Repository Issue Activity (beta)

#5419 burn-cpu: stack overflow in internal worker thread on tall [N,3] x [3,3] matmul (N >= ~5.6M)

#5416 burn-tch: load libtorch at runtime (dlopen), no build-time link

#5332 No feature on `burn` forwards `burn-ndarray/multi-threads`, so `default-features = false` silently disables CPU threading

#5373 softplus overflows to inf for large inputs, producing NaN gradients

#5371 burn-vision: filter2d returns a larger image than the input for even-sized kernels

#5264 Bug in Linear on CUDA with Col layout: 0.21.0

#5382 burn-nn GroupNorm: f16 div_scalar underflows for large per-group element counts

#5413 burn-pack: Reader does not validate a descriptor's byte range against its shape and dtype

#5412 burn-store: PyTorch loader silently substitutes zeros for failed or truncated tensor reads

#4950 RemoteBackend occur error

#5219 burn_pack::Writer requires eager materialization of all tensor bytes before writing

#5399 burn-pack: make Writer::write_to_file atomic by default

#5402 Add exponential moving average (EMA) of model weights

#5393 Create a PR to add Burnpack visualization support to Netron

#5385 conv2d: burn-flex returns a finite value or NaN at padded positions for the same convolution, depending on channel count

#5388 NHWC relayout tests fail on CUDA + fusion

#5384 burn-ndarray 0.21: reading module weights after `Module::into_record()` on a cheap clone returns garbage (first-materialization aliasing)

#5368 Generic Tensor Ops supertrait of TensorMetadata

#5383 burn-store PytorchReader ignores tensor strides (non-contiguous .pth loads scrambled)

#5378 burn-store: replace TensorSnapshot with burn_pack::Tensor

#5380 burn-pack: test the streaming memory bound with a Drop hook instead of a global allocator

#5379 burn-optim: OptimizerRecord should hold TensorData, not burn_pack::Tensor

#5069 Flex: broadcast element-wise ops miss all SIMD fast paths and fall back to a per-element scalar odometer

#5375 burn-tch: permute/swap_dims/flip mint a fresh storage Arc for a shared buffer, defeating can_mut — libtorch panics on in-place logical_and_

#5351 Metal (wgpu-msl): LoRA adapter gradients explode exponentially with depth on M1 Max (Apple7) at seq_len >= 1024 — correct on newer Apple GPUs