v0.29.0

Highlights

• Support for mxfp4 quantization (Metal, CPU)
• More performance improvements, bug fixes, features in CUDA backend
• mx.distributed supports NCCL back-end for CUDA

What's Changed

• [CUDA] Optimize set_mm_device_pointers for small ndim by @zcbenz in #2473
• Fix logsumexp/softmax not fused for some cases by @zcbenz in #2474
• Use CMake <4.1 to avoid the nvpl error by @angeloskath in #2489
• Fix incorrect interpretation of unsigned dtypes in reduce ops by @abeleinin in #2477
• make code blocks copyable by @Dan-Yeh in #2480
• Rename cu::Matmul to CublasGemm by @zcbenz in #2488
• Faster general unary op by @awni in #2472
• The naive_conv_2d is no longer used by @zcbenz in #2496
• Remove the hack around SmallVector in cpu compile by @zcbenz in #2494
• Clean up code handling both std::vector and SmallVector by @zcbenz in #2493
• [CUDA] Fix conv grads with groups by @zcbenz in #2495
• Update cuDNN Frontend to v1.14 by @zcbenz in #2505
• Ensure small sort doesn't use indices if not argsort by @angeloskath in #2506
• Ensure no oob read in gemv_masked by @angeloskath in #2508
• fix custom kernel test by @awni in #2510
• No segfault with uninitialized array.at by @awni in #2514
• Fix lapack svd by @awni in #2515
• Split cuDNN helpers into a separate header by @zcbenz in #2491
• [CUDA] Add GEMM-based fallback convolution kernels by @zcbenz in #2511
• Fix docs by @russellizadi in #2518
• Fix overflow in large filter small channels by @angeloskath in #2520
• [CUDA] Fix stride of singleton dims before passing to cuDNN by @zcbenz in #2521
• Custom cuda kernel by @angeloskath in #2517
• Fix docs omission by @angeloskath in #2524
• Fix power by @awni in #2523
• NCCL backend by @nastya236 in #2476
• [CUDA] Nccl pypi dep + default for cuda by @awni in #2526
• Fix warning 186-D from nvcc by @zcbenz in #2527
• [CUDA] Update calls to cudaMemAdvise and cudaGraphAddDependencies for CUDA 13 by @andportnoy in #2525
• nccl default for backend=any by @awni in #2528
• Fix allocation bug in NCCL by @awni in #2530
• Enable COMPILE_WARNING_AS_ERROR for linux builds in CI by @zcbenz in #2534
• [CUDA] Remove thrust in arange by @zcbenz in #2535
• Use nccl header only when nccl is not present by @awni in #2539
• Allow pathlib.Path to save/load functions by @awni in #2541
• Remove nccl install in release by @awni in #2542
• [CUDA] Implement DynamicSlice/DynamicSliceUpdate by @zcbenz in #2533
• Remove stream from average grads so it uses default by @awni in #2532
• Enable cuda graph toggle by @awni in #2545
• Tests for save/load with Path by @awni in #2543
• Run CPP tests for CUDA build in CI by @zcbenz in #2544
• Separate cpu compilation cache by versions by @zcbenz in #2548
• [CUDA] Link with nccl by @awni in #2546
• [CUDA] Use ConcurrentContext in concatenate_gpu by @zcbenz in #2549
• [CUDA] fix sort by @awni in #2550
• Add mode parameter for quantization by @awni in #2499
• Bump xcode in circle by @awni in #2551
• Fix METAL quantization in JIT + fix release build by @awni in #2553
• Faster contiguous gather for indices in the first axis by @awni in #2552
• version bump by @awni in #2554
• Fix quantized vjp for mxfp4 by @awni in #2555

New Contributors

• @Dan-Yeh made their first contribution in #2480
• @russellizadi made their first contribution in #2518
• @andportnoy made their first contribution in #2525

Full Changelog: v0.28.0...v0.29.0