Search results

MXFP4 quantization extremely slow on distributed GLM-5.1 (DeepSeek-V3.2 arch) — 0.27 tok/s vs 16 tok/s expected

Status: Open.

#3402 In ml-explore/mlx;

· playtheace opened on Apr 12, 2026

FSDP saving state improvement

cuda

distributed

enhancement

Status: Open.

#3401 In ml-explore/mlx;

· nastya236 opened on Apr 12, 2026

[Bug] Numerical divergence / Token repetition in fast.scaled_dot_product_attention for 4-bit Quantized Models

Status: Open.

#3384 In ml-explore/mlx;

· Mengzhen-Zhang opened on Apr 7, 2026

SIGSEGV / hang in mx.fast.scaled_dot_product_attention under vmap with GQA/MQA shapes (n_heads != n_kv_heads)

bug

Status: Open.

#3383 In ml-explore/mlx;

· ishaansingh22 opened on Apr 7, 2026

[BUG] get_shape() has no bounds check allowing read of outside data structure

Status: Open.

#3358 In ml-explore/mlx;

· MillaFleurs opened on Apr 3, 2026

[BUG] mx.fast.metal_kernel: use-after-free when multiple custom kernels compose in lazy graph

bug

Status: Open.

#3347 In ml-explore/mlx;

· Ziqiao-git opened on Mar 31, 2026

[BUG] mx.compile produces wrong results when chaining unsigned->signed astype with matmul

bug

Status: Open.

#3338 In ml-explore/mlx;

· tillahoffmann opened on Mar 30, 2026

Add support for computing matrix determinants

enhancement

Status: Open.

#3335 In ml-explore/mlx;

· jessegrabowski opened on Mar 30, 2026

Kernels for scipy.special statistical functions

enhancement

Status: Open.

#3330 In ml-explore/mlx;

· jessegrabowski opened on Mar 29, 2026

SIGSEGV in Compiled::eval_gpu when auto-fused kernel receives lazy/mmap'd array with null MTLBuffer

Status: Open.

#3329 In ml-explore/mlx;

· FHult opened on Mar 29, 2026

[BUG] Arrays with >= 2^31 elements fail materialization and some indexing on Metal

bug

Status: Open.

#3327 In ml-explore/mlx;

· Aristide021 opened on Mar 26, 2026

"test random uniform" fails for CPU backend

bug

Status: Open.

#3326 In ml-explore/mlx;

· zcbenz opened on Mar 26, 2026