What's Changed
- Bump version by @nathanielsimard in #4102
- Handle empty tensors in cat and slice_assign ops by @antimora in #4095
- Add network utilities to
burn-stdby @laggui in #4104 - Remove RefCell from onnx-ir Arguments by @antimora in #4094
- Fix raspberry pi pico example not compiling by @BjornTheProgrammer in #4034
- Flash Attention module by @louisfd in #4089
- [Breaking] Add
IndexingUpdateOptoscatterandselect_assignby @laggui in #4070 - Feat/improve errors by @nathanielsimard in #4110
- Add 256-byte tensor alignment to burnpack format for mmap zero-copy support by @antimora in #4100
- Add CrossAttention module to burn-nn by @huy209vn in #4101
- Add reflect and edge padding modes to tensor.pad by @antimora in #4105
- Add LSTM operator support with configurable activations by @antimora in #4106
- Add memory-mapped ONNX loading with lazy tensor data by @antimora in #4097
- Refactor
RemoteDeviceto use a thread-safe global address registry. by @crutcher in #4113 - Partial cleanup of RemoteSender api. by @crutcher in #4108
- Move backend traits and types to
burn-backendby @laggui in #4111 - Fix remote sync error by @laggui in #4117
- Small LSTM clean up of unused variable by @antimora in #4116
- Fix/autotune checks by @nathanielsimard in #4114
- Include katex header as symlink by @laggui in #4118
- chore: Update cubecl by @wingertge in #4120
- Fix GLU and quiet softmax activations by @laggui in #4121
- Migrate ONNX import to burnpack format (removing Record type) by @antimora in #4122
- Combined PRs by @github-actions[bot] in #4140
- Chore/pre release 5 by @nathanielsimard in #4141