Repository Issue Activity (beta)

#4298 einsum rejects scalar operands when the empty subscript is not first

#4370 [BUG] `mx.dequantize` produces wrong results on a non-contiguous slice (regression on 0.32.1)

#4389 [BUG] Metal Kernel Compilation Issue on M2 macOS 15 (Sequoia)

#4371 [BUG] DLPackCompatible protocol members are declared as attributes, so np.ndarray fails to match — mx.array(np_arr) no longer type-checks

#3979 conv2d on Metal silently returns all-zero output for large inputs (channel-count dependent, ~2e9 elements)

#4350 [BUG] Metal JIT build: qmv template instantiation passes 6 args to 5-param kernel templates (runtime abort)

#4358 [BUG] `__setitem__` with a single Ellipsis fails

#4386 linalg.inv on CPU stream aborts the process on a singular matrix instead of raising

#4379 [BUG] mx.var / mx.std wrong for offset data

#4341 [Performance] Avoid dense zero work in stride-2 ConvTranspose3d on Metal

#3514 Add full `array-api` tests

#4326 [BUG][CUDA] CommandEncoder remains corrupted after graph-cache thrashing exception

#4377 The compiler cache and tracing state are process-global and unsynchronized in 0.31.x — is the 0.32 thread-local redesign the intended thread-safety contract?

#4239 Add `__array_namespace_info__()` support to `mlx.core`

#4329 [BUG] CPU errors are not propagated through async_eval and abort the process

#4283 [BUG] A jaccl hostfile whose rank order is wrong silently produces N groups of size 1

#1392 [Feature] Metal implementation of linalg APIs

#4327 Regression in 8e00a2d9d (#4248): fatal PyThreadState_Get abort at interpreter exit after mx.compile is used — compile-cache cleanup no longer runs before finalization

#4228 InstanceNorm FP16 variance overflow collapses finite feature maps

#4363 Performance guidance + GPU extended-precision (double-double) support — benchmark data attached

#4285 [BUG] ring sum_scatter's "not supported" error never reaches the caller, the group hangs instead

#3954 [New feature] Support exporting memory snapshots compatible with PyTorch's memory viz format

#4061 Metal: set_vector_bytes passes nullptr to setBytes for empty vectors

#4300 [BUG] Metal scatter does not support 64bit dtype

#4355 Support fusing element-wise operations into reductions