Issues

• Performance profile: where does all the time go?

  ml-explore/mlx#4806 · Filter by author jrp2014 opened on Oct 30, 2026

• Use built in GitHub bypass instead of manually updating PR Limits

  ml-explore/mlx#4804 · Filter by author madrob opened on Oct 30, 2026

• CPU qqmm ignores global_scale_x and global_scale_w (nvfp4)

  ml-explore/mlx#4800 · Filter by author rcfa opened on Oct 30, 2026

• CPU quantized_matmul, gather_qmm and qqmm crash with SIGFPE on x86-64 when M or K is 0

  ml-explore/mlx#4799 · Filter by author rcfa opened on Oct 30, 2026

• CPU arange in floating types sums a running total: float32 values stop growing past 2^24 elements

  ml-explore/mlx#4798 · Filter by author rcfa opened on Oct 30, 2026

• CPU integer divide, remainder and divmod crash with SIGFPE on x86-64 for a signed minimum divided by -1

  ml-explore/mlx#4797 · Filter by author rcfa opened on Oct 30, 2026

• [BUG] slice_update with reductions causes silent memory corruption / mutates input when output offset exceeds INT32_MAX

  ml-explore/mlx#4782 · Filter by author jniimi opened on Oct 28, 2026

• [BUG] Metal expm1 (and erf for |x| > 0.93) reads up to 336 ulp high: the argument reduction drops ln 2's low part

  ml-explore/mlx#4765 · Filter by author belisoful opened on Oct 25, 2026

• [BUG] Metal layer_norm writes outside thread_x and runs 14x to 70x slower on rows of 64 and 16 than on rows of 256

  ml-explore/mlx#4759 · Filter by author ArupKantiDas opened on Oct 24, 2026

• [BUG] FP4 quantization on CPU rounds ties toward zero while Metal rounds to even

  ml-explore/mlx#4753 · Filter by author sw23 opened on Oct 23, 2026

• [BUG] mx.{to,from}_fp8 w/ float64 returns incorrect results on CPU

  ml-explore/mlx#4752 · Filter by author sw23 opened on Oct 23, 2026
  1

• Proposal : Benchmarking Improvement - Aligning measure_runtime with time_fn in time_utils.py

  ml-explore/mlx#4747 · Filter by author manig1729 opened on Oct 22, 2026

• CPU segmented_mm gives wrong values for bfloat16 and float16

  ml-explore/mlx#4739 · Filter by author rcfa opened on Oct 21, 2026

• [CUDA] gather_mm with sorted_indices=True gives wrong rows when x's size-1 axis comes from expand_dims

  bugcuda
  ml-explore/mlx#4727 · Filter by author IamCooller opened on Oct 19, 2026

• [BUG] qmm_splitk is chosen over qmm_nax on M5 for N=4096 projections M=16-64

  bugperformance
  ml-explore/mlx#4720 · Filter by author sbryngelson opened on Oct 17, 2026
  1

• [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
  3

• Use fused attention for head dim 256 from 512 query rows with enough heads

  enhancementlow priorityperformance
  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

  enhancementlow priority
  ml-explore/mlx#4693 · Filter by author jvmenen opened on Oct 14, 2026
  5

• [RFC] A Structured Optimization Pipeline for MLX Compile

  enhancement
  ml-explore/mlx#4681 · Filter by author XXXXRT666 opened on Oct 11, 2026
  4

• 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
  3

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

  low priorityperformance
  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

• [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