# Extra deps for the GPU training/inference path (Colab / Modal / RunPod).
# Install on top of requirements.txt on a CUDA box:
#   pip install -r requirements.txt -r requirements-gpu.txt
#
# NOTE: install a CUDA build of torch first (Colab ships one). The CPU box uses
# torch==*+cpu, which cannot run 4-bit QLoRA.
bitsandbytes>=0.43        # 4-bit NF4 kernels for QLoRA (requires CUDA)
# Optional speedup (UNCONFIRMED against transformers 5.13 — pin transformers down if you use it):
#   pip install unsloth