Issues

• [CUDA] 0.32.3: repeated forward pass of a 4-bit quantized model is not bitwise reproducible on GB10 (sm_121); not reproduced with 0.32.4

  ml-explore/mlx#4843 · Filter by author bazucas opened on Nov 6, 2026
  1

• Metal cache admission can exceed the configured cache limit

  ml-explore/mlx#4841 · Filter by author tijs opened on Nov 6, 2026

• mx.checkpoint does not bound backward memory of a chunked recurrence when per-step inputs need gradients

  ml-explore/mlx#4837 · Filter by author hatim-mlf opened on Nov 5, 2026
  1

• [BUG] Unsynchronized reads of MetalAllocator::active_memory_ and Scheduler::n_active_tasks_ in eval_impl (TSan)

  ml-explore/mlx#4835 · Filter by author goodmai opened on Nov 5, 2026

• [BUG] JIT build: attention_nax_dsplit fails to compile (extra bv template argument)

  ml-explore/mlx#4832 · Filter by author kainosnoema opened on Nov 4, 2026
  1

• [BUG] mx.expand_dims docs: the Args section names axes, but the parameter is axis

  ml-explore/mlx#4829 · Filter by author Redempt1onzzZZ opened on Nov 4, 2026

• Quantized matmul: opt-in mode where a batched row matches M=1

  ml-explore/mlx#4819 · Filter by author Augustas11 opened on Nov 2, 2026

• [BUG] mx.qqmm docs: the Args section names global_scale, but the parameter is global_scale_x

  ml-explore/mlx#4815 · Filter by author Redempt1onzzZZ opened on Nov 2, 2026

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

  bug
  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

• [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
  6

• [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
  1

• 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