Currently work in progress, some parts are very crappy.

This is an attempt to migrate Karpathy's llm.c to safe rust. By saying safe, I mean no unsafe is introduced in the code of this repository (without considering the dependencies).

For now, the performance of each training step of this rust version (opt-level=3) is similar to llm.c (in -O3). Note that more parallelism has been added to the rust version than C, so this is not actually a fair comparison.

Firstly, follow the CPU quick start steps in llm.c. Note the python scripts are under scripts/ and are the same as in llm.c.

pip install -r requirements.txt
python scripts/prepro_tinyshakespeare.py
python scripts/train_gpt2.py

Then just build and run with cargo:

cargo run --release

The rayon is used for parallelism, and most iterations have been parallelized. The number of threads is the default number in rayon.

To compare with the C version, change the optimization level in llm.c makefile to -O3 and make.

make train_gpt2
OMP_NUM_THREADS=12 ./train_gpt2

Haven't figured out how to compile with any equivalent of -Ofast in rust yet.

No serious benchmarking has been done yet but the performance of the rust version is similar to the C version for each training step on a MacBook M2 Max. Each step takes ~4.4s to finish.

There are more parallelism in this rust version than llm.c, so the performance should be (at least slightly) better.

Some parallelization strategies are not optimal, and some parts are not parallelized yet. The performance can be further improved.

Additionally, maybe wgpu can be used to accelerate the process.

I would like to express my sincere gratitude to ChatGPT for the assistance and support in this project.