Issues

• burn-vision morphology reads past its row buffer on 64-bit lanes on the scalar fallback

  tracel-ai/burn#6411 · Filter by author antimora opened on Nov 5, 2026

• Adam/AdamW produce NaN on f16 wgpu backend

  tracel-ai/burn#6401 · Filter by author Kostyanikov opened on Nov 4, 2026
  2

• Moving a bool tensor between CubeCL devices keeps Bool(U8) on a WGSL device, whose kernels then fail to compile

  tracel-ai/burn#6391 · Filter by author Jannik0 opened on Nov 4, 2026
  1

• flex: opt-in fast f32 erf and gelu (~6x faster, up to 5.5 ulp)

  flexperformance
  tracel-ai/burn#6362 · Filter by author Kostyanikov opened on Oct 31, 2026
  2

• flex: native CTC loss path to cut evaluation time and temporary memory

  flexperformance
  tracel-ai/burn#6356 · Filter by author Mikyx-1 opened on Oct 31, 2026

• cuda: sum_dim over a size-1 dim of a reshaped tensor returns scrambled values

  bug
  tracel-ai/burn#6350 · Filter by author verglasz opened on Oct 30, 2026
  1

• nn: soft-target cross-entropy and KL divergence gradients are noisy for near-equal logits

  bug
  tracel-ai/burn#6338 · Filter by author istaari opened on Oct 30, 2026
  1

• flex: vectorize the remaining scalar float ops

  flexperformance
  tracel-ai/burn#6334 · Filter by author antimora opened on Oct 30, 2026
  1

  

• flex: conv1d falls back to pseudo-3D im2col for padded, dilated or grouped convolutions

  flexperformance
  tracel-ai/burn#6321 · Filter by author onenewcode opened on Oct 30, 2026

• cuda: conv2d_asymmetric_weight_backward test fails intermittently

  bug
  tracel-ai/burn#6315 · Filter by author vboussot opened on Oct 29, 2026

• cubecl: convolution autotune only tries the plane selector's default tile

  autotuneperformance
  tracel-ai/burn#6314 · Filter by author vboussot opened on Oct 29, 2026

• wgpu: 3x3x3 convolutions on WGSL only use the direct kernel

  performancewgpu
  tracel-ai/burn#6297 · Filter by author vboussot opened on Oct 29, 2026

• cubecl: memory_pool_report is empty on wasm and with exclusive-memory-only

  bugwgpu
  tracel-ai/burn#6296 · Filter by author vboussot opened on Oct 29, 2026

• cubecl: matmul and reductions with few outputs over a long axis run nearly serially

  performance
  tracel-ai/burn#6295 · Filter by author vboussot opened on Oct 29, 2026

• tensor: add eigenvalue and eigenvector decomposition

  feature
  tracel-ai/burn#6253 · Filter by author Redhawk18 opened on Oct 27, 2026
  3

• cuda: bf16 matmul autotune runs a TMA candidate with tile > 256 and fails silently (cuda-cpp)

  autotunebug
  tracel-ai/burn#6234 · Filter by author bayanasar opened on Oct 27, 2026
  2

• cubecl: duplicate-safe numeric scatter_nd reductions (Mul, Min, Max)

  enhancement
  tracel-ai/burn#6213 · Filter by author HarpeLm opened on Oct 24, 2026

• nn: softmax, log_softmax and CrossEntropyLoss zero the target gradient at high confidence

  bug
  tracel-ai/burn#6212 · Filter by author antimora opened on Oct 24, 2026
  3

• cubecl: work-efficient parallel scan for cumulative ops

  performance
  tracel-ai/burn#6122 · Filter by author saiyambharara opened on Oct 18, 2026

• autodiff: JAX-style function transformations (grad of pure functions)

  feature
  tracel-ai/burn#6090 · Filter by author jpssrocha opened on Oct 16, 2026
  2

• docs: note on Tensor::from_data that the int dtype comes from the device default

  documentation
  tracel-ai/burn#6006 · Filter by author altunenes opened on Oct 8, 2026
  6

• docs: docs.rs fails to build burn-cuda, burn-rocm and burn-cpu (tracel-llvm-bundler)

  documentationinfrastructure
  tracel-ai/burn#5974 · Filter by author antimora opened on Oct 4, 2026

• wgpu: wrong, saturated inference outputs on Metal (CPU backend correct)

  bugwgpu
  tracel-ai/burn#5962 · Filter by author knarf-f opened on Oct 3, 2026
  7

• train: make DDP return Result instead of panicking

  refactor
  tracel-ai/burn#5959 · Filter by author Charles23R opened on Oct 2, 2026

• wgpu: system memory grows over multi-day training runs

  bugwgpu
  tracel-ai/burn#5891 · Filter by author joshhansen opened on Sep 28, 2026 · patch-0.22.1
  19