Denis Kuznedelev

DMax: Aggressive Parallel Decoding for dLLMs

Extending the Context of Pretrained LLMs by Dropping Their Positional Embedding

Mixture-of-experts (MoE) training megakernel

Gumbel-Softmax post-training quantization for LLMs (1–3 bit scalar, INT/GGUF-compatible, NVFP4).

Quartet II Official Code

Triton Playground