z-lab/dflash

Health v2 · maintenance only

Adoption not part of the score

5967 stars · 414 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

DFlash is a lightweight block diffusion model used as a draft model for speculative decoding of large language models, drafting entire token blocks in a single parallel forward pass. It ships as a pip-installable Python package with checkpoints for many model families and integrates with Transformers, MLX, SGLang, vLLM, oMLX, and llama.cpp serving backends.

Use cases

• speed up LLM inference with speculative decoding
• draft multiple tokens in parallel for a target LLM
• serve Qwen or Llama models with faster generation in SGLang
• benchmark speculative decoding speedups against EAGLE-3
• run lossless LLM acceleration on Apple Silicon with MLX
• use a diffusion-based draft model with vLLM or llama.cpp

When to choose

• you want lossless LLM inference speedups beyond autoregressive drafters like EAGLE-3
• your target model has an official DFlash checkpoint (Qwen, Gemma, Llama, Kimi, MiniMax, GLM, GPT-OSS)
• you serve models via SGLang, vLLM, oMLX, or llama.cpp OpenAI-compatible servers

When to avoid

• your target model family has no available DFlash checkpoint
• you need training or fine-tuning of the draft model rather than inference
• you cannot run a GPU or Apple Silicon backend

Facets

library · maturity active

llm-inference machine-learning deep-learning large-language-models machine-learning deep-learning performance python speculative-decoding block-diffusion llm-acceleration draft-model sglang vllm mlx transformers inference-optimization linux macos gpu docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="z-lab/dflash")