mlx · Issues· 152 open

[BUG] Metal `col_reduce_longcolumn` produces wrong results for negative-stride views

ParallelFileReader: uncatchable fatal error when pread lands at EOF

macOS 26 wheel is built for 26.2 but tagged macosx_26_0 — pip installs it on 26.0/26.1, where NAX is enabled on M5-class GPUs

mx.dequantize returns garbage on CPU past 2^31 unpacked bits (int size in cpu/binary.h)

Metal command-buffer limits (50 ops / 50 MB) cost 5–8% decode throughput on launch-heavy models (M3 Ultra measurements)

`nn.Upsample(mode="linear")` is 2-4x slower than a matmul-based bilinear resize when upsampling

[BUG] threads and streams issues

[Performance] gather_qmm at M=1 streams 4-bit expert weights at ~half the bandwidth of M=4 (MoE decode)

mx.compile inlines float scalar constants with 7 significant digits, so compiled results differ from eager by 1 ulp

[BUG] `argsort` diverges between cpu and gpu on complexes with NaN

[Enhancement] Group.split() support for JACCL and Ring backends (parity with NCCL #3172)

[BUG] Inconsistent behaviour of the floot_divide integer arrays truncates towards 0

quantized_matmul is slower than dense bf16 when the output dimension is not a multiple of 8, and ~7x slower below 8

[BUG] JACCL segfaults in mx.distributed.init(backend="jaccl") on 2-node Mac-Mac with RDMA devices present

[ENH]: Using `IndexError` instead of `ValueError` at appropriate places?