Skip to content

README.md

Latest commit

聽

History

History
64 lines (49 loc) 路 3.69 KB

README.md

File metadata and controls

64 lines (49 loc) 路 3.69 KB
Raw
ant's friendly logo

Description

ant is a PyTorch-based Deep Learning framework. It is inspired by nanoGPT, although it does not borrow code from it. It tries to keep being research-friendly, while integrating (way) more features, following modern best practices and being modular. Specifically, among others, ant supports:

  1. The CIFAR-10, OpenWebText, FineWeb-Edu and ClimbMix (SoTA) datasets
  2. Any Hugging Face or TokenMonster (SoTA) tokenizer
  3. ResNet, GPT2, Llama 2, nGPT, and OLMo 2 models
  4. The PSGD, DistributedShampoo, AdEMAMix, SOAP, Muon and Scion optimizers
  5. 渭P for zero-shot hyperparameter transfer
  6. Downstream evaluations (HellaSwag, ARC etc.) through lm_eval during training
  7. Model summary (parameters and FLOPS) and computational graph visualization
  8. Offline logging of gradients and weights in simple .dat files
  9. Attention heatmaps
  10. Plotting through PGFPlots
  11. Pure PyTorch attention, FlashAttention, FlexAttention and cuDNN attention with RoPE, ALiBi and Sliding Window Attention (SWA)
  12. Distributed Data Parallel, torch.compile and Automatic Mixed Precision
Model summary, learning rate schedule and logging in the terminal.

Computational graph of a Transformer with RoPE.
Gradient and weight statistics of the Embeddings Table.
Attention weights for a Transformer block with twelve heads.

Getting Started

Tip

All scripts have a help menu available via --help. In cases this is not enough, it is recommended that you look at the code (self-documenting) directly. Alternatively, if you are a vibe coder 馃, you can try feeding the whole codebase to an LLM (e.g. via https://uithub.com/gvlassis/ant), or a coding agent (e.g. Codex CLI).

  1. Clone the repo:

     git clone https://github.com/gvlassis/ant.git

  2. Install PyTorch and FlashAttention.

  3. Install requirements.txt:

     pip install -r requirements.txt

  4. Prepare the dataset via ./src/data/make.py. The dataset is first downloaded from Hugging Face, processed, and then saved as tensors in .pt files. If you are lazy 馃Ε, you can also directly download the artifacts of the following command.

     python ./src/data/make.py --dataset climbmix10m

  5. Train a neural network via ./src/train.py. For training on cuda:0:

     # If you are not using 渭P, k_input is the learning rate
 python ./src/train.py --opt muon --micro_batch_size 32 --train_batches 2000 --k_input 3e-2 --momentum 0.95 --model_device_index 0 ./out/test

    A lot of settings (e.g. depth, number of heads) are configured in ./src/models/utils_models.py/get_model_opts() and ./src/models/transformer.py. For one node with 4 GPUs:

      OMP_NUM_THREADS=1 torchrun --standalone --nproc_per_node=4 ./src/train.py --opt shampoo --micro_batch_size 32 --train_batches 2000 --k_input 3e-3 --momentum 0.95 --beta2 0.95 --eps 1e-10 ./out/test