tracel-ai/burn · main · activity merged
1–31 August 2026 (UTC−4)
In August, Burn continued retiring older backends, directing users toward Flex for CPU computation and the CubeCL family for execution across hardware targets. The month also brought work on training speed, quantized fine-tuning, importing PyTorch weights, and saving models without damaging existing files.
A new capability records a model’s computation graph. Other changes isolate execution failures and ensure complete metrics are available before saving a model or stopping training. Seven developments below explain what changed and what it means for Burn users.
A backend determines how Burn runs computations on a device. Flex is a Rust CPU backend with multithreading and SIMD to process data in parallel. CubeCL underpins backends for CUDA, ROCm, WGPU and CPU, allowing the Burn API to target different devices.
August removed Candle and deprecated NdArray and LibTorch, recommending migration to Flex or CubeCL. Candle users need to change backend; NdArray and LibTorch continue to be built, tested and published during deprecation. This is a staged transition that retains support for some existing configurations.
Sources: #5343, #5344, #5525. Backend roles and transition details.
CubeCL gained a way to compute convolution weight gradients through matrix multiplication. It avoids an expensive calculation that effectively used an entire feature map as a convolution kernel. The new algorithm needs additional memory, so it participates in autotuning rather than being selected unconditionally.
The PR author reports that an EfficientNet V2-S training step at 384×384 fell from 1.235 to 1.085 seconds, about 12%, on a Radeon 8060S with Vulkan, f32 and batch size 4. That result applies to this configuration. Separately, Flex pooling operations, which reduce spatial dimensions, were reworked for more efficient CPU execution.
Sources: #5428, #5361. Algorithms, measurement conditions and limits.
A normal model run executes operations on data. The new capture backend instead records the requested operations, their connections and the initial values they use. The result is a computation graph: a description of the calculation that another tool can read.
Export tools can use this structure to represent a model in another format; capture does not perform that conversion itself. August follow-ups also preserve initialized values after a recording scope ends and allow those values to move between capture sessions.
Sources: #5377, #5408, #5409. How graph capture works.
QLoRA trains small additional matrices, or adapters, over frozen model weights stored in quantized form. Burn already offered this API; August’s changes fix problems in its execution.
Gradients now flow through the trainable adapters while leaving the quantized base frozen. Parameter dtype and allocation fixes keep adapter factors at the intended precision and alive across temporary memory-pool rebuilding. The shared parameter transformation mechanism was also generalized for methods beyond LoRA.
Sources: #5317, #5311, #5362, #5364, #5365. QLoRA changes and tested scenarios.
BurnpackStore now obtains and writes tensor data one tensor at a time. Host memory for these payloads is bounded by the largest tensor instead of the whole model. Saving through a temporary file also preserves the previous file if obtaining a tensor’s data fails.
PyTorch weight loading now respects tensor strides, including views whose elements are not stored consecutively. Missing or truncated storage returns an error instead of silently filling values with zeros. Additional TensorData size validation rejects malformed deserialized data and protects NdArray construction.
Sources: #5349, #5392, #5360. Saving and loading correctness.
Fusion combines operations for efficient execution. Previously, a panic during an operation could corrupt queue bookkeeping and cause later computations to fail even when they did not depend on the failed operation.
Errors now belong to affected tensors. Dependent operations are skipped with the root cause preserved; independent work can continue. Explicit error-returning data readback APIs also give callers a way to distinguish successful results from data that was never computed.
Sources: #5535, #5316. Failure handling and its boundaries.
Burn processes training events asynchronously. Previously, model saving and early stopping could query metrics before their computation finished. Those decisions now wait for already-submitted events to be processed, so completed epoch metrics are available.
Dice, a segmentation metric, now aggregates raw statistics across batches rather than averaging individual ratios. ROUGE-L, used in text evaluation, now returns its final score instead of panicking. These fixes affect how model quality is evaluated and which training state is saved.
Sources: #5320, #5515, #5517. Metric calculation and regression tests.
User migration and model export. NdArray and LibTorch remain supported; migration cannot be treated as complete. A concrete use of capture is a tool that converts its recorded graph into a particular export format. The readiness of that integration needs to be checked in the relevant project.
Real workloads. The reported speed improvement comes from one configuration. Other models and devices need their own measurements. QLoRA, metric and failure tests cover particular scenarios; their scope is described in the technical review.
Integration still pending at the cutoff. Atomic
SafeTensors saving, Fusion WriteScope, dispatch restructuring and
batched SVD entered main in September. Their August status
and subsequent integration dates appear in the WIP section.
The window contains 137 merged PRs and one direct commit to
main. This memo is selective: the seven developments above
are the ones worth management attention. The technical review covers all merged
work in the window, connecting claims to code and tests, with WIP
recorded separately.
Public Git does not show private work, revenue or adoption. PR measurements are attributed to their authors; tests were inspected but not independently executed for this review.
Burn and Tracel AI are trademarks of their respective owners. PRA is not affiliated with Tracel AI.