Guillaume Lagrange

I would leave this issue open, I agree that the error should be communicated more clearly. Probably a `Result` where an invalid return can only be for a `DowncastMismatch` at...

If the param id is not in the container, I still think it makes sense to return `None`. This behavior is in line with other standard containers like hash maps....

> If the param id is not in the container, I still think it makes sense to return `None` Hmm right, after thinking about it it wouldn't be as ergonomic...

Sorry I was not subscribed to this older issue, didn't get notifications for the first question 😅 1. This is not usually required for loss functions since they are a...

Yeah the gather op translates to `tensor.select(dim, indices)` but the indices are expected to be 1d. There's a discrepancy between ONNX spec and the Burn tensor method.

@raphaelmenges yepp, we do not support ceil mode for pooling in burn right now.. so any ONNX models that use this configuration will panic 🥲 https://github.com/tracel-ai/burn/blob/89de0e13926291028817983d65c242906d3b649b/crates/burn-nn/src/modules/pool/avg_pool2d.rs#L33-L55

Hmmm it seems wgpu is failing to find a suitable Vulkan adapter on Alpine. Might be missing a runtime dep in your env setup, but I am not sure tbh.

I didn't mean a Burn runtime dependency, but rather a wgpu/Vulkan setup issue on your Alpine Linux setup. So it seems that it is failing to find a suitable Vulkan...

> I find it odd that I game with Vulkan no problem, vkcube runs fine, and burn and llamacpp have issues. It's a real conundrum to me. This is most...

Thanks for the bug report! Will take a look. /edit: this is clearly a fusion bug ``` default_program(27): Error: Formal parameter space overflowed (36800 bytes required, max 32764 bytes allowed)...