What's Changed
- Make TransformerEncoderLayer fields public by @Mnwa in #4053
- Feature muon by @NewBornRustacean in #3925
- Implement
FromStrforSlicewith parsing and error handling by @crutcher in #3983 - chore: Update to cubecl scalar refactor by @wingertge in #4062
- refactor: cubecl Runtime trait by @wingertge in #4065
- Fix scatter values backward by @khoek in #4064
- Refactor/autotuner by @nathanielsimard in #4068
- Fix MPS "Placeholder storage has not been allocated" error for embedding operations by @antimora in #4073
- Remove burn-import abstraction layer and use onnx-ir types directly by @antimora in #4033
- More correctness fixes in autodiff ops by @khoek in #4069
- Fix transaction read by @laggui in #4074
- Feat/error handling cubecl by @nathanielsimard in #4076
- Move types from
burn-tensorby @laggui in #4050 - burn-store enhancements for troubleshooting and new enum skip flag by @antimora in #4051
- Re-enabled no-std support for safetensors store by @antimora in #4071
- Fix tch bf16 kind by @laggui in #4088
- Feat/runtime error by @nathanielsimard in #4079
- Fix ConstantOfShape output size determination by @antimora in #4085
- Fix reduce codegen to use turbofish for squeeze_dims by @antimora in #4086
- Fix Expand operation to use ONNX max-semantics by @antimora in #4082
- Add ONNX GridSample op support and tests by @antimora in #4084
- Fix Slice operation to handle empty ranges by @antimora in #4083
- Add RF-DETR model check for burn-import by @antimora in #4087
- Fix cubecl by @BjornTheProgrammer in #4092