Version 0.25.3

What's Changed

• @dshan4585 Prevent premature destructuring of closures & Add atan2 (#286)
• @Vlad-Shcherbina Fix not one but two leaks related to gradients (#296)
• @scttfrdmn Fix: Add missing Float64 pattern in safetensors conversion (#295)
• @Vlad-Shcherbina Add missing #[param] attributes to InstanceNorm (#300)

Version 0.25.1

Merge pull request #272 from oxideai/release/0.25.1

release: Version 0.25.1

Version 0.25.0

Update upstream dep to MLX 0.25.0

Version 0.23.0

What's Changed

• fix: add dilation and groups field to conv layers by @minghuaw in #204
• update: Update mlx-c to v0.1.2 (mlx v0.23.1) by @minghuaw in #210
• Exploration: use macro to mimic named args for optional args by @minghuaw in #157
• fix: Fix generated macro doc link by @minghuaw in #217
• feat: Separate grad into two functions by @minghuaw in #218
• docs: Port some of the python docs by @minghuaw in #212
• release: Version 0.23.0 by @minghuaw in #221
• Revert "release: Version 0.23.0" by @minghuaw in #222
• Update/version 0.23.0 by @minghuaw in #223

Full Changelog: v0.21.2...v0.23.0

Version 0.21.2

What's Changed

• fix: Fix wrong condition in smooth l1 loss by @minghuaw in #213

Full Changelog: v0.21.1...v0.21.2

Version 0.21.1

What's Changed

• feat(examples): add linear regression by @dcvz in #199
• fix(docs): Fix crates badge in README by @dcvz in #201
• chore: Use comparison requirement for mlx-sys version when used as dep by @minghuaw in #208

Full Changelog: v0.21.0...v0.21.1

Version 0.21.0

What's Changed

• reorganize workspace by @minghuaw in #96
• Setting up empty mlx-nn crate for future NN works by @minghuaw in #99
• Trust synchronization on the cpp side by @minghuaw in #102
• chore(example): add automatic_differentiation example by @dcvz in #103
• chore(mlx-c): update to 0.0.9 by @dcvz in #105
• Restrict use of ScalarOrArray trait to operator impl and ClipBound by @minghuaw in #108
• Allow using named shape parameter in array!() macro by @minghuaw in #109
• Use custom c shim for fallible closure by @minghuaw in #116
• Initial attempt at implementing mlx-nn NN modules by @minghuaw in #100
• Change GenarateBuilder and GenerateTestCases to function-like proc_macro by @minghuaw in #120
• Allow indexing with &Array by @minghuaw in #122
• feat(nn): embedding by @dcvz in #121
• Implement AdaGrad optimizer by @minghuaw in #125
• chore(mlx-c): Update MLX-C to 0.10.0 by @dcvz in #126
• Implement AdaDelta optimizer by @minghuaw in #127
• chore(mlx-c): Add remaining additions by @dcvz in #128
• feat(mlx-fast): Implement NN part of mlx-fast by @dcvz in #131
• Move losses and optimizers into mlx-rs by @minghuaw in #132
• feat(api): Add load ops by @dcvz in #133
• implement Adam, AdamW and Adamax optimizers by @minghuaw in #134
• Add remaining optimizers by @minghuaw in #135
• feat: Allow recursive freeze/unfreeze by @minghuaw in #143
• fix: Fix error with checking if parameter is frozen on empty module by @minghuaw in #144
• feat: Change Module::forward to take &mut self by @minghuaw in #145
• Revamp internal macro generate_builder!{} by @minghuaw in #150
• Add generic type Args in Module trait to allow binary modules by @minghuaw in #146
• update: Update to revamped c binding by @minghuaw in #147
• feat(nn): Add normalizations by @minghuaw in #138
• feat(nn): Add convolution_transpose mod by @minghuaw in #137
• feat(nn): Add transformer mod by @minghuaw in #151
• feat(nn): Add upsample mod by @minghuaw in #153
• feat(nn): Add recurrent module by @minghuaw in #149
• fix: Remove unnecessary logical ops in optimizer by @minghuaw in #163
• chore: Renamed value_and_grad_with_hashmap to keyed_value_and_grad by @minghuaw in #165
• chore: Add missing docs by @minghuaw in #167
• chore: Improve closure error message by @minghuaw in #164
• fix: Fix vec of module not listing all params by @minghuaw in #162
• feat(example): update mnist example by @minghuaw in #159
• feat(nn): Part 1 of quantization by @minghuaw in #142
• feat(nn): Another attempt at positional encoding by @minghuaw in #172
• chore: replace &Array with impl AsRef<Array> by @minghuaw in #170
• feat(nn): Add pooling by @minghuaw in #139
• refactor: move mlx-nn into mlx-rs by @minghuaw in #173
• feat: trait Quantizable and enum wrapper MaybeQuantized by @minghuaw in #174
• feat: compile with state by @minghuaw in #166
• feat: Add conversion between safetensors::tensor::TensorView and Array by @minghuaw in #176
• chore: bump and unify version by @minghuaw in #175
• chore: version 0.21.0-rc.1 by @minghuaw in #180
• chore: mlx-sys version 0.1.0-rc.1 by @minghuaw in #181
• chore: move mlx-internal-macros tests to mlx-tests by @minghuaw in #182
• chore: change root README to sym link to mlx-rs/README.md by @minghuaw in #183
• fix & example: mistral by @minghuaw in #178
• refactor: Use c-API io for loading/saving module parameters by @minghuaw in #189
• chore(docs): Update README by @dcvz in #191
• chore: bump up version to 0.21.0-rc.2 by @minghuaw in #192
• refactor: save/load optimizer state by @minghuaw in #190
• chore(readme): update MSRV and license badge by @dcvz in #195
• refactor: Remove prelude mod by @minghuaw in #196
• chore: Minor update to readme by @minghuaw in #197
• chore: Bump up version and updated changelog by @minghuaw in #198

Full Changelog: v0.14.0...v0.21.0

Version 0.14.0

Rust binding for mlx 0.14.0 based on mlx-c 0.0.8.