mHC kernels implemented in CUDA

This project aims to replicate mainstream open-source model architectures with limited computational resources, implementing mini models with 100-200M parameters.

Official style files for papers submitted to venues of the Association for Computational Linguistics

Official implementation of Log-linear Sparse Attention (LLSA).

A compact implementation of SGLang, designed to demystify the complexities of modern LLM serving systems.

Learning about CUDA by writing PTX code.

TAAC2025初赛第十四名O_o队伍代码

Tile-Based Runtime for Ultra-Low-Latency LLM Inference

A collection of GPU experiments and benchmarks for my personal understanding and research.

Distributed MoE in a Single Kernel [NeurIPS '25]

https://bbuf.github.io/gpu-glossary-zh/

GRID: Generative Recommendation with Semantic IDs

Minimal reproduction of OneRec

Triton implementation of FlashAttention2 that adds Custom Masks.

轻量级大语言模型MiniMind的源码解读，包含tokenizer、RoPE、MoE、KV Cache、pretraining、SFT、LoRA、DPO等完整流程

Efficient Distributed GPU Programming for Exascale, an SC/ISC Tutorial

MedicalGPT: Training Your Own Medical GPT Model with ChatGPT Training Pipeline. 训练医疗大模型，实现了包括增量预训练(PT)、有监督微调(SFT)、RLHF、DPO、ORPO、GRPO。

大模型训练框架，支持现代大模型训练的全流程，包括基础预训练、长上下文预训练、思维链推理微调、强化学习训练、指令混合微调、直接偏好优化、模型权重合并、Web 推理服务

Efficient Triton Kernels for LLM Training

Benchmarking code for running quantized kernels from vLLM and other libraries

Implement a reasoning LLM in PyTorch from scratch, step by step

🤗 Optimum ONNX: Export your model to ONNX and run inference with ONNX Runtime

[ICML 2025] Official PyTorch implementation of "FlatQuant: Flatness Matters for LLM Quantization"

QQQ is an innovative and hardware-optimized W4A8 quantization solution for LLMs.

Reduce kernel based on CUTLASS CuTe and TMA.

The first W4A4KV4 quantized + 50% sparse LLMs!

A minimal PyTorch re-implementation of Qwen3 VL with a fancy CLI

VAE from Scratch in Pure C