Guillaume Lagrange

Thanks for flagging! For even kernel sizes the "same" padding scheme actually requires asymmetric padding in order to produce the same output shape. We only support symmetric padding right now,...

The metal bug is different, tracked in #3780 . `ReferenceNotFound` should not happen with autotune / fusion, I will have a look.

We're aware of this, but the suggested solution by the compiler is appropriate. The upgrade to wgpu v24 seems to have increased the complexity of some types. So when using...

> Not the best bug to run into after renting 8 H100s 😅 Oh no, sorry to hear that 😬 Given the last portion of the experiment logs ``` 2025-07-23T10:24:36.812607Z...

Probably just need to have a function in the backend trait, e.g. ```rust fn supports_dtype(device: &Self::Device, dtype: DType) -> bool; fn supports_element(device: &Self::Device) -> bool { Self::supports_dtype(device, E::dtype()) } ```

I think this can be an extension of the current `Backend` trait instead, not a new trait. The current support should be limited to the currently defined backend types (e.g.,...

That's awesome! It would be a great addition to the examples if you want to submit a PR for that 😎 Since these types of examples are a bit more...

Complex numbers have different semantics than floats. I would opt for a different kind (option 2).

Seems like a rust nightly issue. We don't test against nightly, but if you use stable there shouldn't be an issue.

Take a look at the [text classification example](https://github.com/tracel-ai/burn/tree/main/examples/text-classification) 🙂 more specifically, how the inputs are [moved to the same device](https://github.com/tracel-ai/burn/blob/main/examples/text-classification/src/model.rs#L68-L71) as the model.