mlx 0.32.0

[BrewTestBot]

Created by brew bump

---

Created with brew bump-formula-pr.

• resource blocks have been checked for updates.

release notes

## What's Changed
* Generate qmm implementaions with cmake by @zcbenz in https://github.com/ml-explore/mlx/pull/3424
* Enable swap for all CI building CUDA by @zcbenz in https://github.com/ml-explore/mlx/pull/3437
* Bump minor by @angeloskath in https://github.com/ml-explore/mlx/pull/3438
* [CUDA] Fix qmm_naive K-tail dispatch for FP quantized kernels by @Lyxot in https://github.com/ml-explore/mlx/pull/3445
* Keep gguflib input-validation asserts active in release builds by @qflen in https://github.com/ml-explore/mlx/pull/3436
* Reuse nightly build's ccache for release by @zcbenz in https://github.com/ml-explore/mlx/pull/3458
* Add barrier to JACCL by @Isalia20 in https://github.com/ml-explore/mlx/pull/3459
* Add determinant and sign-log-determinant functions to mlx.core.linalg by @abhilashreddys in https://github.com/ml-explore/mlx/pull/3416
* Define ST_F8_E8M0 by @pcuenca in https://github.com/ml-explore/mlx/pull/3448
* Clearer error when shape dimension overflows int32 by @serenposh in https://github.com/ml-explore/mlx/pull/3425
* [CUDA] Fix half type matmul in cutlass kernels by @zcbenz in https://github.com/ml-explore/mlx/pull/3469
* Fix indexing bug in slice update with op by @angeloskath in https://github.com/ml-explore/mlx/pull/3483
* Make device_count() return 0 when there is no GPU by @zcbenz in https://github.com/ml-explore/mlx/pull/3486
* Compute contiguity from the actual occupied data by @sofinvalery in https://github.com/ml-explore/mlx/pull/3475
* Do not use prebuilt cpu compile preamble when headers are installed by @zcbenz in https://github.com/ml-explore/mlx/pull/3463
* [CUDA] Separate main loop into a function in qmm by @zcbenz in https://github.com/ml-explore/mlx/pull/3443
* test: Upcast the random numbers before computing their average by @sofinvalery in https://github.com/ml-explore/mlx/pull/3488
* Pass deployment target when linking metallib by @dhiltgen in https://github.com/ml-explore/mlx/pull/3501
* Fix rope single token multiple sequences by @angeloskath in https://github.com/ml-explore/mlx/pull/3498
* Fix qvm_split_k incorrect batch stride calculation by @angeloskath in https://github.com/ml-explore/mlx/pull/3497
* Fix scatter_prod GPU hang on NaN with contention by @tillahoffmann in https://github.com/ml-explore/mlx/pull/3492
* [jaccl] Fix race on local_staging in MeshImpl::all_reduce by @kernelpool in https://github.com/ml-explore/mlx/pull/3451
* Add MLX_SDPA_BLOCKS env var for 2-pass vector kernel block-count override by @adurham in https://github.com/ml-explore/mlx/pull/3455
* mlx launch clean by @nastya236 in https://github.com/ml-explore/mlx/pull/3513
* [CUDA] Fix gather_mm by @zcbenz in https://github.com/ml-explore/mlx/pull/3503
* [CUDA] Guard qmm_naive scale and bias loads at tile boundaries by @Lyxot in https://github.com/ml-explore/mlx/pull/3509
* win: fix cuda build by @dhiltgen in https://github.com/ml-explore/mlx/pull/3532
* Improve DLPack-compatible array imports by @XXXXRT666 in https://github.com/ml-explore/mlx/pull/3495
* Apply the same thread-local approach to the CPU by @angeloskath in https://github.com/ml-explore/mlx/pull/3537
* removed automatic prepending python in mlx launch by @nastya236 in https://github.com/ml-explore/mlx/pull/3536
* ci: Show stack trace on crash by @zcbenz in https://github.com/ml-explore/mlx/pull/3538
* Handle non-multiple-of-8 spatial dims in depthwise conv2d Metal path by @qflen in https://github.com/ml-explore/mlx/pull/3446
* Remove reference to groups in Conv3D extra_repr by @GriffinMB in https://github.com/ml-explore/mlx/pull/3559
* Add buffer caching to no_gpu CPU allocator by @dhiltgen in https://github.com/ml-explore/mlx/pull/3554
* Fix steel GEMM safe load offset by @sofinvalery in https://github.com/ml-explore/mlx/pull/3560
* Fix singleton lifetime issues at process exit by @dhiltgen in https://github.com/ml-explore/mlx/pull/3555
* Fix doubled-word typos in docstrings by @LeSingh1 in https://github.com/ml-explore/mlx/pull/3561
* Fix off_x/off_y typo in steel BaseMMAFrag::load_safe by @mdegans in https://github.com/ml-explore/mlx/pull/3565
* Synchronize no-GPU cache eviction with CPU streams by @dhiltgen in https://github.com/ml-explore/mlx/pull/3566
* Add JIT compiler support for Windows by @dhiltgen in https://github.com/ml-explore/mlx/pull/3556
* [Metal] Reject tensor-scale nvfp4 in qqmm by @Brooooooklyn in https://github.com/ml-explore/mlx/pull/3551
* Detect int32 shape-product overflow at MLX compute-shape boundaries by @qflen in https://github.com/ml-explore/mlx/pull/3524
* Add `copy` keyword to `mx.asarray` by @eyupcanakman in https://github.com/ml-explore/mlx/pull/3510
* Fix typos in backend code comments by @adityasingh2400 in https://github.com/ml-explore/mlx/pull/3581
* Implement output_shapes for GatherMM and GatherQMM by @dexwritescode in https://github.com/ml-explore/mlx/pull/3485
* Fix activation docstring math in log_softmax and gelu_approx by @adityasingh2400 in https://github.com/ml-explore/mlx/pull/3590
* docs: expose missing Python APIs by @XXXXRT666 in https://github.com/ml-explore/mlx/pull/3598
* Fix CUDA all-reduce planning for large inputs by @sofinvalery in https://github.com/ml-explore/mlx/pull/3603
* Use uv in macOS CI by @zcbenz in https://github.com/ml-explore/mlx/pull/3491
* Fix HDRS_LIST building when paths in header inclusion tree include spaces by @lancelotblanchard in https://github.com/ml-explore/mlx/pull/3607
* Enable the Metal backend by default on iOS by @eyupcanakman in https://github.com/ml-explore/mlx/pull/3617
* Fix int32 overflow in matvec row offset and gather-MM batch stride by @aicayzer in https://github.com/ml-explore/mlx/pull/3609
* Fix signed-integer overflow (UB) in roll and tile shape arithmetic by @devYRPauli in https://github.com/ml-explore/mlx/pull/3604
* Fix complex VJPs for log and exp by @CameronChurchwell in https://github.com/ml-explore/mlx/pull/3605
* added env variables to propagate width by @nastya236 in https://github.com/ml-explore/mlx/pull/3618
* Correct p-norm computation in triplet_loss by @pchintar in https://github.com/ml-explore/mlx/pull/3613
* Fix threaded compile cache cleanup by @lucasnewman in https://github.com/ml-explore/mlx/pull/3628
* Handle invalid dimensions in SinusoidalPositionalEncoding by @pchintar in https://github.com/ml-explore/mlx/pull/3615
* Correct grammatical typo in label smoothing ValueError by @madhav1k in https://github.com/ml-explore/mlx/pull/3641
* Roll back compile cache entry when the first trace throws by @adityasingh2400 in https://github.com/ml-explore/mlx/pull/3635
* Raise on arange with step == 0 instead of undefined behavior by @devYRPauli in https://github.com/ml-explore/mlx/pull/3640
* Fix data race in tracing state by making it thread-local by @aicayzer in https://github.com/ml-explore/mlx/pull/3638
* Add missing break in bool_ switch case that would result in falling through to the uint8_t case by @psolanki in https://github.com/ml-explore/mlx/pull/3655
* Fix axis param in nll_loss by @ishtihoss in https://github.com/ml-explore/mlx/pull/3651
* Fix Adafactor factored update for params with more than 2 dims by @ishtihoss in https://github.com/ml-explore/mlx/pull/3652
* Emit valid kernel source for non-finite float constants by @tillahoffmann in https://github.com/ml-explore/mlx/pull/3648
* Validate inputs in GroupNorm and InstanceNorm by @ishtihoss in https://github.com/ml-explore/mlx/pull/3653
* Fix flaky test_siblings_without_eval by @zcbenz in https://github.com/ml-explore/mlx/pull/3621
* Fix intermittent wrong bias gradient in fast::layer_norm VJP (Metal WAR hazard) by @tillahoffmann in https://github.com/ml-explore/mlx/pull/3630
* NAX requires setting MACOSX_DEPLOYMENT_TARGET=26.2 by @zcbenz in https://github.com/ml-explore/mlx/pull/3622
* Catch error in CommandBuffer and poison the events by @zcbenz in https://github.com/ml-explore/mlx/pull/3523
* [Improvement] CopyType::Vector in concatenate if axis=0 and contiguous   by @nastya236 in https://github.com/ml-explore/mlx/pull/3663
* Fix Q4_1 GGUF loading by @ricky-chaoju in https://github.com/ml-explore/mlx/pull/3664
* Fix tree_map_with_path for namedtuples by @chrismicah in https://github.com/ml-explore/mlx/pull/3674
* Fix gather_qmm NAX kernel name mismatch by @scyyh11 in https://github.com/ml-explore/mlx/pull/3632
* Add bits and smallest_normal to finfo by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3679
* Fix use-after-free when a custom Metal kernel is called with different dtypes in one graph by @discobot in https://github.com/ml-explore/mlx/pull/3662
* Fix dimension count in orthogonal initializer error message by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3693
* fix grid for large uncontiguous input by @nastya236 in https://github.com/ml-explore/mlx/pull/3666
* Ignore numpy matmul warnings in test_blas by @zcbenz in https://github.com/ml-explore/mlx/pull/3657
* Fix repeat zero axis shape by @chrismicah in https://github.com/ml-explore/mlx/pull/3698
* Cache JIT-compiled CUDA kernels by @zcbenz in https://github.com/ml-explore/mlx/pull/3587
* [CUDA] JIT-compile qmm_naive by @zcbenz in https://github.com/ml-explore/mlx/pull/3576
* Support namedtuple and tuple subclasses in tree_merge by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3703
* Fix mx.sort vjp to transpose the permutation by @obchain in https://github.com/ml-explore/mlx/pull/3700
* Add isdtype, result_type, and can_cast by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3681
* [fix] allow exporting custom Metal kernels without a Metal backend by @xthomaswang in https://github.com/ml-explore/mlx/pull/3650
* Make step activation inclusive at the threshold by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3694
* Add new_thread_unsafe_stream API by @zcbenz in https://github.com/ml-explore/mlx/pull/3578
* Fix JVPs of power, divmod, and slice_update with partly traced inputs by @qflen in https://github.com/ml-explore/mlx/pull/3636
* Add flip and unstack by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3683
* Fix int16 overflow in NAX qmm edge-tile bounds by @scyyh11 in https://github.com/ml-explore/mlx/pull/3631
* Make gemv JIT compilable by @angeloskath in https://github.com/ml-explore/mlx/pull/3705
* Fix gaussian_nll_loss documented default reduction (none -> mean) by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3715
* Fix optimizer epsilon validation message and AdaDelta eps doc by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3712
* Add array API standard aliases for existing ops by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3678
* Fix ConvTranspose kernel size repr by @ishtihoss in https://github.com/ml-explore/mlx/pull/3724
* [CUDA] Disable managed memory on Tegra by @gmondada in https://github.com/ml-explore/mlx/pull/3701
* Update nanobind to 2.13.0 by @XXXXRT666 in https://github.com/ml-explore/mlx/pull/3722
* Fix JVPs of select, arctan2, masked_scatter, and bitwise ops by @qflen in https://github.com/ml-explore/mlx/pull/3633
* array API: add empty, empty_like (pure aliases), astype, matrix_transpose by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3729
* Add metal::set_metallib_path() to override the metallib search path by @olilarkin in https://github.com/ml-explore/mlx/pull/3597
* Add Metal DLPack zero-copy sharing by @XXXXRT666 in https://github.com/ml-explore/mlx/pull/3531
* Bump actions/checkout from 6 to 7 by @dependabot[bot] in https://github.com/ml-explore/mlx/pull/3741
* Fix shapeless compile eliding reductions over size-1 dimensions by @discobot in https://github.com/ml-explore/mlx/pull/3672
* [CUDA] JIT-compile qmm_sm80/qmm_sm90/gather_gemm by @zcbenz in https://github.com/ml-explore/mlx/pull/3706
* [CUDA] Make qmv support global scale by @zcbenz in https://github.com/ml-explore/mlx/pull/3723
* Add VJP for cumulative max and min scans by @devin-lai in https://github.com/ml-explore/mlx/pull/3599
* Fix abs jvp for complex inputs by @obchain in https://github.com/ml-explore/mlx/pull/3745
* Fix update_modules crash on over-length module lists by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3743
* Add matrix_norm to array API namespace by @wgu9 in https://github.com/ml-explore/mlx/pull/3749
* Guard scatter_axis against 64-bit outputs on the GPU by @obchain in https://github.com/ml-explore/mlx/pull/3695
* Fix build on Apple clang 15 in cpu_nd_array_to_mlx by @obchain in https://github.com/ml-explore/mlx/pull/3751
* metal: cache x in registers in rms_single_row to avoid redundant global read by @will-march in https://github.com/ml-explore/mlx/pull/3754
* Document row-wise sparsity in nn.init.sparse by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3752
* [CUDA] [Improvement] Rope without copy by @nastya236 in https://github.com/ml-explore/mlx/pull/3704
* Add small-batch quantized matvec kernel (qmv_wide) by @jessegross in https://github.com/ml-explore/mlx/pull/3764
* Fix docstring nits by @ishtihoss in https://github.com/ml-explore/mlx/pull/3758
* Fix logsumexp jvp to reduce along the axis by @obchain in https://github.com/ml-explore/mlx/pull/3708
* Add vecdot to array API namespace by @wgu9 in https://github.com/ml-explore/mlx/pull/3748
* Fix CPU dynamic slice copy bound for collapsed shapes by @Lyxot in https://github.com/ml-explore/mlx/pull/3739
* Enable fused SDPA vector kernel for asymmetric Q/V head dims (192, 128) by @yohann-bearzi in https://github.com/ml-explore/mlx/pull/3637
* array API: add positive, logical_xor, trunc, count_nonzero, diff, full_like by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3730
* array API: add cumulative_sum and cumulative_prod by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3731
* Fix complex vjps for several unary ops by @obchain in https://github.com/ml-explore/mlx/pull/3766
* Fix Upsample align_corners singleton output by @ishtihoss in https://github.com/ml-explore/mlx/pull/3769
* Add CI for Windows CUDA build by @zcbenz in https://github.com/ml-explore/mlx/pull/3775
* Fix multi-wire recv prefill deadlock in jaccl ring backend by @jasonpaulso in https://github.com/ml-explore/mlx/pull/3654
* fix: Quote hostname in mlx.launch ssh commands by @varshneydevansh in https://github.com/ml-explore/mlx/pull/3783
* Fix BatchNorm docstring to cover 4D input and test the NHWC path by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3782
* Fix kl_div_loss docstring formula and two doc nits by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3788
* Fix HardShrink to accept its documented lambd argument by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3786
* Refactor macOS CI by @zcbenz in https://github.com/ml-explore/mlx/pull/3806
* Bypass the PyTorch index TLS issues by @angeloskath in https://github.com/ml-explore/mlx/pull/3808
* Fix fp quantized matvec for output dim < 8 by @katlun-lgtm in https://github.com/ml-explore/mlx/pull/3804
* Fix CUDA RMSNorm small-row dispatch by @dhiltgen in https://github.com/ml-explore/mlx/pull/3792
* Fix Conv3d docstring argument order (padding before dilation) by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3793
* Add math mode option for custom Metal kernels by @shubhxho in https://github.com/ml-explore/mlx/pull/3728
* Fix binary_cross_entropy docstring argument order (weights before with_logits) by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3807
* Fix compiled kernel correctness for negative-strided inputs by @lyonsno in https://github.com/ml-explore/mlx/pull/3720
* Use large runner for Windows CUDA builds by @zcbenz in https://github.com/ml-explore/mlx/pull/3809
* Fix wrong type parameter passed to gemm_splitk_nax by @zcbenz in https://github.com/ml-explore/mlx/pull/3810
* Document default value of dilation in conv layer docstrings by @Pablosinyores in https://github.com/ml-explore/mlx/pull/3811
* Fix CPU gather transposing column-contiguous slices by @tillahoffmann in https://github.com/ml-explore/mlx/pull/3647
* Fix infinite norm negative axis mismatch bug for matrices with 2 or more dimensions. by @danlee2002 in https://github.com/ml-explore/mlx/pull/3756
* Fix conv2 gradients in grouped strided case on Metal by @ericphanson in https://github.com/ml-explore/mlx/pull/3800
New Contributors

@qflen made their first contribution in Keep gguflib input-validation asserts active in release builds ml-explore/mlx#3436
@Isalia20 made their first contribution in Add barrier to JACCL ml-explore/mlx#3459
@abhilashreddys made their first contribution in Add determinant and sign-log-determinant functions to mlx.core.linalg ml-explore/mlx#3416
@serenposh made their first contribution in Clearer error when shape dimension overflows int32 ml-explore/mlx#3425
@sofinvalery made their first contribution in Compute contiguity from the actual occupied data ml-explore/mlx#3475
@tillahoffmann made their first contribution in Fix scatter_prod GPU hang on NaN with contention ml-explore/mlx#3492
@adurham made their first contribution in Add MLX_SDPA_BLOCKS env var for 2-pass vector kernel block-count override ml-explore/mlx#3455
@GriffinMB made their first contribution in Remove reference to groups in Conv3D extra_repr ml-explore/mlx#3559
@LeSingh1 made their first contribution in Fix doubled-word typos in docstrings ml-explore/mlx#3561
@mdegans made their first contribution in Fix off_x/off_y typo in steel BaseMMAFrag::load_safe ml-explore/mlx#3565
@eyupcanakman made their first contribution in Add copy keyword to mx.asarray ml-explore/mlx#3510
@adityasingh2400 made their first contribution in Fix typos in backend code comments ml-explore/mlx#3581
@dexwritescode made their first contribution in Implement output_shapes for GatherMM and GatherQMM ml-explore/mlx#3485
@lancelotblanchard made their first contribution in Fix HDRS_LIST building when paths in header inclusion tree include spaces ml-explore/mlx#3607
@aicayzer made their first contribution in Fix int32 overflow in matvec row offset and gather-MM batch stride ml-explore/mlx#3609
@devYRPauli made their first contribution in Fix signed-integer overflow (UB) in roll and tile shape arithmetic ml-explore/mlx#3604
@pchintar made their first contribution in Correct p-norm computation in triplet_loss ml-explore/mlx#3613
@madhav1k made their first contribution in Correct grammatical typo in label smoothing ValueError ml-explore/mlx#3641
@psolanki made their first contribution in Add missing break in bool_ switch case that would result in falling through to the uint8_t case ml-explore/mlx#3655
@ishtihoss made their first contribution in Fix axis param in nll_loss ml-explore/mlx#3651
@ricky-chaoju made their first contribution in Fix Q4_1 GGUF loading ml-explore/mlx#3664
@chrismicah made their first contribution in Fix tree_map_with_path for namedtuples ml-explore/mlx#3674
@scyyh11 made their first contribution in Fix gather_qmm NAX kernel name mismatch ml-explore/mlx#3632
@katlun-lgtm made their first contribution in Add bits and smallest_normal to finfo ml-explore/mlx#3679
@discobot made their first contribution in Fix use-after-free when a custom Metal kernel is called with different dtypes in one graph ml-explore/mlx#3662
@Pablosinyores made their first contribution in Fix dimension count in orthogonal initializer error message ml-explore/mlx#3693
@obchain made their first contribution in Fix mx.sort vjp to transpose the permutation ml-explore/mlx#3700
@xthomaswang made their first contribution in [fix] allow exporting custom Metal kernels without a Metal backend ml-explore/mlx#3650
@gmondada made their first contribution in [CUDA] Disable managed memory on Tegra ml-explore/mlx#3701
@olilarkin made their first contribution in Add metal::set_metallib_path() to override the metallib search path ml-explore/mlx#3597
@devin-lai made their first contribution in Add VJP for cumulative max and min scans ml-explore/mlx#3599
@wgu9 made their first contribution in Add matrix_norm to array API namespace ml-explore/mlx#3749
@will-march made their first contribution in metal: cache x in registers in rms_single_row to avoid redundant global read ml-explore/mlx#3754
@yohann-bearzi made their first contribution in Enable fused SDPA vector kernel for asymmetric Q/V head dims (192, 128) ml-explore/mlx#3637
@jasonpaulso made their first contribution in Fix multi-wire recv prefill deadlock in jaccl ring backend ml-explore/mlx#3654
@varshneydevansh made their first contribution in fix: Quote hostname in mlx.launch ssh commands ml-explore/mlx#3783
@shubhxho made their first contribution in Add math mode option for custom Metal kernels ml-explore/mlx#3728
@lyonsno made their first contribution in Fix compiled kernel correctness for negative-strided inputs ml-explore/mlx#3720
@danlee2002 made their first contribution in Fix infinite norm negative axis mismatch bug for matrices with 2 or more dimensions. ml-explore/mlx#3756
@ericphanson made their first contribution in Fix conv2 gradients in grouped strided case on Metal ml-explore/mlx#3800

Full Changelog: ml-explore/mlx@v0.31.2...v0.32.0

View the full release notes at https://github.com/ml-explore/mlx/releases/tag/v0.32.0.

---

[github-actions]

🤖 An automated task has requested bottles to be published to this PR.

Caution

Please do not push to this PR branch before the bottle commits have been pushed, as this results in a state that is difficult to recover from. If you need to resolve a merge conflict, please use a merge commit. Do not force-push to this PR branch.