In the spirit of NanoGPT, we created Picotron: The minimalist & most-hackable repository for pre-training Llama-like models with 4D Parallelism (Data, Tensor, Pipeline, Context parallel). It is designed with simplicity and educational purposes in mind, making it an excellent tool for learning and experimentation.
-
The code itself is simple and readable:
train.py,model.pyand[data|tensor|pipeline|context]_parallel.pyare all under 300 lines of code. -
Performance is not the best but still under active development. We observed 38% MFU on a LLaMA-2-7B model using 64 H100 GPUs and nearly 50% MFU on the SmolLM-1.7B model with 8 H100 GPUs. Benchmarks will come soon
- A step by step tutorial on how to build Picotron distributed training framework form scratch:
pip install -e .
-
Get a HF token here to download models from HuggingFace
-
GPU
# To create a config file in json format under tmp by default python create_config.py --out_dir tmp --exp_name llama-1B --dp 8 --model_name HuggingFaceTB/SmolLM-1.7B --num_hidden_layers 15 --grad_acc_steps 32 --mbs 4 --seq_len 1024 --hf_token <HF_TOKEN> # Locally torchrun --nproc_per_node 8 train.py --config tmp/llama-1B/config.json # 3D Parallelism python create_config.py --out_dir tmp --dp 4 --tp 2 --pp 2 --pp_engine 1f1b --exp_name llama-7B --model_name meta-llama/Llama-2-7b-hf --grad_acc_steps 32 --mbs 4 --seq_len 1024 --hf_token <HF_TOKEN> # Slurm python submit_slurm_jobs.py --inp_dir tmp/llama-7B --qos high --hf_token <HF_TOKEN>
-
CPU (expect it to be slow)
# 3D Parallelism on CPU python create_config.py --out_dir tmp --exp_name llama-1B-cpu --dp 2 --tp 2 --pp 2 --pp_engine 1f1b --model_name HuggingFaceTB/SmolLM-1.7B --num_hidden_layers 5 --grad_acc_steps 2 --mbs 4 --seq_len 128 --hf_token <HF_TOKEN> --use_cpu # Locally torchrun --nproc_per_node 8 train.py --config tmp/llama-1B-cpu/config.json