Repository Issue Activity (beta)

#3803 Metal GPU timeout when loading large MLX models from NFS / external volumes; same snapshots work from internal SSD

#3777 [BUG] JACCL segfaults in ibv_reg_mr (null PD) when RDMA absent; distributed_config says "RDMA not enabled" despite rdma_ctl status=enabled, survives cold boot (macOS 26.5.1 / 25F80)

#3799 withdrawn

#3797 [BUG] NAX split-K GEMM (steel_gemm_splitk_axpby_nax) returns garbage for bf16 on M5 Max at the M·N ≥ 2048² dispatch boundary

#3186 Kernel panic (IOGPUMemory.cpp:550) on M4 Max with large context prefill (~173K tokens)

#2185 How to implement custom multi-thread cpu kernels?

#3795 `mx.fast.rope` returns wrong values for batch rows > 0 when seq_len == 1 and batch > 1

#3702 A19 (iPhone 17 Pro) GPU returns numerically wrong results in float32 workloads — deterministic HF corruption in a Demucs separation pipeline; M2/M3 clean

#3789 Decode performance optimizations for MLX inference

#1439 [BUG] ValueError: [scatter] Cannot calculate VJP with respect to indices.

#3787 [BUG] ScatterAxis::vjp throws "Cannot calculate VJP with respect to indices" for computed indices (blocks autograd through MoE routing)

#3784 mx.sum over a non-last axis returns all zeros for certain large tensors (e.g. [1, 32, 4096, 131072])

#451 [BUG] value_and_grad is really slow

#3484 [BUG] Missing `array-api` methods

#1851 Add support for 1.5bit quantization

#2457 [BUG] Enable multiprocessing queues with mx.array

#2583 [BUG] Performance regression when yaml opening many files

#2770 Slower training speed for ResNet-18 compared to Torch

#2887 [BUG] Get 'kernel not found' error when doing quantized_matmul with mxfp8

#3200 [BUG] test_gather_qmm_sorted consistently fails in some aarch64-darwin environments

#3496 MLXNN.RoPE produces row-asymmetric output for byte-identical row inputs at [B>1, H, 1, D]

#3330 Kernels for scipy.special statistical functions

#2878 Feature Request: On-disk (out-of-core) streaming of model weights for low-RAM inference

#3573 [BUG] mx.fast.scaled_dot_product_attention: position-0 output changes at K=8 → K=9, all dtypes

#3776 mlx.launch should properly quote the hostname