Guillaume Lagrange
Guillaume Lagrange
Just added additional support for 16-byte uuid. Turns out the binary model doesn't have to be updated, just had to add backward compatibility for the parameter ids 🙂
@hexd0t you can also review since you contributed the last PR
Seems like the model generated for the test actually is a `TensorType` with dim 0 and not a scalar type. The error was reported to me by a user, and...
Can you share a MWE? Looks like an operation might not be respecting the device limits somehow.
Cross-posting my discord answer: This actually appears to be happening during autotuning for convolutions. The `im2col` kernel is trying to use too much memory, but ideally this should result in...
> Learning burn rn and happen to have an M4 on MacOS 26, so I tried to reproduce this (what i can gather) with a simple norm.forward on wpgu with...
> IMO maybe we can have something like `pub mod safetensors;` under a new feature gate in crate `burn-import` so users could build exactly what they need, since safetensors does...
> **I would prioritize Rust code de-duplication first**. We can leave the example and test duplicated for now (because it will take time). Agreed for the bold part.
> Just a feature request, can ignore this if you think it is unnessesary😂: since safetensors now supports `no_std`, could we also have a `SafeTensorBytesRecorder`? A bit hesitant on adding...
> Thanks for reviewing the long PR. Takes up lots of context switching. Of course! Longer PRs take more time to review carefully so time-to-merge is usually longer as well...