handdl

Minimal tensor, sequence, and pipeline parallelism for LLaMA.

Layer-wise weight offloading with profiling-driven optimizations.

Optimizing LLM throughput via binned padding, sequence packing, and Flash Attention.

Fused and tiled ops, torch.compiled and memory-optimized kernels, and the math behind it.

Minimal implementation of DeepSeek-R1-Zero.

Why 'King-Queen=Man-Woman' or a simulation for the paper 'Analogies Explained: Towards Understanding Word Embeddings'

A minimal mech-interp project for steering LLMs

torch-like convolution over binary trees