Skip to content

Releases: tracel-ai/burn

v0.20.1

23 Jan 17:43
@laggui laggui
v0.20.1
75b7881

Choose a tag to compare

View all tags
v0.20.1 Latest
Latest

Bug Fixes & Improvement

Contributors

laggui
Assets 2
Loading

v0.20.0

15 Jan 16:08
@laggui laggui
v0.20.0
3475ba8

Choose a tag to compare

View all tags
v0.20.0

Summary

This release marks a major turning point for the ecosystem with the introduction of CubeK. Our goal was to solve a classic challenge in deep learning: achieving peak performance on diverse hardware without maintaining fragmented codebases.

By unifying CPU and GPU kernels through CubeCL, we've managed to squeeze maximum efficiency out of everything from NVIDIA Blackwell GPUs to standard consumer CPUs.

Beyond performance, this release makes the library more robust, flexible, and significantly easier to debug.

This release also features a complete overhaul of the ONNX import system, providing broader support for a wide range of ONNX models. In addition, various bug fixes and new tensor operations enhance stability and usability.

For more details, check out the release post on our website.

Changelog

Breaking

We've introduced a couple of breaking API changes with this release. The affected interfaces are detailed in the sections below.

Training

We refactored burn-train to better support different abstractions and custom training strategies. As part of this,
the LearnerBuilder has been replaced by the LearningParadigm flow:

- let learner = LearnerBuilder::new(ARTIFACT_DIR)
+ let training = SupervisedTraining::new(ARTIFACT_DIR, dataloader_train, dataloader_valid)
        .metrics((AccuracyMetric::new(), LossMetric::new()))
        .num_epochs(config.num_epochs)
-       .learning_strategy(burn::train::LearningStrategy::SingleDevice(device))
-       .build(model, config.optimizer.init(), lr_scheduler.init().unwrap());
+       .summary();
 
- let result = learner.fit(dataloader_train, dataloader_valid);
+ let result = training.launch(Learner::new(
+      model,
+      config.optimizer.init(),
+      lr_scheduler.init().unwrap(),
+ ));

Interface Changes

The scatter and select_assign operations now require an IndexingUpdateOp to specify the update behavior.

- let output = tensor.scatter(0, indices, values);
+ let output = tensor.scatter(0, indices, values, IndexingUpdateOp::Add);

API calls for slice, slice_assign, and slice_fill no longer require const generics for dimensions, which cleans up the syntax quite a bit:

- let prev_slice = tensor.slice::<[Range<usize>; D]>(slices.try_into().unwrap());
+ let prev_slice = tensor.slice(slices.as_slice());

The grid_sample_2d operation now supports different options.
To preserve the previous behavior, make sure to specify the matching options:

- let output = tensor.grid_sample_2d(grid, InterpolateMode::Bilinear);
+ let options = GridSampleOptions::new(InterpolateMode::Bilinear)
+     .with_padding_mode(GridSamplePaddingMode::Border)
+     .with_align_corners(true);
+ let output = tensor.grid_sample_2d(grid, options);

The QuantStore variants used in QuantScheme have been updated to support a packing dimension.

  pub enum QuantStore {
      /// Native quantization doesn't require packing and unpacking.
      Native,
+     /// Store packed quantized values in a natively supported packing format (i.e. e2m1x2).
+     PackedNative(usize),
      /// Store packed quantized values in a 4-byte unsigned integer.
-     U32,
+     PackedU32(usize),
 }

Finally, Shape no longer implements IntoIterator. If you need to iterate by-value over dimensions, access the dims field directly.

- for s in shape {
+ for s in shape.dims {

Module & Tensor

Datasets & Training

Backends

Bug Fixes

Documentation & Examples

Fixes

ONNX Support

Enhancements

Refactoring

  • chore: Update to batch caching PR for cubecl (#3948) @wingertge
  • Refactor IR to define outputs as a function of the operation (#3877) ...
Read more

Contributors

khoek, antimora, and 19 other contributors
Loading

v0.20.0-pre.6

18 Dec 21:27
@nathanielsimard nathanielsimard
v0.20.0-pre.6
91dd62c
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.20.0-pre.6 Pre-release
Pre-release

What's Changed

Contributors

antimora, wingertge, and 9 other contributors
Loading

v0.20.0-pre.5

08 Dec 14:53
@nathanielsimard nathanielsimard
v0.20.0-pre.5
42edc63
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.20.0-pre.5 Pre-release
Pre-release

What's Changed

Contributors

antimora, wingertge, and 6 other contributors
Loading

v0.20.0-pre.4

01 Dec 19:15
@nathanielsimard nathanielsimard
v0.20.0-pre.4
c9af669

Choose a tag to compare

View all tags
v0.20.0-pre.4 Pre-release
Pre-release

What's Changed

Contributors

khoek, antimora, and 7 other contributors
Loading

v0.20.0-pre.3

24 Nov 17:37
@nathanielsimard nathanielsimard
v0.20.0-pre.3
88d662d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.20.0-pre.3 Pre-release
Pre-release

What's Changed

Contributors

antimora, laggui, and 6 other contributors
Loading

v0.20.0-pre.2

17 Nov 15:26
@nathanielsimard nathanielsimard
v0.20.0-pre.2
cc0f22a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.20.0-pre.2 Pre-release
Pre-release

What's Changed

Contributors

antimora, wingertge, and 5 other contributors
Loading

v0.20.0-pre.1

11 Nov 15:35
@nathanielsimard nathanielsimard
v0.20.0-pre.1
913ddc0

Choose a tag to compare

View all tags
v0.20.0-pre.1 Pre-release
Pre-release

Summary

This release includes significant performance improvements, bug fixes, and architectural refactoring.
Key Improvements:

  • TMA autotuning and MMA matmul tuning enabled for better performance
  • ONNX-IR refactored to an op/node-centric architecture IR refactored to define outputs as a function of the operation

Bug Fixes:

  • Fixed autodiff graph cleanup issues (multiple fixes for deferred/consumed nodes)
  • Fixed Linear layer panic when output size is one
  • Fixed PyTorch pickle reader regression with integer dict keys
  • Fixed RoPE sum_dim calculation
  • Fixed tensor *_like dtype preservation
  • Fixed squeeze check for D2 > 0
  • Fixed QLinear implementation
  • Fixed async barrier & TMA checks

New Features:

  • Added matvec operation
  • Added support for custom learning strategies
  • Added Candle device seeding
  • Added Shape::ravel_index for row-major raveling
  • Generalized linalg::outer semantics with new linalg::outer_dim
  • Implemented error handling for DataError
  • Added square() optimization where appropriate
Loading

v0.19.1

06 Nov 16:18
@laggui laggui
v0.19.1
a6da424

Choose a tag to compare

View all tags
v0.19.1

Bug Fixes & Improvements

Contributors

wingertge, laggui, and 3 other contributors
Loading

v0.19.0

28 Oct 17:00
@laggui laggui
v0.19.0
0767a9a

Choose a tag to compare

View all tags
v0.19.0

Summary

This release brings major improvements to enable efficient distributed training, quantization, and CPU support in Burn.

To achieve true multi-GPU parallelism, we had to rethink several core systems: we implemented multi-stream execution to keep all GPUs busy, optimized device transfers to avoid unnecessary synchronization, and redesigned our locking strategies to eliminate bottlenecks in autotuning, fusion, and autodiff. We also introduced burn-collective for gradient synchronization and refactored our training loop to support different distributed training strategies.

Additionally, we added comprehensive quantization support, allowing models to use significantly less memory while maintaining performance through fused dequantization and optimized quantized operations.

Finally, we introduced a new CPU backend powered by MLIR and LLVM, bringing the same JIT compilation, autotuning, and fusion capabilities from our GPU backends to CPU execution.

As with previous releases, this version includes various bug fixes, further optimizations and enhanced documentation. Support for ONNX models has also been expanded, with additional operators and bug fixes for better operator coverage.

For more details, check out the release post on our website.

Changelog

Breaking

We've introduced a couple of breaking API changes with this release. The affected interfaces are detailed in the sections below.

Learning Strategy

We refactored the Learner to support better distributed training strategies. Instead of registering a list of device(s), you now specify a training strategy.

  let learner = LearnerBuilder::new(artifact_dir)
      .metric_train_numeric(AccuracyMetric::new())
      .metric_valid_numeric(AccuracyMetric::new())
      .metric_train_numeric(LossMetric::new())
      .metric_valid_numeric(LossMetric::new())
      .with_file_checkpointer(CompactRecorder::new())
-     .devices(vec![device.clone()])
+     .learning_strategy(LearningStrategy::SingleDevice(device.clone()))
      .num_epochs(config.num_epochs)
      .summary()
      .build(
          config.model.init::<B>(&device),
          config.optimizer.init(),
          config.learning_rate,
      );

Learner Training Result

The Learner previously lacked an evaluation loop. We extended its return type to include all training states in a TrainingResult, which includes the trained model and a metrics renderer.

- let model_trained = learner.fit(dataloader_train, dataloader_valid);
+ let result = learner.fit(dataloader_train, dataloader_valid);

- model_trained
+ result
+    .model
     .save_file(format!("{artifact_dir}/model"), &CompactRecorder::new())
     .expect("Trained model should be saved successfully");

This enables the renderer to be reused by the new evaluator so that training and evaluation metrics appear together in the TUI dashboard:

let mut renderer = result.renderer;
let evaluator = EvaluatorBuilder::new(artifact_dir)
    .renderer(renderer)
    .metrics((AccuracyMetric::new(), LossMetric::new()))
    .build(result.model.clone());

evaluator.eval(name, dataloader_test);

Interface Changes

Config

The Config trait now requires Debug:

- #[derive(Config)]
+ #[derive(Config, Debug)]
  pub struct TrainingConfig {
      // ...
  }

BatchNorm

BatchNorm no longer requires the spatial dimension generic:

  #[derive(Module, Debug)]
  pub struct ConvBlock<B: Backend> {
      conv: nn::conv::Conv2d<B>,
-     norm: BatchNorm<B, 2>,
+     norm: BatchNorm<B>,
      pool: Option<MaxPool2d>,
      activation: nn::Relu,
  }

Backend::seed

Seeding is now device-specific:

- B::seed(seed);
+ B::seed(&device, seed);

Tensor

For consistency with other methods like unsqueeze() / unsqueeze_dim(dim), squeeze(dim) was renamed:

- tensor.squeeze(dim)
+ tensor.squeeze_dim(dim)

We've also added a tensor.squeeze() method which squeezes all singleton dimensions.

Finally, we removed tensor ^ T syntax, which was clunky.

- use burn::tensor::T;
- tensor ^ T
+ tensor.t()

tensor.t() is also a simple alias for tensor.transpose().

Module & Tensor

Datasets & Training

Backends

Bug Fixes

Read more

Contributors

antimora, wingertge, and 35 other contributors
Loading