ray
56 lines
1.6 KiB
Markdown
56 lines
1.6 KiB
Markdown
# cu126 PyTorch for Pascal GPUs (CC < 7.5)
|
|
|
|
ComfyUI's `comfy --skip-prompt install --nvidia` installs the latest cu130
|
|
PyTorch build, which only supports compute capability >= 7.5 (Volta and
|
|
later). GPUs like the GTX 1050 Ti (CC 6.1, Pascal) produce a cryptic error:
|
|
|
|
```
|
|
RuntimeError: Found GPU0 which is of compute capability (CC) 6.1. The current
|
|
PyTorch installation supports... [sm_75, sm_80, sm_86, sm_90, sm_100, sm_120]
|
|
```
|
|
|
|
## Fix
|
|
|
|
After `comfy install --nvidia`, reinstall torch from the cu126 index:
|
|
|
|
```bash
|
|
/path/to/comfy/ComfyUI/.venv/bin/pip install --force-reinstall \
|
|
torch torchvision torchaudio \
|
|
--index-url https://download.pytorch.org/whl/cu126
|
|
```
|
|
|
|
cu126 supports CC 5.0 and up (sm_50, sm_60, sm_70, sm_75, sm_80, sm_86, sm_90),
|
|
covering Pascal, Maxwell, and all later architectures.
|
|
|
|
## Verify
|
|
|
|
```bash
|
|
/path/to/comfy/ComfyUI/.venv/bin/python -c "
|
|
import torch
|
|
print(f'CUDA: {torch.cuda.is_available()}')
|
|
print(f'Device: {torch.cuda.get_device_name(0)}')
|
|
print(f'CC support: {torch.cuda.get_arch_list()}')
|
|
"
|
|
```
|
|
|
|
Expected output for GTX 1050 Ti:
|
|
```
|
|
CUDA: True
|
|
Device: NVIDIA GeForce GTX 1050 Ti
|
|
CC support: ['sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90']
|
|
```
|
|
|
|
## Affected GPUs
|
|
|
|
All Pascal and earlier GPUs:
|
|
- GTX 10xx series (1050, 1060, 1070, 1080)
|
|
- GTX 9xx and earlier (Maxwell, Kepler, Fermi)
|
|
- Any GPU with CC < 7.5
|
|
|
|
## Why cu126 works
|
|
|
|
cu12.6 ships with nvidia-cublas, nvidia-cudnn, and all CUDA toolkit
|
|
components for multiple CC targets including sm_60. Later CUDA releases
|
|
(12.8+) dropped pre-compiled kernels for pre-Volta architectures to reduce
|
|
wheel size.
|