wehub-resource-sync
20 lines
663 B
Plaintext
20 lines
663 B
Plaintext
# 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)
|