Issues

• Feature request: Metal command-queue priority / isolation API for concurrent workloads

  ml-explore/mlx#4716 · Filter by author dahai80 opened on Oct 17, 2026

• [BUG] fft2/fftn over a length 1 axis can return wrong results on large Metal inputs

  ml-explore/mlx#4704 · Filter by author StockerMC opened on Oct 15, 2026

• [BUG] [CUDA] Illegal __global__ write in copy_s under CUDA graphs on consumer Blackwell (sm_120) → corrupted output

  bugcuda
  ml-explore/mlx#4700 · Filter by author zhanpengwang2000 opened on Oct 14, 2026
  1

• Fused attention for head dim 256 from 512 query rows

  ml-explore/mlx#4695 · Filter by author jvmenen opened on Oct 14, 2026

• Return the logsumexp from fused SDPA, for merging attention over KV chunks

  ml-explore/mlx#4693 · Filter by author jvmenen opened on Oct 14, 2026

• Larger M tile for the NAX affine quantized matmul during prefill

  ml-explore/mlx#4692 · Filter by author jvmenen opened on Oct 14, 2026

• [BUG] fast.cross_entropy VJP returns wrong gradients for large non-contiguous inputs on Metal (0.32.3 regression)

  bug
  ml-explore/mlx#4686 · Filter by author clam004 opened on Oct 13, 2026
  2

• [RFC] A Structured Optimization Pipeline for MLX Compile

  enhancement
  ml-explore/mlx#4681 · Filter by author XXXXRT666 opened on Oct 11, 2026
  3

• [Metal] bf16 affine gather_qmm precision discrepancy versus quantized_matmul for multi-token MoE inputs

  ml-explore/mlx#4631 · Filter by author griffinwork40 opened on Oct 5, 2026

• qmm: stage dequantized weights as int8/fp8 instead of fp16 to use the 8-bit matrix lanes

  ml-explore/mlx#4627 · Filter by author magnuslundstedt opened on Oct 5, 2026
  2

• [BUG] Metal softmax / logsumexp / rms_norm / layer_norm drop rows when grid width exceeds 2^32

  buglow priority
  ml-explore/mlx#4624 · Filter by author jniimi opened on Oct 4, 2026

• quantized_matmul takes 14- 41% longer than dequantize + bf16 matmul at prefill sizes (> 512 tokens)

  ml-explore/mlx#4621 · Filter by author mikkokam opened on Oct 4, 2026

• mx.fast.gated_delta_update VJP: wrong gradient w.r.t. g (and ~1e-3 error on all inputs) on the NAX path for Hk=16

  bug
  ml-explore/mlx#4614 · Filter by author kevincaldwellgordon-bot opened on Oct 2, 2026
  

• Custom kernel (mx.fast.metal_kernel, MathModeSafe) returns wrong output rows on M3 Ultra / macOS 27.0 — exact under Relaxed and on M2; MLX-free reproducer

  bug
  ml-explore/mlx#4603 · Filter by author Pietroski opened on Oct 1, 2026
  1

• Option to opt out of TF32 for fp32 conv: small-kD conv3d is ~5000× less precise on M5 since 0.32

  enhancement
  ml-explore/mlx#4595 · Filter by author xocialize opened on Sep 29, 2026
  5

• [BUG] custom_function.jvp and .vmap are ignored when the body contains a Metal kernel

  bug
  ml-explore/mlx#4580 · Filter by author alford-auld opened on Sep 28, 2026
  2

• fp32 matmul with small N is ~5.5x slower at M=2..8 than M=1 (GEMM path vs gemv): (M,16384)@(16384,24) 120us vs 22us on M5 Ultra

  performance
  ml-explore/mlx#4571 · Filter by author sethforprivacy opened on Sep 27, 2026

• [BUG] mlx.sort return value incorrect for shape (65537, 65536) on GPU

  buglow priority
  ml-explore/mlx#4538 · Filter by author kvcache670 opened on Sep 20, 2026

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

  performance
  ml-explore/mlx#4521 · Filter by author freddyhaddad opened on Sep 16, 2026
  1

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

  performance
  ml-explore/mlx#4509 · Filter by author freddyhaddad opened on Sep 14, 2026
  1

• [BUG] threads and streams issues

  bug
  ml-explore/mlx#4506 · Filter by author davidkoski opened on Sep 14, 2026
  2

• mx.distributed ring backend: gradient-shaped CPU-stream all_sum deadlocks probabilistically (~3-4%/call) across a training loop; GPU-stream variant crashes deterministically at ~5s

  bugdistributed
  ml-explore/mlx#4475 · Filter by author ictashik opened on Sep 8, 2026
  1

• Metal: distributed collectives run on the CPU stream, and with the default fence each crossing costs ~180 µs (MLX_METAL_FAST_SYNCH: ~4 µs)

  distributedperformance
  ml-explore/mlx#4438 · Filter by author avlp12 opened on Aug 31, 2026
  2

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

  performance
  ml-explore/mlx#4435 · Filter by author bahaehmimdi opened on Aug 30, 2026
  5

• [BUG] NAX MoE simdgroup skipping in 0.32.2 triggers lost JACCL completion under heterogeneous TP

  bugdistributed
  ml-explore/mlx#4403 · Filter by author jonathan308 opened on Aug 25, 2026
  1