| # GPU-host requirements for scripts/autotune/train_qlora.py (S3, plan_autotune.md). |
| # |
| # Target box: rented Linux GPU (RunPod/Vast, single 4090, 24GB VRAM), |
| # CUDA 12.1+. Do NOT install this on the Windows dev box — there is no GPU |
| # here and these packages are not needed for anything except this one script |
| # (see the import guards in train_qlora.py: torch/unsloth/trl/datasets are |
| # imported lazily inside the training function precisely so the rest of the |
| # repo — ruff, mypy, pytest — stays green without them). |
| # |
| # Install order on the rented node: |
| # 1) pip install --upgrade pip |
| # 2) pip install unsloth |
| # # unsloth's installer auto-detects CUDA/torch and pulls a matching |
| # # torch build itself. If auto-detect misfires on the rented image, |
| # # unsloth prints the exact extras string to use instead — see |
| # # https://github.com/unslothai/unsloth#installation-instructions |
| # 3) pip install -r scripts/autotune/requirements_gpu.txt |
| # |
| # Versions below are lower-bounded, not tightly pinned: this is a one-shot |
| # rental image, not a maintained environment, so let pip resolve within the |
| # range unless a specific combo breaks — tighten only then. |
|
|
| datasets>=3.0 |
| peft>=0.13 |
| trl>=0.12 |
| bitsandbytes>=0.44 |
| accelerate>=1.0 |
| transformers>=4.46 |
|
|
| # vLLM is for the serving step after training, not for training itself — |
| # only needed if you serve from this same box instead of a separate step. |
| vllm>=0.6.3 |
|
|
