Upgrade to torch 2.9.1 and allow python 3.13

[dxqb]

git fetch origin pull/1266/head:pr-1266
git switch pr-1266

then update.sh / update.bat

based on #1189 and #1125 (comment)

Torch 2.9.1 #1189 has been well-tested but has not provided a clear benefit for us over torch 2.8, so we haven't upgraded yet. Maybe we skip 2.9.1 and go straight to 2.10 if that provides a benefit?

Does it fix

• - [Bug]: Double the ram usage after upgrading to Torch 2.8 #1125 ? - No.
• - [Bug]: Multi GPU error on Windows: unsupported gloo device #1122 ? - No.
• - [Bug]: Compile transformer blocks when on Windows / MAX_PATH #1120 ? - Yes, by upgrading triton
• -

  OneTrainer/modules/util/compile_util.py

  Line 8 in 661872c

  #can be removed once https://github.com/pytorch/pytorch/pull/169726 is merged into a torch version we use

  ? - leave it there for now - doesn't hurt
• should we also upgrade to cuda 13.0?
• upgrade -default and -rocm dependencies (to the best of our knowledge)

[dxqb]

don't upgrade rocm: pytorch/pytorch#173308

[dxqb]

on windows I was getting an error by training loras with Klein-9B and torch triton, I'm not sure if one these package upgrades solves it or by changing the kernel name from "__mm_kernel" to "_mm_kernel", but these changes solve it 🤷‍♂️

this was a torch bug introduced in torch 2.9 and is worked around by this PR. I guess you didn't use the torch version installed by OneTrainer before (2.8)

[dxqb]

@O-J1 IIRC you have already tested this PR. can you confirm?
I'm using it on linux successfully, but a test on Windows (especially Triton version) should be done before merge

[O-J1]

@O-J1 IIRC you have already tested this PR. can you confirm? I'm using it on linux successfully, but a test on Windows (especially Triton version) should be done before merge

Yes. Compiled and non compiled with SDXL and Flux 1 both worked on windows with an NVIDIA GPU on 2.10. The more recent upgrade ‘all dependencies’ PR is broken however.

[dxqb]

torch 2.10 has a significant vram regression, that would break user settings and presets:
pytorch/pytorch#175058

We can upgrade to 2.9.1 instead, or wait for torch 2.11 and/or a workaround