Issues

autotune panics instead of degrading when a key has no viable candidate — f16 `Sum` on wgpu-Metal has none (`Atomic add not supported by the client for f16`)

bug

#1538

· scott-wyatt opened

on Aug 20, 2026

cubecl-wgpu: autotune OOMs the device benchmarking an oversized matmul, then corrupts the server

#1531

· zachelnet opened

on Aug 18, 2026

cubecl-cpu never reads CompilationConfig.cache — no persistent kernel cache exists, full MLIR compile bill on every process start

Feature

#1527

· andoriyu opened

on Aug 16, 2026

Metal (native runtime): macOS GPU watchdog (ImpactingInteractivity) kills large-model workloads on display-attached Apple Silicon; batch limits have no override and are not the lever

#1523

· scott-wyatt opened

on Aug 14, 2026

Metal (native runtime): generated MSL for f32->bf16 conversion does not compile — 'cannot initialize const bfloat with an rvalue of type float'

#1522

· scott-wyatt opened

on Aug 14, 2026

Unroll pass panics on vector casts: CastOp is not TriviallyUnrollable

#1503

· ThierryCantin-Demers opened

on Aug 12, 2026

cubecl-environment's sqlite cache forces libsqlite3-sys on all downstream users (links conflict with sqlx)

#1488

· TsaoLun opened

on Aug 8, 2026

Support Metal Performance Primitives for accelerated MMA on Apple Silicon

#1480

· olukowski opened

on Aug 6, 2026

why disable Plane::Ops for all wgpu cpu backends?

#1467

· jyi2ya opened

on Aug 2, 2026

`sync_plane` does not establish shared memory visibility on spirv

#1466

· jyi2ya opened

on Aug 2, 2026

expose `PLANE_COUNT`

#1464

· jyi2ya opened

on Aug 1, 2026

A failed fire-and-forget device task leaves the buffer zero-filled and the next call succeeds

#1461

· ThierryCantin-Demers opened

on Jul 31, 2026

Navigation Menu

Uh oh!

All issues

autotune panics instead of degrading when a key has no viable candidate — f16 `Sum` on wgpu-Metal has none (`Atomic add not supported by the client for f16`)

cubecl-wgpu: autotune OOMs the device benchmarking an oversized matmul, then corrupts the server

cubecl-cpu never reads CompilationConfig.cache — no persistent kernel cache exists, full MLIR compile bill on every process start

Metal (native runtime): macOS GPU watchdog (ImpactingInteractivity) kills large-model workloads on display-attached Apple Silicon; batch limits have no override and are not the lever

Metal (native runtime): generated MSL for f32->bf16 conversion does not compile — 'cannot initialize const bfloat with an rvalue of type float'

Unroll pass panics on vector casts: CastOp is not TriviallyUnrollable

cubecl-environment's sqlite cache forces libsqlite3-sys on all downstream users (links conflict with sqlx)

Support Metal Performance Primitives for accelerated MMA on Apple Silicon

why disable Plane::Ops for all wgpu cpu backends?

`sync_plane` does not establish shared memory visibility on spirv

expose `PLANE_COUNT`

A failed fire-and-forget device task leaves the buffer zero-filled and the next call succeeds

Uh oh!

Search results