What it is
Discrete-diffusion LLM with a lossless multi-token sampler and a Nano-vLLM runtime for faster decoding without a separate draft model. Generate multiple tokens in parallel from an autoregressive model with a provably lossless sampler, plus ready-to-run training, inference, and evaluation workflows for Uno checkpoints.
Key results
- Provably lossless multi-token prediction from the AR distribution
- No separately trained draft model is required for decoding
- Supports linear throughput mode and tree per-request throughput mode
Shows that discrete diffusion can accelerate LLM decoding while exactly preserving the base autoregressive distribution.