Python
523 stars
[ICLR 2026] Official code for TraceRL: Revolutionizing post-training for Diffusion LLMs, powering the SOTA TraDo series.
### Overview
TraceRL is an official code repository for the ICLR 2026 paper titled "Revolutionizing post-training for Diffusion Large Language Models". It provides comprehensive support for various discrete diffusion language models, including TraDo, SDAR, Dream, LLaDA, MMaDA, LLaDA-V, and Diffu-Coder. The repository includes a variety of settings and architectures, making it highly versatile for both deployment and reinforcement learning tasks. The model is optimized for inference acceleration, with features like KV-cache and JetEngine, and supports multi-nodes for efficient inference.
### Key Features
- **Model Support**: Supports TraDo, SDAR, Dream, LLaDA, MMaDA, LLaDA-V, and Diffu-Coder, almost all open-sourced discrete diffusion language models.
- **Diverse Settings**: Supports deployment, Semi-AR SFT, RL (with optional value model for variance reduction and process reward model for fine-grained supervision), and RLHF across diverse settings (math, coding, multimodal), and different architectures (both full/block attention dLLMs).
- **Inference Acceleration**: Improves KV-cache, JetEngine (based on nano-vllm), and supports multi-nodes, making inference faster and more efficient.
- **RL Training**: Supports TraceRL, coupled RL, random masking RL, accelerated sampling, and supports multi-nodes and long-CoT finetune.
- **SFT**: Supports Block SFT, semi-AR SFT, random masking SFT, and long-CoT