I've been trying to get this repo running for hours and I'm hitting a wall of dependency conflicts. Could you clarify the exact verified combination of torch and cuda versions to use?
My setup:
- GPU: NVIDIA RTX 3090
- CUDA: 13.2
- OS: Windows 11
Issues encountered:
AttributeError: module 'torch' has no attribute 'float8_e8m0fnu' when using latest stable torch (2.6.0).- Trying to use
torch nightly (2.13.0+) to fix the above leads to RuntimeError: Configured CUDA binary not found at ... libbitsandbytes_cuda132.dll in bitsandbytes.
pyproject.toml defines torch>=2.11, which doesn't exist on standard CUDA channels, causing pip to pull CPU-only versions or conflict with torchaudio/torchvision.
What is the officially supported environment to run this without patching bitsandbytes or hacking pyproject.toml?
I've been trying to get this repo running for hours and I'm hitting a wall of dependency conflicts. Could you clarify the exact verified combination of
torchandcudaversions to use?My setup:
Issues encountered:
AttributeError: module 'torch' has no attribute 'float8_e8m0fnu'when using latest stable torch (2.6.0).torchnightly (2.13.0+) to fix the above leads toRuntimeError: Configured CUDA binary not found at ... libbitsandbytes_cuda132.dllinbitsandbytes.pyproject.tomldefinestorch>=2.11, which doesn't exist on standard CUDA channels, causingpipto pull CPU-only versions or conflict withtorchaudio/torchvision.What is the officially supported environment to run this without patching
bitsandbytesor hackingpyproject.toml?