Plug-and-play PyTorch transformer research framework. Swap attention, FFN, norm, positional, connection, projection, optimizer, scheduler, l

Agentic RL Training at Scale

a small replication of the paper "q0: Primitives for Hyper-Epoch Pretraining"

Fused MoE kernel

a learning repo for on-policy distillation experiments, notes, and small implementation probes

CUDA kernels for the Muon optimizer: Newton-Schulz polar iteration (quintic + Gram-NS variants), benchmarked and verified against torch.opti

Mixture-of-experts (MoE) training megakernel for NVL72s

cs undergrad

Miles is an enterprise-facing reinforcement learning framework for LLM and VLM post-training, forked from and co-evolving with slime.

comparing length penalty formulations for reasoning models under the same GRPO training setup

Learning cuda (aka gpu programming)

chess model that treats board states as tokens. learns rules from millions of games, studies the legends, then plays itself to get better.

GPU-accelerated Canny edge detector in CUDA C++. Parallelizes Gaussian filtering, gradient computation, non-maximum suppression, and hystere

From-scratch BPE tokenizer implementation in C++ with UTF-8 support including emojis.

Self-playing Connect4 agent inspired by AlphaGo/AlphaZero algorithms using reinforcement learning.