v0.13.0 - `dtypes` module & adds `AMP<F>` dtype

What's Changed

• Make storage_traits::TensorToArray pub by @AndrejOrsula in #817
• accurate-gelu by @jcrist1 in #813
• Fixing examples/04-gradients.rs by @coreylowman in #824
• Moving optim kernels to tensor ops by @coreylowman in #828
• [Breaking] Adds AMP<F> dtype by @coreylowman in #811
• Adding documentation to dtypes module and amp by @coreylowman in #834

New Contributors

• @AndrejOrsula made their first contribution in #817
• @jcrist1 made their first contribution in #813

Full Changelog: v0.12.1...v0.13.0

v0.12.1 - Re-export f16 dtype & making more apis public

What's Changed

• Allow models to be backward compatible through #799 by @nkoppel in #808
• Various small fixes by @nkoppel in #814
• Making all shape traits public by @coreylowman in #816

Full Changelog: v0.12.0...v0.12.1

v0.12.0 - Adds f16 dtype

Breaking changes

• [Breaking] Adding Tensor::try_realize, and Tensor::realize no longer returns Result by @coreylowman in #758
• [Breaking] ReshapeTo::reshape_like and ReshapeTo::try_reshape_like now panic instead of returning option by @coreylowman in #766
• [Breaking] Adding dilation/groups to Conv2D. Adding dilation to Pool2D by @coreylowman in #767
• [Breaking] Use gemm for matmul. Removes support for matrixmultiply & MKL by @coreylowman in #776
• [Breaking] Moving storage GAT to trait level generic. Split DeviceStorage into multiple traits by @coreylowman in #782
• [Breaking] Adding dilation/groups to ConvTranspose2D by @coreylowman in #783

What's Changed

• Adding f16 as Dtype by @coreylowman in #696
• Adding example by @sirandreww in #740
• Adds TryConcatAlong to support Concat along any axis by @coreylowman in #750
• Changed CUDA_ARCH in compatibility.cuh by @jafioti in #752
• Allow broadcast_like to accept tensors OR shapes by @VasanthakumarV in #751
• Removing rerun build.rs for output destination by @coreylowman in #754
• Fixing compatibility for compute cap 70-75 by @coreylowman in #757
• Adds TriangleTensor and CmpKernel traits to Device bound by @coreylowman in #760
• Using Bernoulli distribution in dropout - makes dropout reproducible across dtypes by @coreylowman in #761
• Fixes bug with f16 mean where number of elements reduced was f16::INF by @coreylowman in #763
• Placeholder f16 gemm speedups by @coreylowman in #765
• MultiHeadAttention 3d impl now broadcasts to 4d instead of duplicating logic by @coreylowman in #768
• Moving cudarc?/f16 behind f16 feature by @coreylowman in #774
• impl Clone for Adam, SGD, RMSprop by @coreylowman in #775
• Properly setting read_dst for gemm in forward/backward pass by @coreylowman in #777
• Adds rayon dependency. Using gemm::Parallelism::Rayon(rayon::current_num_threads()) by @coreylowman in #778
• Add LogSoftmax by @kurnevsky in #769
• Moving some tests off nightly. Adding docs to conv2d op by @coreylowman in #779
• Adding better error messages if nvidia-smi/nvcc are not found by @coreylowman in #784
• Using for loop with gridDim.x * blockDim.x as increment by @coreylowman in #787
• Removing __hmax and __hmin compat functions by @coreylowman in #788
• Uses grid striding in fill_with by @coreylowman in #790
• Exposed NumpyDType publicly by @jafioti in #791
• Fixing weight shape for grouped Conv2D by @coreylowman in #797
• Bump half/cudarc versions by @coreylowman in #805
• Using Groups in conv weight init by @coreylowman in #806
• Add scalar support to TensorCollection by @nkoppel in #799

New Contributors

• @sirandreww made their first contribution in #740
• @kurnevsky made their first contribution in #769

Full Changelog: v0.11.2...v0.12.0

v0.11.2 - Tensor caching & other nice features

What's Changed

• Simplify upscale cuda kernels by @coreylowman in #680
• JIT compiling stack/concat cuda kernels by @coreylowman in #684
• Initial merging of nvidia-smi and nvcc checks by @quietlychris in #685
• feat: use Cow when appropriate by @Alexandcoats in #688
• Add const generic NUM_THREADS arg to launch_cfg by @VasanthakumarV in #691
• feat: add Tensorlike to clean up spooky ghosts by @Alexandcoats in #689
• Add contiguous and try_contiguous methods by @VasanthakumarV in #690
• (feat) add device access method by @ccaven in #692
• Add examples of runtime dimensions to examples/02-ops.rs by @VasanthakumarV in #698
• Prevent over-allocation for broadcasted outputs of sum_to by @nkoppel in #699
• Adds caching layer to tensor allocations by @coreylowman in #670
• Handle \r in build.rs by @ViliamVadocz in #702
• Disabling cache by default & adds enable_cache() by @coreylowman in #704
• Typos in feature_flags.rs by @mauvray in #710
• Adds mat * vec impl for matmul by @coreylowman in #716
• Adds better assertion macros for testing by @coreylowman in #714
• Combining multiple github workflows for reuse by @coreylowman in #717
• Changing nn ToDtype to use generic on method by @coreylowman in #719
• Flatten2D now accepts generic batch dim by @coreylowman in #720
• Uses impl Into<E> for scalar binary ops when possible by @coreylowman in #722
• Adds cudnn section to feature flags by @coreylowman in #723
• Impls for (T,) by @opfromthestart in #725
• Fixing dependencies for no-std by @coreylowman in #736
• Adds rust 1.65 as the minimum rust compiler version by @coreylowman in #737
• Moves scalar comparison to use the same method as tensor comparison. Deprecates try_scalar_*/scalar_*. by @coreylowman in #738
• Run CI for all kind of pushes - not only pull request related ones by @YannickFricke in #739
• Adds Tensor::to_device to support sending tensors of any shape to any device by @coreylowman in #741

New Contributors

• @VasanthakumarV made their first contribution in #691
• @ccaven made their first contribution in #692
• @mauvray made their first contribution in #710
• @YannickFricke made their first contribution in #739

Full Changelog: v0.11.1...v0.11.2

v0.11.1 - cudnn, optimizations, and new ops/nn layers

What's Changed

• Fix bug in gather cuda kernel by @nkoppel in #588
• feat(device): introduce AutoDevice type by @kakoc in #579
• Use Recursive Macros to Implement Shape Operation Traits. by @nkoppel in #583
• Add ToDtype tensor operation by @nkoppel in #582
• Using 128 threads by default for cuda kernels by @coreylowman in #599
• Add Slice tensor operation. by @nkoppel in #602
• Optimizing conv kernels a bit by @coreylowman in #605
• feat: add upper/lower triangles (tril and triu) allocations by @Alexandcoats in #568
• Adds Tensor::roll by @coreylowman in #608
• Using multiple streams for matmul with cuda by @coreylowman in #610
• Fix no-std support by @Alexandcoats in #615
• Adds matrixmultiply/std to std feature by @kstavro in #618
• Implement concat for usize arrays; add concat to Device. by @nkoppel in #621
• Allow conv2d and pool2d to use dynamic dimensions for width and height. by @nkoppel in #620
• Switch to using nvcc --list-gpu-code for build.rs compute_cap by @quietlychris in #619
• Fix bug in reshape on cuda by @nkoppel in #622
• Don't always do try_min in pool_global.rs by @nkoppel in #623
• Revert "Switch to using nvcc --list-gpu-code for build.rs compute_cap… by @coreylowman in #624
• Adds restrided in favor of get_unstrided_index -> get_strided_index by @coreylowman in #628
• Combines multiple calls to get_strided_index into a single loop by @coreylowman in #629
• Reducing number of buffers sent to cuda for some operations by @coreylowman in #611
• Optimizing conv2d more by @coreylowman in #631
• Add ability to include smaller last batch by @nkoppel in #632
• Upscale2D and ConvTrans2d by @opfromthestart in #603
• impl Dtype for all Unit types except bool by @coreylowman in #635
• Allow convtrans2d to use dynamic dimensions by @nkoppel in #639
• JIT compiling kernel for to_dtype & reshape by @coreylowman in #634
• Optimize conv transpose kernels to do same thing as conv by @coreylowman in #641
• Reworking crate level documentation by @coreylowman in #644
• Adds synchronize to DeviceStorage by @coreylowman in #645
• adding usize dtype to cuda_kernel by @zojeda in #648
• Add PReLU and LeakyReLU by @opfromthestart in #586
• Moving logsumexp normalization off of graph by @coreylowman in #652
• Adding CmpKernels to Device, more documentation by @coreylowman in #653
• Removing bounds checking from cpu conv kernel folding by @coreylowman in #650
• Allow upscale2d to use dynamic dimensions by @nkoppel in #654
• Adding integration test for resnet18 by @coreylowman in #655
• Removing some un-necessary blanket impls by @coreylowman in #656
• Fixes conv transpose stride bug, adds more docs to upscale2d by @coreylowman in #658
• Some QOL fixes by @opfromthestart in #659
• Optimizing softmax & log_softmax by @coreylowman in #660
• Reuse f(x) for unary operations when possible. by @coreylowman in #661
• Allocating gradients in backward op by @coreylowman in #663
• Adds Tensor::recip (1 / x) by @coreylowman in #665
• Reshape layer by @opfromthestart in #666
• Re-using tensor storage when possible by @coreylowman in #664
• Adds cudnn feature flag. Removes "test-cuda" feature flag. Using cuDNN for convolutions by @coreylowman in #651
• Always attempting allocation reuse during inference by @coreylowman in #673
• Clarify reshape behavior in docs by @coreylowman in #674
• Have SplitInto keep tapes of each head seperate by @nkoppel in #676
• Using arch option in nvrtc by @coreylowman in #675

New Contributors

• @kakoc made their first contribution in #579
• @quietlychris made their first contribution in #619
• @opfromthestart made their first contribution in #603
• @zojeda made their first contribution in #648

Full Changelog: v0.11.0...v0.11.1

v0.11.0 - Cuda support, mixed const/runtime tensors, and device rewrite

What's Changed

• AddInto by @Dimev in #256
• added 5d & 6d tensors by @M1ngXU in #283
• Remove phantom by @M1ngXU in #282
• remove tensor bound by @Dimev in #297
• Adding nightly to cargo-test by @JYudelson1 in #294
• Devices/Dyn dimensions refactor by @coreylowman in #304
• Add instructions for running the mnist example. by @infalmo in #310
• Removes Dyn. Use usize directly by @coreylowman in #315
• Making f32 default dtype for Tensor, updating examples/docstrings by @coreylowman in #316
• Only running gha on push by @coreylowman in #317
• Adding Unit and HasUnitType. Reducing bounds for Dtype by @coreylowman in #313
• Removing build_test_device. Using TestDevice everywhere by @coreylowman in #324
• Adding SampleTensor, Removing RandTensor/RandnTensor by @coreylowman in #327
• Removing usages of tensor aliases by @coreylowman in #328
• Moving intel-mkl stuff into sub module in build.rs by @coreylowman in #329
• Adding Cuda device and skeleton cuda kernel impls by @coreylowman in #322
• Implementing abs/exp/div/sum_to cuda kernels by @coreylowman in #331
• permute_to and broadcast_to cuda kernels by @coreylowman in #343
• Add cuda implementations for unary and binary tensor operations in #341 and #334 by @nkoppel in #346
• Using atomicAdd in binary op backwards to properly handle strides by @coreylowman in #350
• Resolve #352 and #347 by @nkoppel in #354
• Implement reshape cuda kernel (resolves #336) by @nkoppel in #356
• Add missing device generic in transformer test by @ViliamVadocz in #358
• Add select and gather cuda kernels. by @nkoppel in #359
• Upgrade to cudarc 0.6.0 by @coreylowman in #361
• Add tests for binary broadcasted add and fix bugs to allow them to pass. by @nkoppel in #357
• run GHA on pull_request by @coreylowman in #364
• matmul cuda kernels by @coreylowman in #342
• Adding dynamic example. by @Narsil in #368
• Add cuda kernels for min_to/max_to by @coreylowman in #370
• Adding dropout cuda kernel by @coreylowman in #372
• Adding ConstDim and ConstShape for tensor creation by @coreylowman in #373
• Fixing computation of lda/ldb/ldc with cblas by @coreylowman in #375
• Modify sum_to cuda kernel to not need atomic adds in backwards by @nkoppel in #367
• Simplifying trait Conv2DKernel and Cpu implementation by @coreylowman in #376
• (#344) Implement cuda kernels for optimizers by @nkoppel in #378
• Fix max_to and min_to edge case with negative zero by @ViliamVadocz in #380
• Add cuda kernels for conv2d by @coreylowman in #369
• Rework pool2d internals & add pool2d cuda kernels by @coreylowman in #384
• Implement Shape for arrays (#377) by @nkoppel in #385
• Efficient cuda kernels for reductions by @nkoppel in #382
• Improving compilation times of deeply nested const generic modules by @coreylowman in #391
• Fixing remainder of cuda tests & fixing cblas/cublas matmul with strides [1,1] by @coreylowman in #393
• Adding Cuda device usage to mnist example by @coreylowman in #396
• Adding GeLU operator (used in Gpt2) by @Narsil in #397
• Removing codecov from workflows/readme by @coreylowman in #403
• Reorganize tensor_ops, and add cuda_utils.cuh by @nkoppel in #398
• Some small optimizations for conv2d on cpu by @coreylowman in #404
• Removing Device generic from Gradients & optimizers by @coreylowman in #402
• Add ToDevice and OnDevice to simplify nn api (#388) by @nkoppel in #394
• Removes ModuleBuilder, Adds BuildModule & BuildOnDevice by @coreylowman in #405
• Enable multi-core matmul by @infalmo in #417
• Fix GELU CUDA kernel compilation by @ViliamVadocz in #409
• Adding nn.Embedding layer. by @Narsil in #406
• Removing defaults for Tensor Dtype & Device generic parameters by @coreylowman in #418
• Removing Default for optimizers & adding &M to constructors by @coreylowman in #422
• Adding runtime assertion in try_binary_op that shapes are equal by @coreylowman in #428
• Add boolean operations and choose. by @nkoppel in #415
• Add TensorFrom trait to create tensors from both vectors and arrays. by @nkoppel in #414
• Adding nn builder structs, dtype generics, and remove device defaults. by @coreylowman in #433
• Upgrade to cudarc==0.7.0 and use alloc_async instead of alloc_zeros_async by @coreylowman in #440
• Add comparison tensor operations by @ViliamVadocz in #386
• Add synchronize method to Cuda device by @ViliamVadocz in #442
• f64 kernels by @coreylowman in #421
• Add stack tensors method by @coreylowman in #449
• cargo check cuda & run f64 tests in CI by @coreylowman in #447
• Fix bug in #451 by @nkoppel in #453
• Add more runtime shape checks by @coreylowman in #454
• Adding ReshapeTo::reshape_like by @coreylowman in #456
• Adding SampleTensor::sample_uniform_like and SampleTensor::sample_normal_like by @coreylowman in #457
• Improve examples (add Cuda) by @TimerErTim in #452
• Dataset iterators - adds batching, collating for iterators by @coreylowman in #462
• Fixing issue with to_device and broadcasted tensors by @coreylowman in #465
• Bump cudarc 0.7.2 by @coreylowman in #466
• Adding index out of bounds checks to select/gather kernels by @coreylowman in #467
• Rename to add_dim. by @infalmo in #471
• impl BuildModule for ZeroSizedModule by @coreylowman in #470
• Adds TensorCollection by @coreylowman in #469
• Fixing cargo doc warnings by @coreylowman in #473
• Using --gpu-architecture native with nvcc by @coreylowman in #474
• using TensorFromVec for OneHotEncode and Arange by @coreylowman in #477
• Small batchnorm optimizations by @coreylowman in #478
• nvcc: fixed type bug by @M1ngXU in #480
• Adds fast_alloc feature and binary kernel optimizations by @coreylowman in #481
• Adding some "benchmarking" scripts by @coreylowman in #483
• Add try_forward and try_forward_mut to Module and ModuleMut. by @nkoppel in #482
• Optimizing cpu kernels for reductions by @coreylowman in #484
• Using alloc_zeros_async and memset_zeros for cuda by @coreylowman in #489
• Making Conv2D unbiased by default, and adding Bias2D module by @coreylowman in #494
• Using image/filter stride in cuda kernel for conv by @coreylowman in #495
• bump cudarc version by @coreylowman in #498
• Adding attention_reshape (inference only) kernels. by @Narsil in #497
• Adding lifetime to gat in Exact...

v0.10.0

What's Changed

Breaking Changes

• Binary ops (add, sub, div, mul, maximum, minimum) take ownership of rhs by @coreylowman in #268
• backwards only allows 0d tensors now by @coreylowman in #206
• Clone now keeps same id, removing Tensor::duplicate by @coreylowman in #249
• Multi axis reductions
  • See docs
  • #189, #190, #194
  • Reduction functions now can reduce across any axis/axes: mean, sum, max, min, stddev, var, softmax, log_softmax, and logsumexp
  • Remove -1 from valid axes, add trait HasLastAxis to use in generic functions instead
  • Adding normalize function that normalizes across any axis
  • Removing single axis reduction functions fn *_axis(): mean_axis, sum_axis, max_axis, min_axis, normalize_axis, std_axis, var_axis
  • Rename HasAxis to HasAxes
  • Add trait BroadcastTo
    • Remove trait Broadcast1, trait Broadcast2, trait Broadcast3, trait Broadcast4
  • Add trait Reduce/trait ReduceTo
    • Remove trait Reduce1
• Batched select & select consistency
  • See docs
  • Renaming SelectTo, using SelectTo for batched select by @coreylowman in #217
  • Add Batched Select for devices and tensor_ops by @coreylowman in #182
• Reduce things in prelude by @coreylowman in #209
• Renaming FlattenImage to Flatten2D by @coreylowman in #243

New features

• Arc in Tensors instead of Rc by @caelunshun in #236
• powi() and powf() functions by @coreylowman in #167
• no_std support
  • See feature flags docs
  • Remove num-traits, no default features on depends by @coreylowman in #200
  • Adding intel-mkl feature and removing the 4 mkl-- features by @coreylowman in #239
  • Adding module that has docs for feature flags by @coreylowman in #240
  • Adding "numpy" feature to make numpy & npz optional by @coreylowman in #241
  • Adding #![no_std] support via no_std_compat by @coreylowman in #244
  • Adding default-features = false to dependencies by @coreylowman in #257
• Adding Axis permutations via trait PermuteTo.
  • See docs
  • #208, #169, #193
• Adding trait ModuleMut
  • See docs
  • #225
  • Removing Module super traits by @coreylowman in #223
  • Rework Dropout/DropoutOneIn to use ModuleMut by @coreylowman in #226
• Adding decoupled/l2 weight decay in optimizers:
  • See docs
  • add HasArrayData to GradientProvider by @cBournhonesque in #261
  • Add weight decay to SGD by @cBournhonesque in #258
  • Adding weight_decay to Adam by @coreylowman in #275
  • Adding weight decay to RMSprop by @coreylowman in #276
• Adding nn::Transformer #175, #173, #180
  • See docs
• Adding nn::MinPool2D, nn::MaxPool2D, nn::AvgPool2D by @coreylowman in #214
  • See docs
• Adding nn::MinPoolGlobal, nn::MaxPoolGlobal, nn::AvgPoolGlobal by @coreylowman in #216
  • See docs
• Adding nn::BatchNorm2D by @coreylowman in #228
  • See docs

Misc changes

• Add tensor() function as a convenient way to make tensors from arrays by @coreylowman in #161
  • See docs
• Remove allocation in dropout implementation by @coreylowman in #164
• Removing Tensor::OwnedTape by @coreylowman in #197
• Revamping examples/ by @coreylowman in #205
• Conv cleanup
  • Moving conv into device and cleaning up a bit by @coreylowman in #212
  • Minifying conv impls by @coreylowman in #213
  • Changing conv2d and conv2d_batched to methods of tensors by @coreylowman in #221
  • Replacing conv2d implementation with matmuls by @coreylowman in #237
• Fix typos by @cBournhonesque in #235
• Combining multiple where clauses with const generics into a single one by @coreylowman in #264
• Checking for null ptr in AllocateZeros by @coreylowman in #271
• Reducing allocations in map_df_uses_fx by @coreylowman in #272
• Adding with_empty_tape and with_diff_tape by @coreylowman in #274

New Contributors

• @cBournhonesque made their first contribution in #235
• @caelunshun made their first contribution in #236

Full Changelog: v0.9.0...v0.10.0

v0.9.0

Breaking Changes

• Add broadcast functions, reductions on any axis, and selecting subtensors (#137, #114, #139) by @coreylowman in #138
• Added normalize axis and removed normalize by @vikigenius in #140
• #67 Optimizer::update now returns Result<(), UnusedParamsError> by @coreylowman in #107

New features

• #34 Add Transformers!!! by @jafioti in #120
• #1 Add Conv2d by @coreylowman in #124
• #55 Added reshape function by #90 #129 #120
• #133 Adding FlattenImage layer that uses reshape by @coreylowman in #133
• #142 Adding Module::forward_mut by @coreylowman in #148
• #80 Adding nn::Softmax by @coreylowman in #81
• #79 Adding smooth_l1_loss and huber_loss by @coreylowman in #82
• #131 matmul now supports batched & broadcasted inputs by @coreylowman in #132
• add macOS MKL support by @yerke in #73
• Adding maximum function by @coreylowman in #143
• Adding min_axis function by @coreylowman in #144

Additional changes

• Simplifying implementation of BCE loss using binary_map by @coreylowman in #75
• Miscellaneous updates by @coreylowman in #76
• Added custom model example by @jafioti in #83
• add Debug and Display support for NpzError by @XBagon in #85
• Added nightly feature by @jafioti in #89
• Added 2d broadcast_first functions and 3d linear forward by @jafioti in #94
• #55 reshape, and #87 additional work on nightly feature by @coreylowman in #90
• #69 adding map_df_uses_fx by @coreylowman in #105
• Fixed a misleading docstring. by @M1ngXU in #109
• Fix Issue #110 Fix (Dropout (test) for non-positive values) by @M1ngXU in #113
• Issue #96 by @M1ngXU in #118

New Contributors

• @jafioti made their first contribution in #83
• @XBagon made their first contribution in #85
• @yerke made their first contribution in #73
• @M1ngXU made their first contribution in #109
• @vikigenius made their first contribution in #140

Full Changelog: v0.8.0...v0.9.0