dllm

Why we included this project

Diffusion language models are moving fast, but the tooling around them tends to be scattered across individual paper codebases. dLLM gathers pretraining, finetuning, and evaluation into a single stack built on the Hugging Face Trainer, with LoRA, DeepSpeed, and FSDP support, so scaling up doesn't mean reimplementing training loops. It wraps EleutherAI's lm-evaluation-harness too, which makes benchmarking open-weight diffusion models much simpler. For researchers and engineers working with LLaDA, Dream, or smaller BERT-based diffusion chatbots, the repo includes recipes for masked and block diffusion, edit flows, and GRPO-based reasoning training. It's a practical, reproducible reference if you want to extend published diffusion LM work rather than build everything yourself.