Guillaume Lagrange

Looks like we maxed out the parameter space with element-wise fusion. elemwise_fuse def ``` extern "C" __global__ void elemwise_fuse ( float_4 buffer_0[], float_4 buffer_1[], float_4 buffer_2[], float_4 buffer_3[], float_4 buffer_4[],...

The split implementation added in #2568 relies on the input shape https://github.com/tracel-ai/burn/blob/37822fdb51a21acff20bae6a9dce24aced0cf575/crates/onnx-ir/src/dim_inference.rs#L912-L915 Now that I think about it, I don't think that's required to get the output rank. The input...

We're always open to PRs! Feel free to open one to add this to the modules in `burn-core` 🙂

No worries! This can be done from a fork of the repository. Check out the details here: https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/proposing-changes-to-your-work-with-pull-requests/creating-a-pull-request-from-a-fork

#2356 #2455 🔥

Thanks for flagging the issue! Do you have a direct link to the onnx model by any chance? I cannot find the file on the linked HF hub repo. The...

Just looked at the model really quick and it is what I suspected. The input to `Squeeze` comes from a `Shape` output. And the code currently does not support this....

Linking https://github.com/tracel-ai/burn/issues/2924 The reason is simply because gradients are not stored on the autodiff backend. The values in the tensor container for the gradients do not need to be tracked,...

I think this PR can focus on the arrow dataset changes, and if you feel like following up for the csv and json sources another PR can be opened after...

Closing due to inactivity, but feel free to re-open if you want to pick this up again! :slightly_smiling_face: