Toolchain: Python and PyTorch

Two version decisions here are load-bearing. Both have failure modes that look like something else, which is why they get their own page.

Do not use the system Python

The host ships Python 3.14, which is ahead of what PyTorch publishes wheels for. Installing into it fails at best and builds something unsupported at worst. We install uv (opens in new tab) into the user’s own ~/.local/bin and pin a 3.12 virtual environment, leaving the system interpreter untouched:

curl -LsSf https://astral.sh/uv/install.sh -o /tmp/uv-install.sh && sh /tmp/uv-install.sh
uv venv ~/projects/comfy-venv --python 3.12

3.12 rather than 3.13 is deliberate: custom ComfyUI nodes are the least-maintained part of this ecosystem, and they lag the newest interpreter by a wide margin.

Match the CUDA build to the driver — not to habit

The card is Blackwell, compute capability sm_120. A PyTorch build without sm_120 kernels does not necessarily error; it can fall back or fail deep inside a generation with a confusing message. Verify the architecture is actually compiled in, rather than trusting that CUDA “works”.

The driver on this box exposes CUDA 13.3, so we take the cu130 wheels:

uv pip install --python ~/projects/comfy-venv/bin/python \
  --index-url https://download.pytorch.org/whl/cu130 \
  torch==2.11.0 torchvision==0.26.0 torchaudio==2.11.0

Why 2.11.0 and not the newest torch

At the time of writing the index carried torch up to 2.13.0, and taking the newest is the obvious move. It is the wrong one here:

torchaudio stops at 2.11.0 on every CUDA index. ComfyUI’s requirements.txt lists torchaudio as a dependency. Installing torch 2.13 leaves that requirement unsatisfiable without a version fight, and resolving it drags torch back down anyway.

2.11.0 is the newest fully-matched trio — torch 2.11.0, torchvision 0.26.0, torchaudio 2.11.0. Take the matched set. The pairing is lockstep: torch 2.x goes with torchvision 0.(x+15) and torchaudio 2.x.

Verify before going further

Detection is not the same as having kernels. Check the architecture list and run a real computation:

import torch
print(torch.__version__, torch.version.cuda, torch.cuda.is_available())
print(torch.cuda.get_arch_list())            # must contain 'sm_120'
a = torch.randn(4096, 4096, device="cuda", dtype=torch.float16)
print(float((a @ a)[0, 0]))                  # a real kernel launch

On this machine that prints 2.11.0+cu130 13.0 True, an arch list ending in sm_120, and a finite number. If the arch list lacks your card’s capability, stop and fix it now — everything downstream will be mystifying otherwise.