Issues

• top-k with a NaN in the row returns the empty-slot identity (-inf, -1) for every slot

  tracel-ai/cubecl#2133 · Filter by author Jannik0 opened on Nov 10, 2026

• let dims = (2usize, 2usize); fails to compile

  tracel-ai/cubecl#2114 · Filter by author Redhawk18 opened on Nov 9, 2026
  1

• f16 transcendentals on SPIR-V/WGSL: accuracy varies by driver (Intel tanh(1e-4) = 0)

  tracel-ai/cubecl#2113 · Filter by author Jannik0 opened on Nov 9, 2026

• powf polyfill: wrong IEEE special cases on SPIR-V and WGSL (1^NaN, NaN^0, 0^0, (-0)^-1, negative base)

  tracel-ai/cubecl#2112 · Filter by author Jannik0 opened on Nov 9, 2026

• SPIR-V: float != lowers to OpFOrdNotEqual, so NaN != x is false

  tracel-ai/cubecl#2111 · Filter by author Jannik0 opened on Nov 9, 2026

• CUDA: launch failure of a kernel over the per-thread local-memory limit is an anonymous INVALID_VALUE; expose kernel resource attributes

  tracel-ai/cubecl#2108 · Filter by author tbetcke opened on Nov 9, 2026

• CUBECL_CUDA_DUMP_PTX: file names collide across comptime variants of a kernel

  tracel-ai/cubecl#2107 · Filter by author tbetcke opened on Nov 9, 2026

• CPU runtime: SIGSEGV/SIGABRT at process exit when a kernel compile is still running on the device server thread

  tracel-ai/cubecl#2101 · Filter by author tbetcke opened on Nov 8, 2026

• CPU cube barriers have severe launch latency under oversubscription

  tracel-ai/cubecl#2084 · Filter by author onenewcode opened on Nov 6, 2026

• Device channel: clients spin while the task queue is full; throughput collapses with more submitting threads than cores

  tracel-ai/cubecl#2066 · Filter by author tbetcke opened on Nov 5, 2026

• wgpu (Vulkan): reading back a buffer of 4 GiB or more panics in the staging buffer

  tracel-ai/cubecl#2065 · Filter by author skewballfox opened on Nov 5, 2026

• wgpu (Vulkan): a failed device allocation is not retried after the memory pools are freed

  tracel-ai/cubecl#2064 · Filter by author skewballfox opened on Nov 5, 2026

• cubecl-llvm: i32::cast_from(u32) >> n lowers to a logical shift instead of an arithmetic one

  tracel-ai/cubecl#2026 · Filter by author lega0208 opened on Nov 1, 2026

• Incorrect matmul/bmm (and reduction) results on wgpu

  tracel-ai/cubecl#2023 · Filter by author Kostyanikov opened on Oct 31, 2026
  1

• WGSL: an infinite float constant is printed as f32(+Inf), which no WGSL compiler accepts

  tracel-ai/cubecl#1997 · Filter by author vboussot opened on Oct 29, 2026

• let (mut a, mut b) = (x, y) in a kernel fails IR verification

  tracel-ai/cubecl#1996 · Filter by author vboussot opened on Oct 29, 2026

• Process aborts at exit: device runner thread drops the persistence (turso) connection during TLS teardown ("thread local panicked on drop")

  tracel-ai/cubecl#1981 · Filter by author jesusdpa1 opened on Oct 29, 2026

• Acquire/release atomics and a fence one unit can issue, for cubes that hand results to each other in order

  tracel-ai/cubecl#1961 · Filter by author bayanasar opened on Oct 28, 2026

• Metal: Metal 4 tensor operations (MetalPerformancePrimitives) as an MMA path for matmul

  tracel-ai/cubecl#1952 · Filter by author bayanasar opened on Oct 28, 2026
  1

• Metal: bf16 x bf16 -> f32 simdgroup matrices are not registered, so bf16 matmuls take no MMA path

  tracel-ai/cubecl#1951 · Filter by author bayanasar opened on Oct 28, 2026

• log1p loses all precision for small |x| wherever the polyfill runs (LLVM backend, Metal, WGSL)

  tracel-ai/cubecl#1950 · Filter by author bayanasar opened on Oct 28, 2026

• 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