decoding. Use the target model's tokenizer.
model.safetensors contains the complete trained draft weights. config.json and the Python files contain the DFlash architecture. The target model is not included. This export was tested with SGLang 0.5.18 using --speculative-algorithm DFLASH, --speculative-draft-model-path pointing to this repository, and --trust-remote-code.
The starting draft architecture and code come from z-lab/Qwen3-8B-DFlash-b16. Our code is in the DynaDFlash repository.
z-lab/Qwen3-8B-DFlash-b16