tracel-ai/burn v0.21.0-pre.3

on GitHub

What's Changed

• Fix select_assign OOB (#4760) @nathanielsimard
• Fix unsqueeze_dims panic (#4755) @softmaximalist
• Fix quantization tests and flaky tolerance (#4743) @laggui
• Fix fusion scalar broadcasting in write_output_aligned (#4741) @laggui
• Feat/implement fusion for irfft (#4736) @Sublime12
• Fix cubecl cuda all-reduce + remove useless check in distributed server (#4720) @Charles23R
• Feat/implement fusion for rfft (#4735) @Sublime12
• update cubek & fix gemv autotune (#4726) @louisfd
• Add more checks for quantized tensor reshape (#4704) @laggui
• feat: support cross-kind tensor casting via .cast() (#4713) @antimora
• chore: Fix some clippy errors, fix quant tests (#4708) @wingertge
• Update cubek (#4714) @louisfd
• Feat/add irfft (#4719) @Sublime12
• Feat/add rfft (#4707) @Sublime12
• Make Param Sync for parallel model inference (#4701) @antimora
• Perf/burn fusion overhead (#4645) @nathanielsimard
• Split TrainingStrategy to decouple the DistributedBackend requirement (#4710) @laggui
• fix: use integer arithmetic for nearest-neighbor coordinate scaling (#4687) @wkrettek
• All reduce in backward (#4650) @Charles23R
• fix output in attention tuner (#4702) @louisfd
• Fix attention_fallback NaN for fully-masked rows (#4697) @antimora
• Add HammingWindow operator to burn-tensor (#4698) @RunjiaChen
• update cubek and cubecl (#4699) @louisfd
• Fix fusion consistency checks and binding estimation in burn-cubecl-fusion (#4695) @nathanielsimard
• Update cubek and fix vecmat autotune (#4682) @louisfd
• Ignore local tests with pre-trained weights (#4676) @laggui
• Fix dispatch when only wgpu is enabled (maps to webgpu) (#4678) @laggui
• update cubek (#4677) @louisfd
• Fix fusion kernel vector_size mismatch on f16 output writes (#4675) @AdrianEddy
• Include new vec2mat routine in matmul autotune (#4673) @louisfd
• Update cubecl & cubek revs (#4672) @laggui
• feat: add categorical sampling for tensors (#4655) @majiayu000
• chore: Update to upstream changes in cubecl (#4670) @wingertge
• Refactor backend tests to set device settings at initialization + use Dispatch (#4666) @laggui
• Add HannWindow operator to burn-tensor (#4631) @walkinggo
• fixup:(burn-ndarray) fix comment and tidy imports (#4668) @TsaoLun
• Fix tch int_zeros dtype in sync (#4664) @laggui
• [Breaking] Use device settings to provide output dtype (#4653) @laggui
• feat: add FID vision metric (#4644) @cong-or
• Add Adan optimizer implementation with tests (#4651) @sepcnt
• [Breaking] Add bool store dtype + remove bool elem from fusion (#4649) @laggui
• Selector/attention (#4648) @louisfd
• fix(burn-ndarray): use owned storage for native heap allocations in from_data (#4647) @TsaoLun
• add utilities fn to FusionServer (#4640) @Charles23R
• Remove int powf and make powi numeric op (#4646) @laggui
• refactor: View launch (#4639) @wingertge
• chore: Update to cubecl changes (#4630) @wingertge
• Dispatch autodiff checkpointing strategy support (#4629) @laggui
• Implement RNNT loss (#4623) @cong-or
• Remove named tensor (#4628) @laggui
• Perf: Improve fusion score (#4511) @nathanielsimard
• refactor: Vector size generic (#4624) @wingertge
• Fix function arg name inconsistencies (#4626) @softmaximalist
• Update building-blocks chapter (#4625) @softmaximalist
• Refactor/device handle (#4593) @nathanielsimard
• feat: Introduce Lanczos3 interpolation method (#4601) @ovr
• Add Gram Matrix Loss for vision tasks (#4595) @softmaximalist
• Fix fusion cumulative op inputs (#4621) @laggui
• fix: replace ValidStep with InferenceStep in training.md (#4620) @TsaoLun
• Update documentation link for burn-store (#4619) @softmaximalist
• Improve module derive + add #[module(skip)] attribute (#4618) @laggui
• Add HalfPrecisionAdapter for F32/F16 mixed-precision storage (#4594) @antimora
• Fix cosine scheduler record in composed scheduler (#4617) @laggui
• Update ONNX import docs for LoadStrategy and from_bytes (#4607) @antimora
• Use shape in TensorData (#4603) @laggui
• Update SSIM float types to f32 (#4602) @softmaximalist
• Fix conv2d_weight_backward w/ strided channels and unit spatial dims (via conv_im2col_1x1) (#4591) @laggui
• Add multi-scale SSIM for image quality assessment (#4555) @softmaximalist
• Remove Clone bound from WindowsDataset item (#4597) @laggui
• Add contributing guidelines with AI-assisted contributions policy (#4569) @antimora
• feat: Implements DISTS metric (#4574) @koreaygj
• Fix dispatch autodiff feature propagation (#4592) @laggui

Full Changelog: v0.21.0-pre.2...v0.21.0-pre.3