I kept using diffusers with python 3.10 for as long as I could bypassing the checks for python but kept noticing more and more issues.
I cleared my 180k file VENV that worked for 20 different training programs over the last two years and did a full update.
Here are the issues. (For windows users of course Linux your fine)
Bitsandbytes - can be compiled for Cuda 13.3 and Pytorch - You have to do this yourself
Xformers - Can be compiled but requires MSLK to work with python 3.13 - this was not the case for 3.10
This is a major speed issue as xformers was a bridge for SDPA, Cutlass and a lot of backend attention GEMMS
I did manage to make CMAKE files for MSLK that work with windows but I have been unable to fully build it. Despite hours of work I am stuck at:
2 errors detected in the compilation of "MSLK-main/csrc/conv/cutlass/f8f8bf16_conv/f8f8bf16_conv_128x128x128_2x1x1.cu".
f8f8bf16_conv_128x128x128_2x1x1.cu
ninja: build stopped: subcommand failed.I think a lot of the complaints I have been getting about model speed is do to me having 3.10 and working XFORMERS, and now I do not I see what the issue is.
Note: You can do
set MSLK_PYTHON_ONLY=1then in the MSLK git clone
pip install -e .However this does not link the cuda files, I am not sure if this allows Xformers to work or not.
[INFO] pytorch version: 2.14.1+cu132
[INFO] MSLK running in python-only mode — skipping native library loading
[INFO] xformers version: 0.0.35Without Flash attention, SDPA, Cutlass, Xformers, training and interface falls back to Pyattention which can double or triple training times.
While I have compiled flash attention for windows before it took so much time to custom build I really do not want to have to do that again.
While the training time looks ok, and triton is working I wonder about what attention its falling back too:
