numpy
tqdm
# [ANVIL2] torch must come from the cu128 wheel index on a CUDA 13 runtime image: the
# attention kernel is the patched FA3 build devenpzak/flash-attn3-12864 (pinned in code at revision 64c1e6d1, the commit behind tag v1), whose one
# published variant is torch-stable-abi29-cu128 and links against libcudart.so.13.
#   pip install torch==2.10 --index-url https://download.pytorch.org/whl/cu128
torch==2.10
huggingface-hub==1.29.0
kernels==0.16.1
setuptools
typing-extensions==4.15.0
datasets
tiktoken