-
Notifications
You must be signed in to change notification settings - Fork 341
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Compatibility with nightly PyTorch for M1 acceleration #542
Comments
That's great, note that #538 adds support for PyTorch 1.13 which is at the RC stage. Once PyTorch 1.13 has been released (this is planned for the 27th of October), this will be merged into the main branch and hopefully this will include the changes that you need for M1 GPU acceleration. |
The 1.13 version is now released and the github tip uses it, so hopefully this will give you M1 acceleration. |
I don't have a local pytorch installation so I suppose the sys crate downloaded a CPU version. It can't compile and I get a bunch of these:
Any ideas? |
The upstream libtorch archives still ship without aarch64 support. Just install the Python package and then:
|
I authored pytorch/pytorch#88911 to hopefully fix this. Unclear if that PR is all that's needed. |
This works for compiling this repo. However, compiling my project that depends on this crate (using the main branch) results in the following error:
|
Could you try with
? |
This works! Thanks. |
Closing this in favor of #488 |
tch
does work on M1 at present (see https://github.com/ssoudan/tch-m1), however MPS acceleration is only implemented for PyTorch Nightly andtch
is only compatible with PyTorch 1.12. Adaptingtch
to the breaking API changes will need to be done anyways once the current Nightly becomes a full release, and it seems to be the only barrier to M1 GPU acceleration which is rather in demand (#524, #518, #517, #501, #488, #464, #440, #412, #346).Here's the full compilation error
The text was updated successfully, but these errors were encountered: