mlx 0.27.1

[BrewTestBot]

Created by brew bump

---

Created with brew bump-formula-pr.

• resource blocks have been checked for updates.

release notes

## Highlights

Initial PyPi release of the CUDA back-end.
CUDA back-end works for well with mlx-lm:

Reasonably fast for LLM inference
Supports single-machine training and LoRA fine-tuning



What's Changed

Avoid invoking allocator::malloc when creating CUDA event by @zcbenz in Avoid invoking allocator::malloc when creating CUDA event ml-explore/mlx#2232
Share more common code in Compiled by @zcbenz in Share more common code in Compiled ml-explore/mlx#2240
Avoid atomic updates across CPU/GPU in CUDA event by @zcbenz in Avoid atomic updates across CPU/GPU in CUDA event ml-explore/mlx#2231
Perf regression fix by @angeloskath in Perf regression fix ml-explore/mlx#2243
Add profiler annotations in common primitives for CUDA backend by @zcbenz in Add profiler annotations in common primitives for CUDA backend ml-explore/mlx#2244
Default strict mode for module update and update_modules by @awni in Default strict mode for module update and update_modules ml-explore/mlx#2239
Fix linux linking error by @awni in Fix linux linking error ml-explore/mlx#2248
Improve metal elementwise kernels by @awni in Improve metal elementwise kernels ml-explore/mlx#2247
CUDA backend: matmul by @zcbenz in CUDA backend: matmul ml-explore/mlx#2241
Change layernorms to two pass algorithm by @angeloskath in Change layernorms to two pass algorithm ml-explore/mlx#2246
Fix unintuitive metal kernel caching by @awni in Fix unintuitive metal kernel caching ml-explore/mlx#2242
Refactor the lu test by @emmanuel-ferdman in Refactor the lu test ml-explore/mlx#2250
CUDA backend: unary ops by @zcbenz in CUDA backend: unary ops ml-explore/mlx#2158
Fix export to work with gather/scatter axis by @awni in Fix export to work with gather/scatter axis ml-explore/mlx#2263
CUDA backend: binary ops by @zcbenz in CUDA backend: binary ops ml-explore/mlx#2259
Report number of missing parameters by @FL33TW00D in Report number of missing parameters ml-explore/mlx#2264
CUDA backend: sort by @zcbenz in CUDA backend: sort ml-explore/mlx#2262
CUDA backend: random by @zcbenz in CUDA backend: random ml-explore/mlx#2261
Fix conv export by @awni in Fix conv export ml-explore/mlx#2265
CUDA backend: copy ops by @zcbenz in CUDA backend: copy ops ml-explore/mlx#2260
Fix building cpp benchmarks on Linux by @zcbenz in Fix building cpp benchmarks on Linux ml-explore/mlx#2268
Add load_safe to the general conv loaders by @angeloskath in Add load_safe to the general conv loaders ml-explore/mlx#2258
start cuda circle config by @awni in start cuda circle config ml-explore/mlx#2256
CUDA backend: reduce by @zcbenz in CUDA backend: reduce ml-explore/mlx#2269
CUDA backend: argreduce by @zcbenz in CUDA backend: argreduce ml-explore/mlx#2270
CUDA backend: softmax by @zcbenz in CUDA backend: softmax ml-explore/mlx#2272
CUDA backend: layernorm by @zcbenz in CUDA backend: layernorm ml-explore/mlx#2271
Fix warnings from latest CUDA toolkit by @zcbenz in Fix warnings from latest CUDA toolkit ml-explore/mlx#2275
Make sliceUpdate general by @awni in Make sliceUpdate general ml-explore/mlx#2282
CUDA backend: compile by @zcbenz in CUDA backend: compile ml-explore/mlx#2276
[CUDA] RMSNorm and VJP by @awni in [CUDA] RMSNorm and VJP ml-explore/mlx#2280
[CUDA] Fix build by @awni in [CUDA] Fix build ml-explore/mlx#2284
[CUDA] ternary with select op by @awni in [CUDA] ternary with select op ml-explore/mlx#2283
CUDA backend: indexing ops by @zcbenz in CUDA backend: indexing ops ml-explore/mlx#2277
Collection of refactors  by @jagrit06 in Collection of refactors  ml-explore/mlx#2274
Fix complex power and print by @awni in Fix complex power and print ml-explore/mlx#2286
fix cuda jit by @awni in fix cuda jit ml-explore/mlx#2287
Fix cuda gemm for bf16 by @awni in Fix cuda gemm for bf16 ml-explore/mlx#2288
Fix cuda arg reduce by @awni in Fix cuda arg reduce ml-explore/mlx#2291
RoPE for CUDA by @angeloskath in RoPE for CUDA ml-explore/mlx#2293
Add python testing for cuda with ability to skip list of tests by @awni in Add python testing for cuda with ability to skip list of tests ml-explore/mlx#2295
[CUDA] Fix back-end bugs and enable corresponding tests by @awni in [CUDA] Fix back-end bugs and enable corresponding tests ml-explore/mlx#2296
Cuda bug fixes 2 by @awni in Cuda bug fixes 2 ml-explore/mlx#2298
[CUDA] Divmod, Partition, and sort fixes by @awni in [CUDA] Divmod, Partition, and sort fixes ml-explore/mlx#2302
[CUDA] synch properly waits for all tasks to finish and clear by @awni in [CUDA] synch properly waits for all tasks to finish and clear ml-explore/mlx#2303
Make ptx cache settable by environment variable by @angeloskath in Make ptx cache settable by environment variable ml-explore/mlx#2304
Build CUDA release in Circle by @awni in Build CUDA release in Circle ml-explore/mlx#2306
Cuda perf tuning by @awni in Cuda perf tuning ml-explore/mlx#2307
Fix update_modules() when providing a subset by @angeloskath in Fix update_modules() when providing a subset ml-explore/mlx#2308
Compile float64 functions on CPU by @awni in Compile float64 functions on CPU ml-explore/mlx#2311
Fix get 2d grid dims by @angeloskath in Fix get 2d grid dims ml-explore/mlx#2316
Split broadcast so it is always fused in compile by @angeloskath in Split broadcast so it is always fused in compile ml-explore/mlx#2318
[CUDA] Fix reductions by @angeloskath in [CUDA] Fix reductions ml-explore/mlx#2314
Fix module update in strict mode by @awni in Fix module update in strict mode ml-explore/mlx#2321
MLX_SWITCH macros to templates by @angeloskath in MLX_SWITCH macros to templates ml-explore/mlx#2320
Use fp32 for testing, add more complex ops by @awni in Use fp32 for testing, add more complex ops ml-explore/mlx#2322
Patch bump by @awni in Patch bump ml-explore/mlx#2324
Allow parameters to be deleted from a module by @awni in Allow parameters to be deleted from a module ml-explore/mlx#2325
Fix compilation error from integral_constant by @zcbenz in Fix compilation error from integral_constant ml-explore/mlx#2326
[CUDA] Switch to CUDA graphs by @awni in [CUDA] Switch to CUDA graphs ml-explore/mlx#2317
[CUDA] Fix graphs for older cuda by @awni in [CUDA] Fix graphs for older cuda ml-explore/mlx#2328
[CUDA] Add MLX_CUDA_GRAPH_CACHE_SIZE env for setting graph cache size by @zcbenz in [CUDA] Add MLX_CUDA_GRAPH_CACHE_SIZE env for setting graph cache size ml-explore/mlx#2329
Fix layernorm race condition by @angeloskath in Fix layernorm race condition ml-explore/mlx#2340
Build with all cpu cores by default by @zcbenz in Build with all cpu cores by default ml-explore/mlx#2336
[CUDA] Do vectorized store/load in binary ops by @zcbenz in [CUDA] Do vectorized store/load in binary ops ml-explore/mlx#2330
Auto build linux release by @awni in Auto build linux release ml-explore/mlx#2341
MoE backward improvements by @angeloskath in MoE backward improvements ml-explore/mlx#2335
Fix compilation with CUDA 11 by @zcbenz in Fix compilation with CUDA 11 ml-explore/mlx#2331
patch bump by @awni in patch bump ml-explore/mlx#2343
Align mlx::core::max op nan propagation with NumPy by @jhavukainen in Align mlx::core::max op nan propagation with NumPy ml-explore/mlx#2339
Add zero for argsort vjp by @awni in Add zero for argsort vjp ml-explore/mlx#2345
[CUDA] Do vectorized store/load in contiguous elementwise ops by @zcbenz in [CUDA] Do vectorized store/load in contiguous elementwise ops ml-explore/mlx#2342
Align mlx::core::min op nan propagation with NumPy by @jhavukainen in Align mlx::core::min op nan propagation with NumPy ml-explore/mlx#2346
[CUDA] Set current device before cudaGraphLaunch by @zcbenz in [CUDA] Set current device before cudaGraphLaunch ml-explore/mlx#2351
[CUDA] Put version in ptx cache dir path by @zcbenz in [CUDA] Put version in ptx cache dir path ml-explore/mlx#2352
Fix type promotion in Adam with bias correction by @angeloskath in Fix type promotion in Adam with bias correction ml-explore/mlx#2350
Fix edge check in QuantizedBlockLoader for qmm_n by @angeloskath in Fix edge check in QuantizedBlockLoader for qmm_n ml-explore/mlx#2355
[CUDA] Implement Scan kernel by @zcbenz in [CUDA] Implement Scan kernel ml-explore/mlx#2347
[Metal] fix copy dispatch by @awni in [Metal] fix copy dispatch ml-explore/mlx#2360
[CUDA] Bundle CCCL for JIT compilation by @zcbenz in [CUDA] Bundle CCCL for JIT compilation ml-explore/mlx#2357
[CUDA] Do not put kernels in annoymous namespace by @zcbenz in [CUDA] Do not put kernels in annoymous namespace ml-explore/mlx#2362
Fix imag() vjp by @angeloskath in Fix imag() vjp ml-explore/mlx#2367
Add Primitive::name and remove Primitive::print by @zcbenz in Add Primitive::name and remove Primitive::print ml-explore/mlx#2365
update linux build by @awni in update linux build ml-explore/mlx#2370
[CUDA] Affine quantize by @awni in [CUDA] Affine quantize ml-explore/mlx#2354
Fix flaky linux test by @awni in Fix flaky linux test ml-explore/mlx#2371
Install linux with mlx[cuda] and mlx[cpu] by @awni in Install linux with mlx[cuda] and mlx[cpu] ml-explore/mlx#2356
[CUDA] Use cuda::std::complex in place of cuComplex by @zcbenz in [CUDA] Use cuda::std::complex in place of cuComplex ml-explore/mlx#2372
lower memory uniform sampling by @awni in lower memory uniform sampling ml-explore/mlx#2361
[CUDA] Fix complex reduce + nan propagation in min and max by @awni in [CUDA] Fix complex reduce + nan propagation in min and max ml-explore/mlx#2377
Rename the copy util in cpu/copy.h to copy_cpu by @zcbenz in Rename the copy util in cpu/copy.h to copy_cpu ml-explore/mlx#2378
fix ring distributed test by @awni in fix ring distributed test ml-explore/mlx#2380
Test with CUDA 12.2 by @awni in Test with CUDA 12.2 ml-explore/mlx#2375
[CUDA] Add work per thread to compile by @angeloskath in [CUDA] Add work per thread to compile ml-explore/mlx#2368
[CUDA] Fix resource leaks in matmul and graph by @awni in [CUDA] Fix resource leaks in matmul and graph ml-explore/mlx#2383
[CUDA] Add more ways finding CCCL headers in JIT by @zcbenz in [CUDA] Add more ways finding CCCL headers in JIT ml-explore/mlx#2382
Add contiguous_copy_gpu util for copying array by @zcbenz in Add contiguous_copy_gpu util for copying array ml-explore/mlx#2379
Adding support for the Muon Optimizer by @Goekdeniz-Guelmez in Adding support for the Muon Optimizer ml-explore/mlx#1914
Patch bump by @awni in Patch bump ml-explore/mlx#2386
Fix release build + patch bump by @awni in Fix release build + patch bump ml-explore/mlx#2387
Fix cuda manylinux version to match others by @awni in Fix cuda manylinux version to match others ml-explore/mlx#2388
[CUDA] speedup handling scalars by @awni in [CUDA] speedup handling scalars ml-explore/mlx#2389
Remove thrust iterators by @zcbenz in Remove thrust iterators ml-explore/mlx#2396
Add contiguous_copy_cpu util for copying array by @zcbenz in Add contiguous_copy_cpu util for copying array ml-explore/mlx#2397
Fix including stubs in wheel by @awni in Fix including stubs in wheel ml-explore/mlx#2398
use size option in binary by @awni in use size option in binary ml-explore/mlx#2399
[CUDA] Simplify allocator by @awni in [CUDA] Simplify allocator ml-explore/mlx#2392
Add cuda gemv by @awni in Add cuda gemv ml-explore/mlx#2400
Fix an error in the comment for mx.dequantize by @csukuangfj in Fix an error in the comment for mx.dequantize ml-explore/mlx#2409
Remove unused code in Convolution::vjp by @zcbenz in Remove unused code in Convolution::vjp ml-explore/mlx#2408
[CUDA] --compress-mode requires CUDA 12.8 by @zcbenz in [CUDA] --compress-mode requires CUDA 12.8 ml-explore/mlx#2407
full row mask in sdpa consistently gives nan by @awni in full row mask in sdpa consistently gives nan ml-explore/mlx#2406
Fix uv install and add dev release by @awni in Fix uv install and add dev release ml-explore/mlx#2411
[Metal] Release metal events by @awni in [Metal] Release metal events ml-explore/mlx#2412
Test on cuda 12.2 and 12.9 by @awni in Test on cuda 12.2 and 12.9 ml-explore/mlx#2413
[CUDA] Initial implementation of Convolution with cuDNN by @zcbenz in [CUDA] Initial implementation of Convolution with cuDNN ml-explore/mlx#2385
[DOCS]: Fix eps placement in Adam and AdamW by @Skonor in [DOCS]: Fix eps placement in Adam and AdamW ml-explore/mlx#2416
[CUDA] Always use batched matmul by @awni in [CUDA] Always use batched matmul ml-explore/mlx#2404
Fix qvm splitk by @awni in Fix qvm splitk ml-explore/mlx#2415
Update install docs and requirements by @awni in Update install docs and requirements ml-explore/mlx#2419
version by @awni in version ml-explore/mlx#2420

New Contributors

@emmanuel-ferdman made their first contribution in Refactor the lu test ml-explore/mlx#2250
@FL33TW00D made their first contribution in Report number of missing parameters ml-explore/mlx#2264
@jhavukainen made their first contribution in Align mlx::core::max op nan propagation with NumPy ml-explore/mlx#2339
@Goekdeniz-Guelmez made their first contribution in Adding support for the Muon Optimizer ml-explore/mlx#1914
@Skonor made their first contribution in [DOCS]: Fix eps placement in Adam and AdamW ml-explore/mlx#2416

Full Changelog: ml-explore/mlx@v0.26.0...v0.27.0

View the full release notes at https://github.com/ml-explore/mlx/releases/tag/v0.27.1.

---

[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.