# Post-training stack for the from-scratch SFT / RM / PPO / DPO / GRPO suite.
# Targets H100 + CUDA 12.x (the base requirements.txt pins cu118 for the original
# pretraining path; this file does NOT overwrite it). Install into a fresh venv:
#
#   python3 -m venv /ephemeral/venv && source /ephemeral/venv/bin/activate
#   pip install -r requirements-post.txt
#
# torchrun (multi-GPU) ships with torch.

--extra-index-url https://download.pytorch.org/whl/cu121

torch
numpy
h5py
tqdm
tiktoken
zstandard
datasets          # HuggingFace dataset downloads (Alpaca, HH-RLHF, GSM8K, ...)
wandb             # optional experiment logging (every script also writes JSONL)