dhcode95

X-R1. minimal-cost for training 0.5B R1-Zero

NSA-pytorch. DeepSeek Native Sparse Attention pytorch implementation

grpo-loss. Jupyter Notebook

easy-dualpipe. Pipeline-Parallel Lecture: Simplest Dualpipe Implementation.

mHC-pytorch. pytorch implementation of Manifold-constrained Hyper Connection (mHC)

DeepSeek-V4-mini. DeepSeek-V4 Lecture

Engram-pytorch. pytorch implementation of DeepSeek Engram

online-softmax. simplest online-softmax notebook for explain Flash Attention

cut-cross-entropy-pytorch. pytorch notebook for implemention for cut-cross-entropy LLM training.

easy-infer. LLM Inference Server Lecture: reproduce Continue Batching、vLLM、PD-disaggreation、ChunkedPrefill

PKU-SafeRLHF-DPO-Demo. Jupyter Notebook