NEW vs LEGACY GPU setup

PyTorch CUDA wheels drop older GPU support over time. Pascal (sm_61) needs a legacy setup.

Decision table

NEW GPUs

Use: requirements_new_gpu.txt (Python 3.12+ fine)

python3.12 -m venv venv
source venv/bin/activate
pip install -U pip
pip install -r requirements_new_gpu.txt

This installs the latest PyTorch with CUDA support. Works with any GPU that has compute capability sm_75 or higher (Turing, Ampere, Ada Lovelace, Hopper).

LEGACY GPUs (Pascal — GTX 1080 / 1080 Ti)

Use: requirements_legacy_gpu.txt (requires Python 3.10)

python3.10 -m venv venv
source venv/bin/activate
pip install -U pip
pip install -r requirements_legacy_gpu.txt

The legacy requirements pin specific PyTorch and CUDA versions that still include sm_61 support. Key constraints:

• Python 3.10 — newer Python versions may not have compatible legacy wheels
• Pinned Torch version — a specific torch+CUDA build that includes Pascal kernels
• Pinned dependencies — numpy, etc. pinned to versions compatible with the older torch

Why Python 3.10?

Modern PyTorch wheels (2.2+) dropped Pascal support. The last compatible builds target Python 3.10 with CUDA 11.8. Using a newer Python may result in torch finding no matching wheel or silently falling back to CPU-only.

Verifying GPU access

After installing, confirm the worker can see the GPU:

python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

Expected output for a GTX 1080 Ti:

True
NVIDIA GeForce GTX 1080 Ti

CPU-only workers

Either install path works for CPU-only workers. The NEW path is simpler. Workers will train on CPU (slower, but functional). The coordinator handles mixed GPU and CPU workers automatically.

python3.12 -m venv venv
source venv/bin/activate
pip install -U pip
pip install -r requirements_new_gpu.txt

Troubleshooting