Issues

• wgpu on Metal: naga binds the metadata as device; with constant, burn's matmul takes 26% less time

  tracel-ai/cubecl#1931 · Filter by author yasumorishima opened on Oct 25, 2026

• cubecl-llvm (AMDGPU): a wave-uniform plane GEMV miscompiles under heavy SGPR spilling (machine-scheduler)

  tracel-ai/cubecl#1929 · Filter by author nathanielsimard opened on Oct 25, 2026

• Device channel: producers spin while the device thread compiles kernels

  tracel-ai/cubecl#1910 · Filter by author skewballfox opened on Oct 22, 2026

• Generated MSL differs between processes (per-process HashMap/HashSet seeds), which defeats Metal's shader cache

  tracel-ai/cubecl#1878 · Filter by author DivagoNM opened on Oct 20, 2026
  1

• wasm: autotune benchmarks plane-op matmul kernels on a device without subgroups, and the validation error panics

  tracel-ai/cubecl#1868 · Filter by author N34AY opened on Oct 19, 2026

• Consider snapshot semantics for lazy device reads

  tracel-ai/cubecl#1857 · Filter by author Liberxue opened on Oct 18, 2026

• modules cannot have the same name as a function.

  tracel-ai/cubecl#1841 · Filter by author Redhawk18 opened on Oct 15, 2026
  3

• Expand for into and from traits

  tracel-ai/cubecl#1839 · Filter by author Redhawk18 opened on Oct 15, 2026
  2

• Build failure on NixOS due to LLVM bundling

  tracel-ai/cubecl#1829 · Filter by author Jroid8 opened on Oct 13, 2026
  6

• CUDA LLVM segfaults when a kernel mixes vec4 and vec2 reductions

  tracel-ai/cubecl#1823 · Filter by author HongchengZhao opened on Oct 13, 2026

• [cubecl-wgpu / Vulkan] Matmul reads uninitialized/racing input buffers in long eager chains — bit-correct in isolation and bit-exact vs PyTorch on first pass, all-NaN when re-materialized

  tracel-ai/cubecl#1822 · Filter by author sebastianshoelace99-ctrl opened on Oct 13, 2026
  2

• IR: incorrect remainder-by-one and unconditional floating-point zero folds

  tracel-ai/cubecl#1812 · Filter by author Enferlain opened on Oct 12, 2026
  1

• Feature request: a cube-wide vote (sync_cube_or / sync_cube_count, CUDA __syncthreads_or) in #[cube]

  tracel-ai/cubecl#1808 · Filter by author terasakisatoshi opened on Oct 11, 2026

• Browser WebGPU: CNN cascade much slower than host CPU/GPU (and MaxPool WGSL unusable)

  tracel-ai/cubecl#1797 · Filter by author vesuvisian opened on Oct 9, 2026

• Shape::reshape panics on a zero-length target dim that isn't a copy placeholder

  tracel-ai/cubecl#1789 · Filter by author jaweed3 opened on Oct 8, 2026

• Add cbrt

  tracel-ai/cubecl#1785 · Filter by author Redhawk18 opened on Oct 8, 2026

• common: Accessing Bytes can cause panics

  bug
  tracel-ai/cubecl#1748 · Filter by author Charles23R opened on Oct 2, 2026
  1

• metal: cmma StoreOp ignores layout, column-major stores come out transposed

  bug
  tracel-ai/cubecl#1726 · Filter by author antimora opened on Sep 28, 2026

• cpp: signed mod_floor goes through f32, wrong for |x| > 2^24 on CUDA/HIP/Metal

  bug
  tracel-ai/cubecl#1725 · Filter by author antimora opened on Sep 28, 2026

• cubecl-metal fails to compile: missing DebugInformation import in context.rs

  tracel-ai/cubecl#1722 · Filter by author sunipkm opened on Sep 28, 2026

• Should add pattern syntax

  enhancement
  tracel-ai/cubecl#1721 · Filter by author wingertge opened on Sep 28, 2026

• CUDA C++: atomic outputs are Dead, failures get lost

  tracel-ai/cubecl#1712 · Filter by author swfsql opened on Sep 25, 2026

• Expose the memory a device has left, so a caller can size its work to what is actually free

  tracel-ai/cubecl#1698 · Filter by author vboussot opened on Sep 23, 2026

• On wasm a kernel whose tune is pending runs candidate 0 of its set, usually the slowest, for as long as the measurements take

  tracel-ai/cubecl#1695 · Filter by author vboussot opened on Sep 23, 2026

• The optimized layout on CUDA pads a 160-channel f16 row to 256 elements: 60% more memory for a channels-last activation

  tracel-ai/cubecl#1694 · Filter by author vboussot opened on Sep 23, 2026