Search results

Intermittent corruption of small int32 arrays under long lazy accumulation chains + in-place slice updates + async_eval (batch KV cache bookkeeping)

Status: Open.

#3866 In ml-explore/mlx;

· nickbc31-sketch opened on Jul 17, 2026

[BUG] nn.quantize: a class_predicate that selects a module without to_quantized makes the whole call fatal — apply the documented "defines to_quantized" floor to explicit predicates too

Status: Open.

#3865 In ml-explore/mlx;

· danielkhalilov opened on Jul 17, 2026

[BUG] distributed ring: SocketThread dies silently on transient connection reset — all ranks wedge forever in Event::wait

Status: Open.

#3862 In ml-explore/mlx;

· Stefanfindeisen opened on Jul 17, 2026

[PERF] [CUDA] fp32 matmul ~36× slower than cuBLAS on sm_120 (RTX 50 / consumer Blackwell)

Status: Open.

#3861 In ml-explore/mlx;

· stoyoda0012-cyber opened on Jul 17, 2026

[BUG] [CUDA] fp32 matmul silently defaults to TF32 (MLX_ENABLE_TF32=1); undocumented, diverges from Metal

Status: Open.

#3860 In ml-explore/mlx;

· stoyoda0012-cyber opened on Jul 17, 2026

[PACKAGING] mlx[cuda13] missing CUDA runtime/CCCL headers — first GPU op fails; silent abort under pytest

Status: Open.

#3859 In ml-explore/mlx;

· stoyoda0012-cyber opened on Jul 17, 2026

[BUG] [CUDA] Batched matrix-vector product crashes when batch > 65,535 (gridDim limit)

Status: Open.

#3858 In ml-explore/mlx;

· stoyoda0012-cyber opened on Jul 17, 2026

Silent numerical corruption in single long forward on quantized MoE when sequence length % 32 != 0 (Qwen3-Coder-30B-A3B-8bit, M5)

Status: Open.

#3856 In ml-explore/mlx;

· pierre427 opened on Jul 17, 2026

2-bit quantized_matmul loses its size advantage over 4-bit at M >= 3

Status: Open.

#3852 In ml-explore/mlx;

· ARahim3 opened on Jul 16, 2026

How can the Metal resource limit be queried or configured on macOS?

Status: Open.

#3849 In ml-explore/mlx;

· Lyxot opened on Jul 16, 2026

mx.random.categorical(logits, num_samples=M) allocates O(N·M) memory

Status: Open.

#3847 In ml-explore/mlx;

· michaelellis003 opened on Jul 15, 2026

qmv_wide: residual small-m cost on larger shapes — m=4–8 still 1.5–2.7x of m=1 on M4 Pro

Status: Open.

#3839 In ml-explore/mlx;

· abcgco opened on Jul 13, 2026