mlx 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Potentially problematic release.


This version of mlx might be problematic. Click here for more details.

Files changed (914) hide show
  1. checksums.yaml +7 -0
  2. data/ext/mlx/CMakeLists.txt +7 -0
  3. data/ext/mlx/Makefile +273 -0
  4. data/ext/mlx/extconf.rb +94 -0
  5. data/ext/mlx/mkmf.log +44 -0
  6. data/ext/mlx/native.bundle +0 -0
  7. data/ext/mlx/native.bundle.dSYM/Contents/Info.plist +20 -0
  8. data/ext/mlx/native.bundle.dSYM/Contents/Resources/DWARF/native.bundle +0 -0
  9. data/ext/mlx/native.bundle.dSYM/Contents/Resources/Relocations/aarch64/native.bundle.yml +5 -0
  10. data/ext/mlx/native.cpp +8027 -0
  11. data/ext/mlx/native.o +0 -0
  12. data/lib/mlx/core.rb +1678 -0
  13. data/lib/mlx/distributed_utils/common.rb +116 -0
  14. data/lib/mlx/distributed_utils/config.rb +600 -0
  15. data/lib/mlx/distributed_utils/launch.rb +490 -0
  16. data/lib/mlx/extension.rb +24 -0
  17. data/lib/mlx/nn/base.rb +388 -0
  18. data/lib/mlx/nn/init.rb +140 -0
  19. data/lib/mlx/nn/layers/activations.rb +336 -0
  20. data/lib/mlx/nn/layers/base.rb +6 -0
  21. data/lib/mlx/nn/layers/containers.rb +20 -0
  22. data/lib/mlx/nn/layers/convolution.rb +120 -0
  23. data/lib/mlx/nn/layers/convolution_transpose.rb +114 -0
  24. data/lib/mlx/nn/layers/distributed.rb +309 -0
  25. data/lib/mlx/nn/layers/dropout.rb +75 -0
  26. data/lib/mlx/nn/layers/embedding.rb +28 -0
  27. data/lib/mlx/nn/layers/linear.rb +79 -0
  28. data/lib/mlx/nn/layers/normalization.rb +216 -0
  29. data/lib/mlx/nn/layers/pooling.rb +167 -0
  30. data/lib/mlx/nn/layers/positional_encoding.rb +126 -0
  31. data/lib/mlx/nn/layers/quantized.rb +215 -0
  32. data/lib/mlx/nn/layers/recurrent.rb +135 -0
  33. data/lib/mlx/nn/layers/transformer.rb +330 -0
  34. data/lib/mlx/nn/layers/upsample.rb +97 -0
  35. data/lib/mlx/nn/layers.rb +18 -0
  36. data/lib/mlx/nn/losses.rb +251 -0
  37. data/lib/mlx/nn/utils.rb +167 -0
  38. data/lib/mlx/nn.rb +12 -0
  39. data/lib/mlx/optimizers/optimizers.rb +808 -0
  40. data/lib/mlx/optimizers/schedulers.rb +62 -0
  41. data/lib/mlx/optimizers.rb +9 -0
  42. data/lib/mlx/utils.rb +171 -0
  43. data/lib/mlx/version +1 -0
  44. data/lib/mlx/version.rb +5 -0
  45. data/lib/mlx.rb +64 -0
  46. data/mlx/.clang-format +87 -0
  47. data/mlx/.git +1 -0
  48. data/mlx/.github/ISSUE_TEMPLATE/bug_report.md +28 -0
  49. data/mlx/.github/actions/build-cuda-release/action.yml +31 -0
  50. data/mlx/.github/actions/build-docs/action.yml +38 -0
  51. data/mlx/.github/actions/build-linux/action.yml +38 -0
  52. data/mlx/.github/actions/build-linux-release/action.yml +42 -0
  53. data/mlx/.github/actions/build-macos/action.yml +80 -0
  54. data/mlx/.github/actions/build-macos-release/action.yml +36 -0
  55. data/mlx/.github/actions/build-windows/action.yml +26 -0
  56. data/mlx/.github/actions/setup-linux/action.yml +93 -0
  57. data/mlx/.github/actions/setup-macos/action.yml +24 -0
  58. data/mlx/.github/actions/setup-windows/action.yml +42 -0
  59. data/mlx/.github/actions/test-linux/action.yml +69 -0
  60. data/mlx/.github/actions/test-windows/action.yml +20 -0
  61. data/mlx/.github/dependabot.yml +6 -0
  62. data/mlx/.github/pull_request_template.md +12 -0
  63. data/mlx/.github/scripts/build-sanitizer-tests.sh +48 -0
  64. data/mlx/.github/scripts/setup+build-cpp-linux-fedora-container.sh +27 -0
  65. data/mlx/.github/workflows/build_and_test.yml +152 -0
  66. data/mlx/.github/workflows/documentation.yml +28 -0
  67. data/mlx/.github/workflows/nightly.yml +104 -0
  68. data/mlx/.github/workflows/release.yml +256 -0
  69. data/mlx/.gitignore +81 -0
  70. data/mlx/.pre-commit-config.yaml +27 -0
  71. data/mlx/ACKNOWLEDGMENTS.md +268 -0
  72. data/mlx/CITATION.cff +24 -0
  73. data/mlx/CMakeLists.txt +437 -0
  74. data/mlx/CODE_OF_CONDUCT.md +132 -0
  75. data/mlx/CONTRIBUTING.md +38 -0
  76. data/mlx/LICENSE +21 -0
  77. data/mlx/MANIFEST.in +6 -0
  78. data/mlx/README.md +121 -0
  79. data/mlx/benchmarks/cpp/CMakeLists.txt +11 -0
  80. data/mlx/benchmarks/cpp/autograd.cpp +39 -0
  81. data/mlx/benchmarks/cpp/compare_devices.cpp +27 -0
  82. data/mlx/benchmarks/cpp/irregular_strides.cpp +201 -0
  83. data/mlx/benchmarks/cpp/single_ops.cpp +288 -0
  84. data/mlx/benchmarks/cpp/time_utils.h +39 -0
  85. data/mlx/benchmarks/numpy/single_ops.py +39 -0
  86. data/mlx/benchmarks/numpy/time_utils.py +20 -0
  87. data/mlx/benchmarks/python/batch_matmul_bench.py +62 -0
  88. data/mlx/benchmarks/python/blas/bench_gemm.py +191 -0
  89. data/mlx/benchmarks/python/blas/bench_gemv.py +220 -0
  90. data/mlx/benchmarks/python/comparative/README.md +15 -0
  91. data/mlx/benchmarks/python/comparative/bench_mlx.py +519 -0
  92. data/mlx/benchmarks/python/comparative/bench_torch.py +482 -0
  93. data/mlx/benchmarks/python/comparative/compare.py +284 -0
  94. data/mlx/benchmarks/python/compile_bench.py +107 -0
  95. data/mlx/benchmarks/python/conv1d_bench.py +123 -0
  96. data/mlx/benchmarks/python/conv2d_bench_cpu.py +127 -0
  97. data/mlx/benchmarks/python/conv2d_train_bench_cpu.py +143 -0
  98. data/mlx/benchmarks/python/conv2d_transpose_bench_cpu.py +129 -0
  99. data/mlx/benchmarks/python/conv3d_bench_cpu.py +110 -0
  100. data/mlx/benchmarks/python/conv3d_train_bench_cpu.py +143 -0
  101. data/mlx/benchmarks/python/conv3d_transpose_bench_cpu.py +116 -0
  102. data/mlx/benchmarks/python/conv_bench.py +135 -0
  103. data/mlx/benchmarks/python/conv_transpose_bench.py +135 -0
  104. data/mlx/benchmarks/python/conv_unaligned_bench.py +107 -0
  105. data/mlx/benchmarks/python/distributed_bench.py +66 -0
  106. data/mlx/benchmarks/python/einsum_bench.py +84 -0
  107. data/mlx/benchmarks/python/fft_bench.py +118 -0
  108. data/mlx/benchmarks/python/gather_bench.py +52 -0
  109. data/mlx/benchmarks/python/gather_mm_bench.py +74 -0
  110. data/mlx/benchmarks/python/gather_qmm_bench.py +84 -0
  111. data/mlx/benchmarks/python/hadamard_bench.py +70 -0
  112. data/mlx/benchmarks/python/large_gemm_bench.py +119 -0
  113. data/mlx/benchmarks/python/layer_norm_bench.py +82 -0
  114. data/mlx/benchmarks/python/masked_scatter.py +212 -0
  115. data/mlx/benchmarks/python/rms_norm_bench.py +63 -0
  116. data/mlx/benchmarks/python/rope_bench.py +35 -0
  117. data/mlx/benchmarks/python/scatter_bench.py +96 -0
  118. data/mlx/benchmarks/python/sdpa_bench.py +223 -0
  119. data/mlx/benchmarks/python/sdpa_vector_bench.py +95 -0
  120. data/mlx/benchmarks/python/single_ops.py +132 -0
  121. data/mlx/benchmarks/python/synchronize_bench.py +55 -0
  122. data/mlx/benchmarks/python/time_utils.py +38 -0
  123. data/mlx/cmake/FindCUDNN.cmake +177 -0
  124. data/mlx/cmake/FindNCCL.cmake +54 -0
  125. data/mlx/cmake/Findnvpl.cmake +3 -0
  126. data/mlx/cmake/extension.cmake +50 -0
  127. data/mlx/docs/.clang-format +2 -0
  128. data/mlx/docs/.gitignore +3 -0
  129. data/mlx/docs/.nojekyll +0 -0
  130. data/mlx/docs/Doxyfile +51 -0
  131. data/mlx/docs/Makefile +18 -0
  132. data/mlx/docs/README.md +54 -0
  133. data/mlx/docs/index.html +1 -0
  134. data/mlx/docs/requirements.txt +5 -0
  135. data/mlx/docs/src/_static/distributed/m3-ultra-mesh-broken.png +0 -0
  136. data/mlx/docs/src/_static/distributed/m3-ultra-mesh.png +0 -0
  137. data/mlx/docs/src/_static/metal_debugger/capture.png +0 -0
  138. data/mlx/docs/src/_static/metal_debugger/schema.png +0 -0
  139. data/mlx/docs/src/_static/mlx_logo.png +0 -0
  140. data/mlx/docs/src/_static/mlx_logo_dark.png +0 -0
  141. data/mlx/docs/src/_static/tp_inference/all-to-sharded-linear.png +0 -0
  142. data/mlx/docs/src/_static/tp_inference/column-row-tp.png +0 -0
  143. data/mlx/docs/src/_static/tp_inference/llama-transformer.png +0 -0
  144. data/mlx/docs/src/_static/tp_inference/sharded-to-all-linear.png +0 -0
  145. data/mlx/docs/src/_templates/module-base-class.rst +33 -0
  146. data/mlx/docs/src/_templates/nn-module-template.rst +20 -0
  147. data/mlx/docs/src/_templates/optimizers-template.rst +20 -0
  148. data/mlx/docs/src/conf.py +99 -0
  149. data/mlx/docs/src/cpp/ops.rst +7 -0
  150. data/mlx/docs/src/dev/custom_metal_kernels.rst +445 -0
  151. data/mlx/docs/src/dev/extensions.rst +811 -0
  152. data/mlx/docs/src/dev/metal_debugger.rst +68 -0
  153. data/mlx/docs/src/dev/metal_logging.rst +40 -0
  154. data/mlx/docs/src/dev/mlx_in_cpp.rst +121 -0
  155. data/mlx/docs/src/examples/data_parallelism.rst +91 -0
  156. data/mlx/docs/src/examples/linear_regression.rst +77 -0
  157. data/mlx/docs/src/examples/llama-inference.rst +382 -0
  158. data/mlx/docs/src/examples/mlp.rst +134 -0
  159. data/mlx/docs/src/examples/tensor_parallelism.rst +239 -0
  160. data/mlx/docs/src/index.rst +96 -0
  161. data/mlx/docs/src/install.rst +340 -0
  162. data/mlx/docs/src/python/array.rst +65 -0
  163. data/mlx/docs/src/python/cuda.rst +9 -0
  164. data/mlx/docs/src/python/data_types.rst +78 -0
  165. data/mlx/docs/src/python/devices_and_streams.rst +21 -0
  166. data/mlx/docs/src/python/distributed.rst +22 -0
  167. data/mlx/docs/src/python/export.rst +14 -0
  168. data/mlx/docs/src/python/fast.rst +16 -0
  169. data/mlx/docs/src/python/fft.rst +24 -0
  170. data/mlx/docs/src/python/linalg.rst +27 -0
  171. data/mlx/docs/src/python/memory_management.rst +16 -0
  172. data/mlx/docs/src/python/metal.rst +12 -0
  173. data/mlx/docs/src/python/nn/distributed.rst +30 -0
  174. data/mlx/docs/src/python/nn/functions.rst +40 -0
  175. data/mlx/docs/src/python/nn/init.rst +45 -0
  176. data/mlx/docs/src/python/nn/layers.rst +74 -0
  177. data/mlx/docs/src/python/nn/losses.rst +25 -0
  178. data/mlx/docs/src/python/nn/module.rst +38 -0
  179. data/mlx/docs/src/python/nn.rst +186 -0
  180. data/mlx/docs/src/python/ops.rst +184 -0
  181. data/mlx/docs/src/python/optimizers/common_optimizers.rst +22 -0
  182. data/mlx/docs/src/python/optimizers/optimizer.rst +23 -0
  183. data/mlx/docs/src/python/optimizers/schedulers.rst +15 -0
  184. data/mlx/docs/src/python/optimizers.rst +78 -0
  185. data/mlx/docs/src/python/random.rst +48 -0
  186. data/mlx/docs/src/python/transforms.rst +22 -0
  187. data/mlx/docs/src/python/tree_utils.rst +23 -0
  188. data/mlx/docs/src/usage/compile.rst +516 -0
  189. data/mlx/docs/src/usage/distributed.rst +572 -0
  190. data/mlx/docs/src/usage/export.rst +288 -0
  191. data/mlx/docs/src/usage/function_transforms.rst +191 -0
  192. data/mlx/docs/src/usage/indexing.rst +194 -0
  193. data/mlx/docs/src/usage/launching_distributed.rst +234 -0
  194. data/mlx/docs/src/usage/lazy_evaluation.rst +144 -0
  195. data/mlx/docs/src/usage/numpy.rst +124 -0
  196. data/mlx/docs/src/usage/quick_start.rst +67 -0
  197. data/mlx/docs/src/usage/saving_and_loading.rst +81 -0
  198. data/mlx/docs/src/usage/unified_memory.rst +78 -0
  199. data/mlx/docs/src/usage/using_streams.rst +18 -0
  200. data/mlx/examples/cmake_project/CMakeLists.txt +22 -0
  201. data/mlx/examples/cmake_project/README.md +26 -0
  202. data/mlx/examples/cmake_project/example.cpp +14 -0
  203. data/mlx/examples/cpp/CMakeLists.txt +12 -0
  204. data/mlx/examples/cpp/distributed.cpp +22 -0
  205. data/mlx/examples/cpp/linear_regression.cpp +54 -0
  206. data/mlx/examples/cpp/logistic_regression.cpp +54 -0
  207. data/mlx/examples/cpp/metal_capture.cpp +31 -0
  208. data/mlx/examples/cpp/timer.h +20 -0
  209. data/mlx/examples/cpp/tutorial.cpp +99 -0
  210. data/mlx/examples/export/CMakeLists.txt +22 -0
  211. data/mlx/examples/export/README.md +49 -0
  212. data/mlx/examples/export/eval_mlp.cpp +25 -0
  213. data/mlx/examples/export/eval_mlp.py +52 -0
  214. data/mlx/examples/export/train_mlp.cpp +35 -0
  215. data/mlx/examples/export/train_mlp.py +76 -0
  216. data/mlx/examples/extensions/CMakeLists.txt +78 -0
  217. data/mlx/examples/extensions/README.md +24 -0
  218. data/mlx/examples/extensions/axpby/axpby.cpp +306 -0
  219. data/mlx/examples/extensions/axpby/axpby.h +90 -0
  220. data/mlx/examples/extensions/axpby/axpby.metal +47 -0
  221. data/mlx/examples/extensions/bindings.cpp +39 -0
  222. data/mlx/examples/extensions/mlx_sample_extensions/__init__.py +5 -0
  223. data/mlx/examples/extensions/pyproject.toml +8 -0
  224. data/mlx/examples/extensions/requirements.txt +4 -0
  225. data/mlx/examples/extensions/setup.py +18 -0
  226. data/mlx/examples/extensions/test.py +12 -0
  227. data/mlx/examples/python/linear_regression.py +46 -0
  228. data/mlx/examples/python/logistic_regression.py +49 -0
  229. data/mlx/examples/python/qqmm.py +117 -0
  230. data/mlx/mlx/3rdparty/.clang-format +2 -0
  231. data/mlx/mlx/3rdparty/pocketfft.h +3581 -0
  232. data/mlx/mlx/CMakeLists.txt +107 -0
  233. data/mlx/mlx/allocator.h +75 -0
  234. data/mlx/mlx/api.h +29 -0
  235. data/mlx/mlx/array.cpp +354 -0
  236. data/mlx/mlx/array.h +647 -0
  237. data/mlx/mlx/backend/common/CMakeLists.txt +9 -0
  238. data/mlx/mlx/backend/common/binary.h +97 -0
  239. data/mlx/mlx/backend/common/broadcasting.cpp +24 -0
  240. data/mlx/mlx/backend/common/broadcasting.h +11 -0
  241. data/mlx/mlx/backend/common/buffer_cache.h +158 -0
  242. data/mlx/mlx/backend/common/common.cpp +305 -0
  243. data/mlx/mlx/backend/common/compiled.cpp +243 -0
  244. data/mlx/mlx/backend/common/compiled.h +77 -0
  245. data/mlx/mlx/backend/common/copy.h +50 -0
  246. data/mlx/mlx/backend/common/hadamard.h +109 -0
  247. data/mlx/mlx/backend/common/load.cpp +57 -0
  248. data/mlx/mlx/backend/common/matmul.h +67 -0
  249. data/mlx/mlx/backend/common/reduce.cpp +154 -0
  250. data/mlx/mlx/backend/common/reduce.h +59 -0
  251. data/mlx/mlx/backend/common/slicing.cpp +71 -0
  252. data/mlx/mlx/backend/common/slicing.h +20 -0
  253. data/mlx/mlx/backend/common/ternary.h +85 -0
  254. data/mlx/mlx/backend/common/unary.h +29 -0
  255. data/mlx/mlx/backend/common/utils.cpp +231 -0
  256. data/mlx/mlx/backend/common/utils.h +205 -0
  257. data/mlx/mlx/backend/cpu/CMakeLists.txt +88 -0
  258. data/mlx/mlx/backend/cpu/arange.h +28 -0
  259. data/mlx/mlx/backend/cpu/arg_reduce.cpp +124 -0
  260. data/mlx/mlx/backend/cpu/binary.cpp +269 -0
  261. data/mlx/mlx/backend/cpu/binary.h +517 -0
  262. data/mlx/mlx/backend/cpu/binary_ops.h +98 -0
  263. data/mlx/mlx/backend/cpu/binary_two.h +166 -0
  264. data/mlx/mlx/backend/cpu/cholesky.cpp +85 -0
  265. data/mlx/mlx/backend/cpu/compiled.cpp +357 -0
  266. data/mlx/mlx/backend/cpu/compiled_preamble.h +12 -0
  267. data/mlx/mlx/backend/cpu/conv.cpp +1351 -0
  268. data/mlx/mlx/backend/cpu/copy.cpp +386 -0
  269. data/mlx/mlx/backend/cpu/copy.h +36 -0
  270. data/mlx/mlx/backend/cpu/device_info.cpp +113 -0
  271. data/mlx/mlx/backend/cpu/device_info.h +28 -0
  272. data/mlx/mlx/backend/cpu/distributed.cpp +103 -0
  273. data/mlx/mlx/backend/cpu/eig.cpp +281 -0
  274. data/mlx/mlx/backend/cpu/eigh.cpp +241 -0
  275. data/mlx/mlx/backend/cpu/encoder.cpp +16 -0
  276. data/mlx/mlx/backend/cpu/encoder.h +67 -0
  277. data/mlx/mlx/backend/cpu/eval.cpp +40 -0
  278. data/mlx/mlx/backend/cpu/eval.h +12 -0
  279. data/mlx/mlx/backend/cpu/fft.cpp +120 -0
  280. data/mlx/mlx/backend/cpu/gemm.h +26 -0
  281. data/mlx/mlx/backend/cpu/gemms/bnns.cpp +214 -0
  282. data/mlx/mlx/backend/cpu/gemms/cblas.cpp +134 -0
  283. data/mlx/mlx/backend/cpu/gemms/simd_bf16.cpp +45 -0
  284. data/mlx/mlx/backend/cpu/gemms/simd_fp16.cpp +45 -0
  285. data/mlx/mlx/backend/cpu/gemms/simd_gemm.h +139 -0
  286. data/mlx/mlx/backend/cpu/hadamard.cpp +121 -0
  287. data/mlx/mlx/backend/cpu/indexing.cpp +854 -0
  288. data/mlx/mlx/backend/cpu/inverse.cpp +160 -0
  289. data/mlx/mlx/backend/cpu/jit_compiler.cpp +166 -0
  290. data/mlx/mlx/backend/cpu/jit_compiler.h +20 -0
  291. data/mlx/mlx/backend/cpu/lapack.h +80 -0
  292. data/mlx/mlx/backend/cpu/logsumexp.cpp +139 -0
  293. data/mlx/mlx/backend/cpu/luf.cpp +120 -0
  294. data/mlx/mlx/backend/cpu/make_compiled_preamble.ps1 +38 -0
  295. data/mlx/mlx/backend/cpu/make_compiled_preamble.sh +41 -0
  296. data/mlx/mlx/backend/cpu/masked_mm.cpp +608 -0
  297. data/mlx/mlx/backend/cpu/matmul.cpp +166 -0
  298. data/mlx/mlx/backend/cpu/primitives.cpp +478 -0
  299. data/mlx/mlx/backend/cpu/qrf.cpp +147 -0
  300. data/mlx/mlx/backend/cpu/quantized.cpp +1370 -0
  301. data/mlx/mlx/backend/cpu/reduce.cpp +587 -0
  302. data/mlx/mlx/backend/cpu/scan.cpp +338 -0
  303. data/mlx/mlx/backend/cpu/select.cpp +95 -0
  304. data/mlx/mlx/backend/cpu/simd/accelerate_fp16_simd.h +56 -0
  305. data/mlx/mlx/backend/cpu/simd/accelerate_simd.h +329 -0
  306. data/mlx/mlx/backend/cpu/simd/base_simd.h +319 -0
  307. data/mlx/mlx/backend/cpu/simd/math.h +193 -0
  308. data/mlx/mlx/backend/cpu/simd/neon_fp16_simd.h +212 -0
  309. data/mlx/mlx/backend/cpu/simd/simd.h +4 -0
  310. data/mlx/mlx/backend/cpu/simd/type.h +11 -0
  311. data/mlx/mlx/backend/cpu/slicing.h +21 -0
  312. data/mlx/mlx/backend/cpu/softmax.cpp +170 -0
  313. data/mlx/mlx/backend/cpu/sort.cpp +481 -0
  314. data/mlx/mlx/backend/cpu/svd.cpp +289 -0
  315. data/mlx/mlx/backend/cpu/ternary.h +154 -0
  316. data/mlx/mlx/backend/cpu/threefry.cpp +31 -0
  317. data/mlx/mlx/backend/cpu/threefry.h +21 -0
  318. data/mlx/mlx/backend/cpu/unary.cpp +238 -0
  319. data/mlx/mlx/backend/cpu/unary.h +281 -0
  320. data/mlx/mlx/backend/cpu/unary_ops.h +175 -0
  321. data/mlx/mlx/backend/cuda/CMakeLists.txt +265 -0
  322. data/mlx/mlx/backend/cuda/allocator.cpp +451 -0
  323. data/mlx/mlx/backend/cuda/allocator.h +94 -0
  324. data/mlx/mlx/backend/cuda/arange.cu +68 -0
  325. data/mlx/mlx/backend/cuda/arg_reduce.cu +189 -0
  326. data/mlx/mlx/backend/cuda/bin2h.cmake +150 -0
  327. data/mlx/mlx/backend/cuda/binary/CMakeLists.txt +21 -0
  328. data/mlx/mlx/backend/cuda/binary/add.cu +7 -0
  329. data/mlx/mlx/backend/cuda/binary/arctan2.cu +7 -0
  330. data/mlx/mlx/backend/cuda/binary/binary.cuh +383 -0
  331. data/mlx/mlx/backend/cuda/binary/bitwise_binary.cu +27 -0
  332. data/mlx/mlx/backend/cuda/binary/divide.cu +7 -0
  333. data/mlx/mlx/backend/cuda/binary/equal.cu +15 -0
  334. data/mlx/mlx/backend/cuda/binary/greater.cu +7 -0
  335. data/mlx/mlx/backend/cuda/binary/greater_equal.cu +7 -0
  336. data/mlx/mlx/backend/cuda/binary/less.cu +7 -0
  337. data/mlx/mlx/backend/cuda/binary/less_equal.cu +7 -0
  338. data/mlx/mlx/backend/cuda/binary/log_add_exp.cu +7 -0
  339. data/mlx/mlx/backend/cuda/binary/logical_and.cu +7 -0
  340. data/mlx/mlx/backend/cuda/binary/logical_or.cu +7 -0
  341. data/mlx/mlx/backend/cuda/binary/maximum.cu +7 -0
  342. data/mlx/mlx/backend/cuda/binary/minimum.cu +7 -0
  343. data/mlx/mlx/backend/cuda/binary/multiply.cu +7 -0
  344. data/mlx/mlx/backend/cuda/binary/not_equal.cu +7 -0
  345. data/mlx/mlx/backend/cuda/binary/power.cu +7 -0
  346. data/mlx/mlx/backend/cuda/binary/remainder.cu +7 -0
  347. data/mlx/mlx/backend/cuda/binary/subtract.cu +7 -0
  348. data/mlx/mlx/backend/cuda/binary_two.cu +412 -0
  349. data/mlx/mlx/backend/cuda/compiled.cpp +357 -0
  350. data/mlx/mlx/backend/cuda/conv/conv.h +126 -0
  351. data/mlx/mlx/backend/cuda/conv/gemm_conv.cu +217 -0
  352. data/mlx/mlx/backend/cuda/conv/gemm_grouped_conv.cu +231 -0
  353. data/mlx/mlx/backend/cuda/conv.cpp +403 -0
  354. data/mlx/mlx/backend/cuda/copy/copy.cuh +55 -0
  355. data/mlx/mlx/backend/cuda/copy/copy_contiguous.cu +88 -0
  356. data/mlx/mlx/backend/cuda/copy/copy_general.cu +171 -0
  357. data/mlx/mlx/backend/cuda/copy/copy_general_dynamic.cu +118 -0
  358. data/mlx/mlx/backend/cuda/copy/copy_general_input.cu +229 -0
  359. data/mlx/mlx/backend/cuda/copy.cu +132 -0
  360. data/mlx/mlx/backend/cuda/cublas_utils.cpp +222 -0
  361. data/mlx/mlx/backend/cuda/cublas_utils.h +95 -0
  362. data/mlx/mlx/backend/cuda/cuda.h +21 -0
  363. data/mlx/mlx/backend/cuda/cuda_utils.h +90 -0
  364. data/mlx/mlx/backend/cuda/cudnn_utils.cpp +133 -0
  365. data/mlx/mlx/backend/cuda/cudnn_utils.h +187 -0
  366. data/mlx/mlx/backend/cuda/custom_kernel.cpp +379 -0
  367. data/mlx/mlx/backend/cuda/cutlass_utils.cuh +46 -0
  368. data/mlx/mlx/backend/cuda/delayload.cpp +80 -0
  369. data/mlx/mlx/backend/cuda/device/atomic_ops.cuh +63 -0
  370. data/mlx/mlx/backend/cuda/device/binary_ops.cuh +300 -0
  371. data/mlx/mlx/backend/cuda/device/cast_op.cuh +118 -0
  372. data/mlx/mlx/backend/cuda/device/complex.cuh +60 -0
  373. data/mlx/mlx/backend/cuda/device/config.h +12 -0
  374. data/mlx/mlx/backend/cuda/device/fp16_math.cuh +96 -0
  375. data/mlx/mlx/backend/cuda/device/gather.cuh +53 -0
  376. data/mlx/mlx/backend/cuda/device/gather_axis.cuh +65 -0
  377. data/mlx/mlx/backend/cuda/device/indexing.cuh +30 -0
  378. data/mlx/mlx/backend/cuda/device/scatter.cuh +68 -0
  379. data/mlx/mlx/backend/cuda/device/scatter_axis.cuh +67 -0
  380. data/mlx/mlx/backend/cuda/device/scatter_ops.cuh +44 -0
  381. data/mlx/mlx/backend/cuda/device/ternary_ops.cuh +13 -0
  382. data/mlx/mlx/backend/cuda/device/unary_ops.cuh +350 -0
  383. data/mlx/mlx/backend/cuda/device/utils.cuh +464 -0
  384. data/mlx/mlx/backend/cuda/device.cpp +522 -0
  385. data/mlx/mlx/backend/cuda/device.h +195 -0
  386. data/mlx/mlx/backend/cuda/device_info.cpp +232 -0
  387. data/mlx/mlx/backend/cuda/distributed.cu +121 -0
  388. data/mlx/mlx/backend/cuda/eval.cpp +66 -0
  389. data/mlx/mlx/backend/cuda/event.cu +415 -0
  390. data/mlx/mlx/backend/cuda/event.h +79 -0
  391. data/mlx/mlx/backend/cuda/fence.cpp +42 -0
  392. data/mlx/mlx/backend/cuda/gemms/cublas_gemm.cpp +233 -0
  393. data/mlx/mlx/backend/cuda/gemms/cublas_gemm.h +114 -0
  394. data/mlx/mlx/backend/cuda/gemms/cublas_gemm_batched_12_0.cpp +77 -0
  395. data/mlx/mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu +329 -0
  396. data/mlx/mlx/backend/cuda/gemms/gemv.cu +327 -0
  397. data/mlx/mlx/backend/cuda/gemms/gemv.h +34 -0
  398. data/mlx/mlx/backend/cuda/gemms/grouped_gemm.h +25 -0
  399. data/mlx/mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu +358 -0
  400. data/mlx/mlx/backend/cuda/indexing.cpp +434 -0
  401. data/mlx/mlx/backend/cuda/jit_module.cpp +443 -0
  402. data/mlx/mlx/backend/cuda/jit_module.h +120 -0
  403. data/mlx/mlx/backend/cuda/kernel_utils.cu +52 -0
  404. data/mlx/mlx/backend/cuda/kernel_utils.cuh +148 -0
  405. data/mlx/mlx/backend/cuda/layer_norm.cu +417 -0
  406. data/mlx/mlx/backend/cuda/load.cpp +60 -0
  407. data/mlx/mlx/backend/cuda/logsumexp.cu +161 -0
  408. data/mlx/mlx/backend/cuda/lru_cache.h +190 -0
  409. data/mlx/mlx/backend/cuda/matmul.cpp +373 -0
  410. data/mlx/mlx/backend/cuda/no_cuda.cpp +47 -0
  411. data/mlx/mlx/backend/cuda/primitives.cpp +46 -0
  412. data/mlx/mlx/backend/cuda/quantized/affine_quantize.cu +329 -0
  413. data/mlx/mlx/backend/cuda/quantized/convert_fp8.cu +19 -0
  414. data/mlx/mlx/backend/cuda/quantized/cublas_qqmm.cpp +206 -0
  415. data/mlx/mlx/backend/cuda/quantized/cublas_qqmm.h +88 -0
  416. data/mlx/mlx/backend/cuda/quantized/cuda_fp4.h +100 -0
  417. data/mlx/mlx/backend/cuda/quantized/fp_quantize.cu +496 -0
  418. data/mlx/mlx/backend/cuda/quantized/mxfp8_quantize.cuh +32 -0
  419. data/mlx/mlx/backend/cuda/quantized/no_qqmm_impl.cpp +26 -0
  420. data/mlx/mlx/backend/cuda/quantized/nvfp4_quantize.cuh +334 -0
  421. data/mlx/mlx/backend/cuda/quantized/qmv.cu +304 -0
  422. data/mlx/mlx/backend/cuda/quantized/qmv.h +21 -0
  423. data/mlx/mlx/backend/cuda/quantized/qqmm.cpp +158 -0
  424. data/mlx/mlx/backend/cuda/quantized/qqmm_impl.cpp +50 -0
  425. data/mlx/mlx/backend/cuda/quantized/qqmm_impl.h +26 -0
  426. data/mlx/mlx/backend/cuda/quantized/qqmm_utils.cu +227 -0
  427. data/mlx/mlx/backend/cuda/quantized/qqmm_utils.h +30 -0
  428. data/mlx/mlx/backend/cuda/quantized/quantized.cpp +85 -0
  429. data/mlx/mlx/backend/cuda/quantized/quantized.h +53 -0
  430. data/mlx/mlx/backend/cuda/quantized/quantized_utils.cuh +88 -0
  431. data/mlx/mlx/backend/cuda/quantized/quantized_utils.h +50 -0
  432. data/mlx/mlx/backend/cuda/random.cu +202 -0
  433. data/mlx/mlx/backend/cuda/reduce/all_reduce.cu +159 -0
  434. data/mlx/mlx/backend/cuda/reduce/col_reduce.cu +510 -0
  435. data/mlx/mlx/backend/cuda/reduce/init_reduce.cu +50 -0
  436. data/mlx/mlx/backend/cuda/reduce/reduce.cuh +71 -0
  437. data/mlx/mlx/backend/cuda/reduce/reduce_ops.cuh +211 -0
  438. data/mlx/mlx/backend/cuda/reduce/reduce_utils.cuh +145 -0
  439. data/mlx/mlx/backend/cuda/reduce/row_reduce.cu +361 -0
  440. data/mlx/mlx/backend/cuda/reduce.cu +73 -0
  441. data/mlx/mlx/backend/cuda/rms_norm.cu +536 -0
  442. data/mlx/mlx/backend/cuda/rope.cu +429 -0
  443. data/mlx/mlx/backend/cuda/scaled_dot_product_attention.cpp +681 -0
  444. data/mlx/mlx/backend/cuda/scaled_dot_product_attention.cu +796 -0
  445. data/mlx/mlx/backend/cuda/scan.cu +468 -0
  446. data/mlx/mlx/backend/cuda/slicing.cpp +111 -0
  447. data/mlx/mlx/backend/cuda/softmax.cu +162 -0
  448. data/mlx/mlx/backend/cuda/sort.cu +1076 -0
  449. data/mlx/mlx/backend/cuda/steel/defines.cuh +9 -0
  450. data/mlx/mlx/backend/cuda/steel/gemm.cuh +101 -0
  451. data/mlx/mlx/backend/cuda/steel/mma.cuh +117 -0
  452. data/mlx/mlx/backend/cuda/steel/tiles.cuh +450 -0
  453. data/mlx/mlx/backend/cuda/steel/utils.cuh +89 -0
  454. data/mlx/mlx/backend/cuda/ternary.cu +271 -0
  455. data/mlx/mlx/backend/cuda/unary/CMakeLists.txt +34 -0
  456. data/mlx/mlx/backend/cuda/unary/abs.cu +7 -0
  457. data/mlx/mlx/backend/cuda/unary/arccos.cu +7 -0
  458. data/mlx/mlx/backend/cuda/unary/arccosh.cu +7 -0
  459. data/mlx/mlx/backend/cuda/unary/arcsin.cu +7 -0
  460. data/mlx/mlx/backend/cuda/unary/arcsinh.cu +7 -0
  461. data/mlx/mlx/backend/cuda/unary/arctan.cu +7 -0
  462. data/mlx/mlx/backend/cuda/unary/arctanh.cu +7 -0
  463. data/mlx/mlx/backend/cuda/unary/bitwise_invert.cu +7 -0
  464. data/mlx/mlx/backend/cuda/unary/ceil.cu +7 -0
  465. data/mlx/mlx/backend/cuda/unary/conjugate.cu +7 -0
  466. data/mlx/mlx/backend/cuda/unary/cos.cu +7 -0
  467. data/mlx/mlx/backend/cuda/unary/cosh.cu +7 -0
  468. data/mlx/mlx/backend/cuda/unary/erf.cu +7 -0
  469. data/mlx/mlx/backend/cuda/unary/erf_inv.cu +7 -0
  470. data/mlx/mlx/backend/cuda/unary/exp.cu +7 -0
  471. data/mlx/mlx/backend/cuda/unary/expm1.cu +7 -0
  472. data/mlx/mlx/backend/cuda/unary/floor.cu +7 -0
  473. data/mlx/mlx/backend/cuda/unary/imag.cu +7 -0
  474. data/mlx/mlx/backend/cuda/unary/log.cu +21 -0
  475. data/mlx/mlx/backend/cuda/unary/log1p.cu +7 -0
  476. data/mlx/mlx/backend/cuda/unary/logical_not.cu +7 -0
  477. data/mlx/mlx/backend/cuda/unary/negative.cu +7 -0
  478. data/mlx/mlx/backend/cuda/unary/real.cu +7 -0
  479. data/mlx/mlx/backend/cuda/unary/round.cu +18 -0
  480. data/mlx/mlx/backend/cuda/unary/sigmoid.cu +7 -0
  481. data/mlx/mlx/backend/cuda/unary/sign.cu +7 -0
  482. data/mlx/mlx/backend/cuda/unary/sin.cu +7 -0
  483. data/mlx/mlx/backend/cuda/unary/sinh.cu +7 -0
  484. data/mlx/mlx/backend/cuda/unary/sqrt.cu +15 -0
  485. data/mlx/mlx/backend/cuda/unary/square.cu +7 -0
  486. data/mlx/mlx/backend/cuda/unary/tan.cu +7 -0
  487. data/mlx/mlx/backend/cuda/unary/tanh.cu +7 -0
  488. data/mlx/mlx/backend/cuda/unary/unary.cuh +224 -0
  489. data/mlx/mlx/backend/cuda/utils.cpp +116 -0
  490. data/mlx/mlx/backend/cuda/utils.h +49 -0
  491. data/mlx/mlx/backend/cuda/vector_types.cuh +48 -0
  492. data/mlx/mlx/backend/cuda/worker.cpp +79 -0
  493. data/mlx/mlx/backend/cuda/worker.h +55 -0
  494. data/mlx/mlx/backend/gpu/CMakeLists.txt +5 -0
  495. data/mlx/mlx/backend/gpu/copy.cpp +89 -0
  496. data/mlx/mlx/backend/gpu/copy.h +57 -0
  497. data/mlx/mlx/backend/gpu/device_info.h +36 -0
  498. data/mlx/mlx/backend/gpu/eval.h +18 -0
  499. data/mlx/mlx/backend/gpu/primitives.cpp +307 -0
  500. data/mlx/mlx/backend/gpu/slicing.cpp +44 -0
  501. data/mlx/mlx/backend/gpu/slicing.h +36 -0
  502. data/mlx/mlx/backend/metal/CMakeLists.txt +144 -0
  503. data/mlx/mlx/backend/metal/allocator.cpp +279 -0
  504. data/mlx/mlx/backend/metal/allocator.h +79 -0
  505. data/mlx/mlx/backend/metal/binary.cpp +257 -0
  506. data/mlx/mlx/backend/metal/binary.h +33 -0
  507. data/mlx/mlx/backend/metal/compiled.cpp +471 -0
  508. data/mlx/mlx/backend/metal/conv.cpp +1118 -0
  509. data/mlx/mlx/backend/metal/copy.cpp +235 -0
  510. data/mlx/mlx/backend/metal/custom_kernel.cpp +430 -0
  511. data/mlx/mlx/backend/metal/device.cpp +816 -0
  512. data/mlx/mlx/backend/metal/device.h +289 -0
  513. data/mlx/mlx/backend/metal/device_info.cpp +58 -0
  514. data/mlx/mlx/backend/metal/distributed.cpp +38 -0
  515. data/mlx/mlx/backend/metal/eval.cpp +97 -0
  516. data/mlx/mlx/backend/metal/event.cpp +62 -0
  517. data/mlx/mlx/backend/metal/fence.cpp +162 -0
  518. data/mlx/mlx/backend/metal/fft.cpp +807 -0
  519. data/mlx/mlx/backend/metal/hadamard.cpp +198 -0
  520. data/mlx/mlx/backend/metal/indexing.cpp +727 -0
  521. data/mlx/mlx/backend/metal/jit/includes.h +58 -0
  522. data/mlx/mlx/backend/metal/jit/indexing.h +76 -0
  523. data/mlx/mlx/backend/metal/jit_kernels.cpp +1118 -0
  524. data/mlx/mlx/backend/metal/kernels/CMakeLists.txt +193 -0
  525. data/mlx/mlx/backend/metal/kernels/arange.h +9 -0
  526. data/mlx/mlx/backend/metal/kernels/arange.metal +20 -0
  527. data/mlx/mlx/backend/metal/kernels/arg_reduce.metal +182 -0
  528. data/mlx/mlx/backend/metal/kernels/atomic.h +345 -0
  529. data/mlx/mlx/backend/metal/kernels/bf16.h +16 -0
  530. data/mlx/mlx/backend/metal/kernels/bf16_math.h +380 -0
  531. data/mlx/mlx/backend/metal/kernels/binary.h +199 -0
  532. data/mlx/mlx/backend/metal/kernels/binary.metal +109 -0
  533. data/mlx/mlx/backend/metal/kernels/binary_ops.h +330 -0
  534. data/mlx/mlx/backend/metal/kernels/binary_two.h +244 -0
  535. data/mlx/mlx/backend/metal/kernels/binary_two.metal +54 -0
  536. data/mlx/mlx/backend/metal/kernels/cexpf.h +134 -0
  537. data/mlx/mlx/backend/metal/kernels/complex.h +173 -0
  538. data/mlx/mlx/backend/metal/kernels/conv.metal +701 -0
  539. data/mlx/mlx/backend/metal/kernels/copy.h +276 -0
  540. data/mlx/mlx/backend/metal/kernels/copy.metal +75 -0
  541. data/mlx/mlx/backend/metal/kernels/defines.h +24 -0
  542. data/mlx/mlx/backend/metal/kernels/erf.h +69 -0
  543. data/mlx/mlx/backend/metal/kernels/expm1f.h +90 -0
  544. data/mlx/mlx/backend/metal/kernels/fence.metal +52 -0
  545. data/mlx/mlx/backend/metal/kernels/fft/radix.h +328 -0
  546. data/mlx/mlx/backend/metal/kernels/fft/readwrite.h +624 -0
  547. data/mlx/mlx/backend/metal/kernels/fft.h +486 -0
  548. data/mlx/mlx/backend/metal/kernels/fft.metal +67 -0
  549. data/mlx/mlx/backend/metal/kernels/fp4.h +48 -0
  550. data/mlx/mlx/backend/metal/kernels/fp8.h +80 -0
  551. data/mlx/mlx/backend/metal/kernels/fp_quantized.h +1850 -0
  552. data/mlx/mlx/backend/metal/kernels/fp_quantized.metal +153 -0
  553. data/mlx/mlx/backend/metal/kernels/fp_quantized_nax.h +1044 -0
  554. data/mlx/mlx/backend/metal/kernels/fp_quantized_nax.metal +79 -0
  555. data/mlx/mlx/backend/metal/kernels/gemv.metal +868 -0
  556. data/mlx/mlx/backend/metal/kernels/gemv_masked.h +827 -0
  557. data/mlx/mlx/backend/metal/kernels/gemv_masked.metal +76 -0
  558. data/mlx/mlx/backend/metal/kernels/hadamard.h +182 -0
  559. data/mlx/mlx/backend/metal/kernels/indexing/gather.h +51 -0
  560. data/mlx/mlx/backend/metal/kernels/indexing/gather_axis.h +44 -0
  561. data/mlx/mlx/backend/metal/kernels/indexing/gather_front.h +24 -0
  562. data/mlx/mlx/backend/metal/kernels/indexing/indexing.h +23 -0
  563. data/mlx/mlx/backend/metal/kernels/indexing/masked_scatter.h +41 -0
  564. data/mlx/mlx/backend/metal/kernels/indexing/scatter.h +59 -0
  565. data/mlx/mlx/backend/metal/kernels/indexing/scatter_axis.h +52 -0
  566. data/mlx/mlx/backend/metal/kernels/layer_norm.metal +433 -0
  567. data/mlx/mlx/backend/metal/kernels/logging.h +26 -0
  568. data/mlx/mlx/backend/metal/kernels/logsumexp.h +140 -0
  569. data/mlx/mlx/backend/metal/kernels/logsumexp.metal +18 -0
  570. data/mlx/mlx/backend/metal/kernels/quantized.h +2508 -0
  571. data/mlx/mlx/backend/metal/kernels/quantized.metal +144 -0
  572. data/mlx/mlx/backend/metal/kernels/quantized_nax.h +1705 -0
  573. data/mlx/mlx/backend/metal/kernels/quantized_nax.metal +106 -0
  574. data/mlx/mlx/backend/metal/kernels/quantized_utils.h +90 -0
  575. data/mlx/mlx/backend/metal/kernels/random.metal +103 -0
  576. data/mlx/mlx/backend/metal/kernels/reduce.h +5 -0
  577. data/mlx/mlx/backend/metal/kernels/reduce.metal +169 -0
  578. data/mlx/mlx/backend/metal/kernels/reduce_utils.h +6 -0
  579. data/mlx/mlx/backend/metal/kernels/reduction/ops.h +275 -0
  580. data/mlx/mlx/backend/metal/kernels/reduction/reduce_all.h +66 -0
  581. data/mlx/mlx/backend/metal/kernels/reduction/reduce_col.h +398 -0
  582. data/mlx/mlx/backend/metal/kernels/reduction/reduce_init.h +8 -0
  583. data/mlx/mlx/backend/metal/kernels/reduction/reduce_row.h +369 -0
  584. data/mlx/mlx/backend/metal/kernels/rms_norm.metal +391 -0
  585. data/mlx/mlx/backend/metal/kernels/rope.metal +229 -0
  586. data/mlx/mlx/backend/metal/kernels/scaled_dot_product_attention.metal +44 -0
  587. data/mlx/mlx/backend/metal/kernels/scan.h +514 -0
  588. data/mlx/mlx/backend/metal/kernels/scan.metal +109 -0
  589. data/mlx/mlx/backend/metal/kernels/sdpa_vector.h +394 -0
  590. data/mlx/mlx/backend/metal/kernels/softmax.h +190 -0
  591. data/mlx/mlx/backend/metal/kernels/softmax.metal +24 -0
  592. data/mlx/mlx/backend/metal/kernels/sort.h +719 -0
  593. data/mlx/mlx/backend/metal/kernels/sort.metal +80 -0
  594. data/mlx/mlx/backend/metal/kernels/steel/attn/attn.h +296 -0
  595. data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.h +471 -0
  596. data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.metal +27 -0
  597. data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.h +481 -0
  598. data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.metal +28 -0
  599. data/mlx/mlx/backend/metal/kernels/steel/attn/loader.h +264 -0
  600. data/mlx/mlx/backend/metal/kernels/steel/attn/mma.h +750 -0
  601. data/mlx/mlx/backend/metal/kernels/steel/attn/nax.h +1076 -0
  602. data/mlx/mlx/backend/metal/kernels/steel/attn/params.h +44 -0
  603. data/mlx/mlx/backend/metal/kernels/steel/attn/transforms.h +71 -0
  604. data/mlx/mlx/backend/metal/kernels/steel/conv/conv.h +13 -0
  605. data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv.h +176 -0
  606. data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv.metal +56 -0
  607. data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv_general.h +225 -0
  608. data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv_general.metal +47 -0
  609. data/mlx/mlx/backend/metal/kernels/steel/conv/loader.h +6 -0
  610. data/mlx/mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_l.h +451 -0
  611. data/mlx/mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_n.h +319 -0
  612. data/mlx/mlx/backend/metal/kernels/steel/conv/loaders/loader_general.h +381 -0
  613. data/mlx/mlx/backend/metal/kernels/steel/conv/params.h +62 -0
  614. data/mlx/mlx/backend/metal/kernels/steel/defines.h +7 -0
  615. data/mlx/mlx/backend/metal/kernels/steel/gemm/gemm.h +295 -0
  616. data/mlx/mlx/backend/metal/kernels/steel/gemm/gemm_nax.h +157 -0
  617. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused.h +346 -0
  618. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused.metal +34 -0
  619. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.h +219 -0
  620. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.metal +30 -0
  621. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather.h +459 -0
  622. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather.metal +59 -0
  623. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.h +143 -0
  624. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.metal +37 -0
  625. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_masked.h +719 -0
  626. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_masked.metal +76 -0
  627. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_segmented.h +266 -0
  628. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_segmented.metal +43 -0
  629. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk.h +227 -0
  630. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk.metal +76 -0
  631. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk_nax.h +152 -0
  632. data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk_nax.metal +30 -0
  633. data/mlx/mlx/backend/metal/kernels/steel/gemm/loader.h +137 -0
  634. data/mlx/mlx/backend/metal/kernels/steel/gemm/mma.h +1146 -0
  635. data/mlx/mlx/backend/metal/kernels/steel/gemm/nax.h +1084 -0
  636. data/mlx/mlx/backend/metal/kernels/steel/gemm/params.h +65 -0
  637. data/mlx/mlx/backend/metal/kernels/steel/gemm/transforms.h +72 -0
  638. data/mlx/mlx/backend/metal/kernels/steel/utils/integral_constant.h +134 -0
  639. data/mlx/mlx/backend/metal/kernels/steel/utils/type_traits.h +55 -0
  640. data/mlx/mlx/backend/metal/kernels/steel/utils.h +42 -0
  641. data/mlx/mlx/backend/metal/kernels/ternary.h +145 -0
  642. data/mlx/mlx/backend/metal/kernels/ternary.metal +48 -0
  643. data/mlx/mlx/backend/metal/kernels/ternary_ops.h +10 -0
  644. data/mlx/mlx/backend/metal/kernels/unary.h +63 -0
  645. data/mlx/mlx/backend/metal/kernels/unary.metal +115 -0
  646. data/mlx/mlx/backend/metal/kernels/unary_ops.h +454 -0
  647. data/mlx/mlx/backend/metal/kernels/utils.h +445 -0
  648. data/mlx/mlx/backend/metal/kernels.h +375 -0
  649. data/mlx/mlx/backend/metal/logsumexp.cpp +95 -0
  650. data/mlx/mlx/backend/metal/make_compiled_preamble.sh +120 -0
  651. data/mlx/mlx/backend/metal/matmul.cpp +2572 -0
  652. data/mlx/mlx/backend/metal/matmul.h +144 -0
  653. data/mlx/mlx/backend/metal/metal.cpp +50 -0
  654. data/mlx/mlx/backend/metal/metal.h +25 -0
  655. data/mlx/mlx/backend/metal/no_metal.cpp +42 -0
  656. data/mlx/mlx/backend/metal/nojit_kernels.cpp +414 -0
  657. data/mlx/mlx/backend/metal/normalization.cpp +433 -0
  658. data/mlx/mlx/backend/metal/primitives.cpp +242 -0
  659. data/mlx/mlx/backend/metal/quantized.cpp +1651 -0
  660. data/mlx/mlx/backend/metal/reduce.cpp +1038 -0
  661. data/mlx/mlx/backend/metal/reduce.h +41 -0
  662. data/mlx/mlx/backend/metal/resident.cpp +100 -0
  663. data/mlx/mlx/backend/metal/resident.h +32 -0
  664. data/mlx/mlx/backend/metal/rope.cpp +165 -0
  665. data/mlx/mlx/backend/metal/scaled_dot_product_attention.cpp +798 -0
  666. data/mlx/mlx/backend/metal/scan.cpp +145 -0
  667. data/mlx/mlx/backend/metal/scan.h +17 -0
  668. data/mlx/mlx/backend/metal/slicing.cpp +99 -0
  669. data/mlx/mlx/backend/metal/softmax.cpp +87 -0
  670. data/mlx/mlx/backend/metal/sort.cpp +368 -0
  671. data/mlx/mlx/backend/metal/ternary.cpp +160 -0
  672. data/mlx/mlx/backend/metal/ternary.h +21 -0
  673. data/mlx/mlx/backend/metal/unary.cpp +161 -0
  674. data/mlx/mlx/backend/metal/unary.h +21 -0
  675. data/mlx/mlx/backend/metal/utils.cpp +77 -0
  676. data/mlx/mlx/backend/metal/utils.h +99 -0
  677. data/mlx/mlx/backend/no_cpu/CMakeLists.txt +7 -0
  678. data/mlx/mlx/backend/no_cpu/compiled.cpp +24 -0
  679. data/mlx/mlx/backend/no_cpu/device_info.cpp +22 -0
  680. data/mlx/mlx/backend/no_cpu/primitives.cpp +146 -0
  681. data/mlx/mlx/backend/no_gpu/CMakeLists.txt +8 -0
  682. data/mlx/mlx/backend/no_gpu/allocator.cpp +134 -0
  683. data/mlx/mlx/backend/no_gpu/apple_memory.h +16 -0
  684. data/mlx/mlx/backend/no_gpu/device_info.cpp +22 -0
  685. data/mlx/mlx/backend/no_gpu/eval.cpp +24 -0
  686. data/mlx/mlx/backend/no_gpu/event.cpp +53 -0
  687. data/mlx/mlx/backend/no_gpu/fence.cpp +54 -0
  688. data/mlx/mlx/backend/no_gpu/linux_memory.h +22 -0
  689. data/mlx/mlx/backend/no_gpu/primitives.cpp +185 -0
  690. data/mlx/mlx/compile.cpp +1243 -0
  691. data/mlx/mlx/compile.h +45 -0
  692. data/mlx/mlx/compile_impl.h +70 -0
  693. data/mlx/mlx/device.cpp +72 -0
  694. data/mlx/mlx/device.h +56 -0
  695. data/mlx/mlx/distributed/CMakeLists.txt +14 -0
  696. data/mlx/mlx/distributed/distributed.cpp +197 -0
  697. data/mlx/mlx/distributed/distributed.h +61 -0
  698. data/mlx/mlx/distributed/distributed_impl.h +59 -0
  699. data/mlx/mlx/distributed/jaccl/CMakeLists.txt +12 -0
  700. data/mlx/mlx/distributed/jaccl/jaccl.cpp +178 -0
  701. data/mlx/mlx/distributed/jaccl/jaccl.h +12 -0
  702. data/mlx/mlx/distributed/jaccl/mesh.cpp +451 -0
  703. data/mlx/mlx/distributed/jaccl/mesh.h +122 -0
  704. data/mlx/mlx/distributed/jaccl/no_jaccl.cpp +20 -0
  705. data/mlx/mlx/distributed/jaccl/ring.cpp +692 -0
  706. data/mlx/mlx/distributed/jaccl/ring.h +178 -0
  707. data/mlx/mlx/distributed/jaccl/utils.cpp +329 -0
  708. data/mlx/mlx/distributed/jaccl/utils.h +342 -0
  709. data/mlx/mlx/distributed/mpi/CMakeLists.txt +5 -0
  710. data/mlx/mlx/distributed/mpi/mpi.cpp +501 -0
  711. data/mlx/mlx/distributed/mpi/mpi.h +12 -0
  712. data/mlx/mlx/distributed/mpi/mpi_declarations.h +28 -0
  713. data/mlx/mlx/distributed/mpi/no_mpi.cpp +20 -0
  714. data/mlx/mlx/distributed/nccl/CMakeLists.txt +26 -0
  715. data/mlx/mlx/distributed/nccl/nccl.cpp +443 -0
  716. data/mlx/mlx/distributed/nccl/nccl.h +12 -0
  717. data/mlx/mlx/distributed/nccl/nccl_stub/CMakeLists.txt +1 -0
  718. data/mlx/mlx/distributed/nccl/nccl_stub/nccl_stubs.cpp +54 -0
  719. data/mlx/mlx/distributed/nccl/no_nccl.cpp +20 -0
  720. data/mlx/mlx/distributed/ops.cpp +186 -0
  721. data/mlx/mlx/distributed/ops.h +57 -0
  722. data/mlx/mlx/distributed/primitives.cpp +95 -0
  723. data/mlx/mlx/distributed/primitives.h +156 -0
  724. data/mlx/mlx/distributed/reduction_ops.h +38 -0
  725. data/mlx/mlx/distributed/ring/CMakeLists.txt +5 -0
  726. data/mlx/mlx/distributed/ring/no_ring.cpp +20 -0
  727. data/mlx/mlx/distributed/ring/ring.cpp +870 -0
  728. data/mlx/mlx/distributed/ring/ring.h +12 -0
  729. data/mlx/mlx/distributed/utils.cpp +206 -0
  730. data/mlx/mlx/distributed/utils.h +67 -0
  731. data/mlx/mlx/dtype.cpp +197 -0
  732. data/mlx/mlx/dtype.h +116 -0
  733. data/mlx/mlx/dtype_utils.cpp +42 -0
  734. data/mlx/mlx/dtype_utils.h +119 -0
  735. data/mlx/mlx/einsum.cpp +941 -0
  736. data/mlx/mlx/einsum.h +23 -0
  737. data/mlx/mlx/event.h +58 -0
  738. data/mlx/mlx/export.cpp +1130 -0
  739. data/mlx/mlx/export.h +137 -0
  740. data/mlx/mlx/export_impl.h +99 -0
  741. data/mlx/mlx/fast.cpp +941 -0
  742. data/mlx/mlx/fast.h +103 -0
  743. data/mlx/mlx/fast_primitives.h +427 -0
  744. data/mlx/mlx/fence.h +39 -0
  745. data/mlx/mlx/fft.cpp +262 -0
  746. data/mlx/mlx/fft.h +159 -0
  747. data/mlx/mlx/graph_utils.cpp +175 -0
  748. data/mlx/mlx/graph_utils.h +67 -0
  749. data/mlx/mlx/io/CMakeLists.txt +25 -0
  750. data/mlx/mlx/io/gguf.cpp +470 -0
  751. data/mlx/mlx/io/gguf.h +20 -0
  752. data/mlx/mlx/io/gguf_quants.cpp +164 -0
  753. data/mlx/mlx/io/load.cpp +397 -0
  754. data/mlx/mlx/io/load.h +175 -0
  755. data/mlx/mlx/io/no_gguf.cpp +20 -0
  756. data/mlx/mlx/io/no_safetensors.cpp +37 -0
  757. data/mlx/mlx/io/safetensors.cpp +234 -0
  758. data/mlx/mlx/io.h +61 -0
  759. data/mlx/mlx/linalg.cpp +708 -0
  760. data/mlx/mlx/linalg.h +115 -0
  761. data/mlx/mlx/memory.h +80 -0
  762. data/mlx/mlx/mlx.h +25 -0
  763. data/mlx/mlx/ops.cpp +6094 -0
  764. data/mlx/mlx/ops.h +1610 -0
  765. data/mlx/mlx/primitives.cpp +5850 -0
  766. data/mlx/mlx/primitives.h +2525 -0
  767. data/mlx/mlx/random.cpp +492 -0
  768. data/mlx/mlx/random.h +283 -0
  769. data/mlx/mlx/scheduler.cpp +73 -0
  770. data/mlx/mlx/scheduler.h +189 -0
  771. data/mlx/mlx/small_vector.h +540 -0
  772. data/mlx/mlx/stream.h +42 -0
  773. data/mlx/mlx/threadpool.h +133 -0
  774. data/mlx/mlx/transforms.cpp +1065 -0
  775. data/mlx/mlx/transforms.h +231 -0
  776. data/mlx/mlx/transforms_impl.h +88 -0
  777. data/mlx/mlx/types/bf16.h +187 -0
  778. data/mlx/mlx/types/complex.h +113 -0
  779. data/mlx/mlx/types/fp16.h +234 -0
  780. data/mlx/mlx/types/half_types.h +58 -0
  781. data/mlx/mlx/types/limits.h +70 -0
  782. data/mlx/mlx/utils.cpp +302 -0
  783. data/mlx/mlx/utils.h +174 -0
  784. data/mlx/mlx/version.cpp +11 -0
  785. data/mlx/mlx/version.h +22 -0
  786. data/mlx/mlx.pc.in +52 -0
  787. data/mlx/pyproject.toml +7 -0
  788. data/mlx/python/mlx/__main__.py +27 -0
  789. data/mlx/python/mlx/_distributed_utils/common.py +135 -0
  790. data/mlx/python/mlx/_distributed_utils/config.py +631 -0
  791. data/mlx/python/mlx/_distributed_utils/launch.py +570 -0
  792. data/mlx/python/mlx/_reprlib_fix.py +16 -0
  793. data/mlx/python/mlx/_stub_patterns.txt +36 -0
  794. data/mlx/python/mlx/extension.py +88 -0
  795. data/mlx/python/mlx/nn/__init__.py +5 -0
  796. data/mlx/python/mlx/nn/init.py +441 -0
  797. data/mlx/python/mlx/nn/layers/__init__.py +105 -0
  798. data/mlx/python/mlx/nn/layers/activations.py +661 -0
  799. data/mlx/python/mlx/nn/layers/base.py +675 -0
  800. data/mlx/python/mlx/nn/layers/containers.py +24 -0
  801. data/mlx/python/mlx/nn/layers/convolution.py +232 -0
  802. data/mlx/python/mlx/nn/layers/convolution_transpose.py +242 -0
  803. data/mlx/python/mlx/nn/layers/distributed.py +601 -0
  804. data/mlx/python/mlx/nn/layers/dropout.py +137 -0
  805. data/mlx/python/mlx/nn/layers/embedding.py +53 -0
  806. data/mlx/python/mlx/nn/layers/linear.py +180 -0
  807. data/mlx/python/mlx/nn/layers/normalization.py +363 -0
  808. data/mlx/python/mlx/nn/layers/pooling.py +398 -0
  809. data/mlx/python/mlx/nn/layers/positional_encoding.py +162 -0
  810. data/mlx/python/mlx/nn/layers/quantized.py +426 -0
  811. data/mlx/python/mlx/nn/layers/recurrent.py +289 -0
  812. data/mlx/python/mlx/nn/layers/transformer.py +354 -0
  813. data/mlx/python/mlx/nn/layers/upsample.py +277 -0
  814. data/mlx/python/mlx/nn/losses.py +610 -0
  815. data/mlx/python/mlx/nn/utils.py +165 -0
  816. data/mlx/python/mlx/optimizers/__init__.py +4 -0
  817. data/mlx/python/mlx/optimizers/optimizers.py +976 -0
  818. data/mlx/python/mlx/optimizers/schedulers.py +158 -0
  819. data/mlx/python/mlx/py.typed +1 -0
  820. data/mlx/python/mlx/utils.py +325 -0
  821. data/mlx/python/src/CMakeLists.txt +96 -0
  822. data/mlx/python/src/array.cpp +1525 -0
  823. data/mlx/python/src/buffer.h +124 -0
  824. data/mlx/python/src/constants.cpp +15 -0
  825. data/mlx/python/src/convert.cpp +504 -0
  826. data/mlx/python/src/convert.h +50 -0
  827. data/mlx/python/src/cuda.cpp +19 -0
  828. data/mlx/python/src/device.cpp +98 -0
  829. data/mlx/python/src/distributed.cpp +352 -0
  830. data/mlx/python/src/export.cpp +356 -0
  831. data/mlx/python/src/fast.cpp +627 -0
  832. data/mlx/python/src/fft.cpp +514 -0
  833. data/mlx/python/src/indexing.cpp +1016 -0
  834. data/mlx/python/src/indexing.h +41 -0
  835. data/mlx/python/src/linalg.cpp +663 -0
  836. data/mlx/python/src/load.cpp +531 -0
  837. data/mlx/python/src/load.h +51 -0
  838. data/mlx/python/src/memory.cpp +125 -0
  839. data/mlx/python/src/metal.cpp +98 -0
  840. data/mlx/python/src/mlx.cpp +51 -0
  841. data/mlx/python/src/mlx_func.cpp +116 -0
  842. data/mlx/python/src/mlx_func.h +31 -0
  843. data/mlx/python/src/ops.cpp +5545 -0
  844. data/mlx/python/src/random.cpp +516 -0
  845. data/mlx/python/src/small_vector.h +76 -0
  846. data/mlx/python/src/stream.cpp +147 -0
  847. data/mlx/python/src/transforms.cpp +1542 -0
  848. data/mlx/python/src/trees.cpp +311 -0
  849. data/mlx/python/src/trees.h +62 -0
  850. data/mlx/python/src/utils.cpp +98 -0
  851. data/mlx/python/src/utils.h +78 -0
  852. data/mlx/python/tests/__main__.py +5 -0
  853. data/mlx/python/tests/cuda_skip.py +62 -0
  854. data/mlx/python/tests/mlx_distributed_tests.py +314 -0
  855. data/mlx/python/tests/mlx_tests.py +116 -0
  856. data/mlx/python/tests/mpi_test_distributed.py +142 -0
  857. data/mlx/python/tests/nccl_test_distributed.py +52 -0
  858. data/mlx/python/tests/ring_test_distributed.py +131 -0
  859. data/mlx/python/tests/test_array.py +2139 -0
  860. data/mlx/python/tests/test_autograd.py +880 -0
  861. data/mlx/python/tests/test_bf16.py +196 -0
  862. data/mlx/python/tests/test_blas.py +1429 -0
  863. data/mlx/python/tests/test_compile.py +1277 -0
  864. data/mlx/python/tests/test_constants.py +41 -0
  865. data/mlx/python/tests/test_conv.py +1198 -0
  866. data/mlx/python/tests/test_conv_transpose.py +810 -0
  867. data/mlx/python/tests/test_device.py +150 -0
  868. data/mlx/python/tests/test_double.py +306 -0
  869. data/mlx/python/tests/test_einsum.py +363 -0
  870. data/mlx/python/tests/test_eval.py +200 -0
  871. data/mlx/python/tests/test_export_import.py +614 -0
  872. data/mlx/python/tests/test_fast.py +923 -0
  873. data/mlx/python/tests/test_fast_sdpa.py +647 -0
  874. data/mlx/python/tests/test_fft.py +323 -0
  875. data/mlx/python/tests/test_graph.py +37 -0
  876. data/mlx/python/tests/test_init.py +139 -0
  877. data/mlx/python/tests/test_linalg.py +621 -0
  878. data/mlx/python/tests/test_load.py +447 -0
  879. data/mlx/python/tests/test_losses.py +427 -0
  880. data/mlx/python/tests/test_memory.py +77 -0
  881. data/mlx/python/tests/test_nn.py +1986 -0
  882. data/mlx/python/tests/test_ops.py +3261 -0
  883. data/mlx/python/tests/test_optimizers.py +584 -0
  884. data/mlx/python/tests/test_quantized.py +1160 -0
  885. data/mlx/python/tests/test_random.py +392 -0
  886. data/mlx/python/tests/test_reduce.py +223 -0
  887. data/mlx/python/tests/test_tree.py +96 -0
  888. data/mlx/python/tests/test_upsample.py +100 -0
  889. data/mlx/python/tests/test_vmap.py +860 -0
  890. data/mlx/setup.py +315 -0
  891. data/mlx/tests/CMakeLists.txt +44 -0
  892. data/mlx/tests/allocator_tests.cpp +41 -0
  893. data/mlx/tests/arg_reduce_tests.cpp +204 -0
  894. data/mlx/tests/array_tests.cpp +663 -0
  895. data/mlx/tests/autograd_tests.cpp +1399 -0
  896. data/mlx/tests/blas_tests.cpp +110 -0
  897. data/mlx/tests/compile_tests.cpp +818 -0
  898. data/mlx/tests/creations_tests.cpp +239 -0
  899. data/mlx/tests/custom_vjp_tests.cpp +55 -0
  900. data/mlx/tests/device_tests.cpp +35 -0
  901. data/mlx/tests/einsum_tests.cpp +85 -0
  902. data/mlx/tests/eval_tests.cpp +93 -0
  903. data/mlx/tests/export_import_tests.cpp +164 -0
  904. data/mlx/tests/fft_tests.cpp +366 -0
  905. data/mlx/tests/gpu_tests.cpp +523 -0
  906. data/mlx/tests/linalg_tests.cpp +639 -0
  907. data/mlx/tests/load_tests.cpp +270 -0
  908. data/mlx/tests/ops_tests.cpp +4159 -0
  909. data/mlx/tests/random_tests.cpp +716 -0
  910. data/mlx/tests/scheduler_tests.cpp +121 -0
  911. data/mlx/tests/tests.cpp +26 -0
  912. data/mlx/tests/utils_tests.cpp +67 -0
  913. data/mlx/tests/vmap_tests.cpp +547 -0
  914. metadata +958 -0
@@ -0,0 +1,570 @@
1
+ # Copyright © 2025 Apple Inc.
2
+
3
+ import argparse
4
+ import base64
5
+ import json
6
+ import os
7
+ import shlex
8
+ import shutil
9
+ import sys
10
+ import tempfile
11
+ import threading
12
+ from collections import Counter
13
+ from itertools import chain
14
+ from pathlib import Path
15
+ from queue import Empty as QueueEmpty
16
+ from queue import Queue
17
+ from select import select
18
+ from subprocess import PIPE, Popen, run
19
+
20
+ import mlx.core as mx
21
+
22
+ from .common import Hostfile, log, log_warning, positive_number
23
+
24
+
25
+ class CommandProcess:
26
+ @property
27
+ def process(self):
28
+ """Return the Popen object that refers to the current command."""
29
+ raise NotImplementedError()
30
+
31
+ @property
32
+ def exit_status(self):
33
+ """Return a tuple (returncode, killed) for the command. It should be
34
+ (None, None) while the command is running normally."""
35
+ raise NotImplementedError()
36
+
37
+ def preprocess_output(self, data: str, is_stdout=False):
38
+ """Preprocess the output of the command so that extra data can be
39
+ capture or the format changed on the fly."""
40
+ raise NotImplementedError()
41
+
42
+ def terminate(self):
43
+ """Terminate or return the exit code."""
44
+ raise NotImplementedError()
45
+
46
+
47
+ class RemoteProcess(CommandProcess):
48
+ def __init__(self, rank, host, python, cwd, files, env, command):
49
+ is_local = host == "127.0.0.1"
50
+ cmd = RemoteProcess.make_launch_script(rank, cwd, files, env, command, is_local)
51
+ if not is_local:
52
+ cmd = f"ssh -tt -o LogLevel=QUIET {host} {shlex.quote(cmd)}"
53
+
54
+ self._host = host
55
+ self._pidfile = None
56
+ self._is_local = is_local
57
+ self._process = Popen(
58
+ cmd,
59
+ shell=True,
60
+ executable="/bin/bash",
61
+ stdin=PIPE,
62
+ stdout=PIPE,
63
+ stderr=PIPE,
64
+ )
65
+
66
+ self._killed = False
67
+
68
+ @property
69
+ def process(self):
70
+ return self._process
71
+
72
+ @property
73
+ def exit_status(self):
74
+ return self._process.poll(), self._killed
75
+
76
+ def preprocess_output(self, data, is_stdout=False):
77
+ if self._pidfile is None:
78
+ pidfile, *rest = data.split("\n", maxsplit=1)
79
+ self._pidfile = pidfile
80
+ return rest[0] if rest else ""
81
+
82
+ return data
83
+
84
+ def terminate(self):
85
+ if self._killed:
86
+ return
87
+
88
+ self._process.terminate()
89
+ self._process.wait()
90
+
91
+ # Kill the remote program if possible
92
+ cmd = RemoteProcess.make_kill_script(self._pidfile)
93
+ if not self._is_local:
94
+ cmd = f"ssh {self._host} {shlex.quote(cmd)}"
95
+ c = run(
96
+ cmd,
97
+ check=True,
98
+ shell=True,
99
+ executable="/bin/bash",
100
+ capture_output=True,
101
+ text=True,
102
+ )
103
+
104
+ self._killed = c.stdout.strip() == "1"
105
+
106
+ @staticmethod
107
+ def make_launch_script(rank, cwd, files, env, command, is_local):
108
+ script = ""
109
+
110
+ # Disable echo
111
+ if not is_local:
112
+ script = "stty -echo; "
113
+
114
+ # Write the PID to a file so we can kill the process if needed
115
+ script += "pidfile=$(mktemp); "
116
+ script += "echo $$ > $pidfile; "
117
+ script += 'printf "%s\\n" $pidfile; '
118
+
119
+ # Change the working directory if one was requested. Otherwise attempt to
120
+ # change to the current one but don't fail if it wasn't possible.
121
+ d = cwd or os.getcwd()
122
+ script += f"if [[ -d {repr(d)} ]]; then "
123
+ script += f" cd {repr(d)}; "
124
+ if cwd is not None:
125
+ script += "else "
126
+ script += f" echo 'Failed to change directory to' {repr(d)} >2; "
127
+ script += "fi; "
128
+
129
+ # Add the environment variables that were requested
130
+ for e in env:
131
+ key, *value = e.split("=", maxsplit=1)
132
+ value = shlex.quote(value[0]) if len(value) > 0 else ""
133
+ if not all(c.isalnum() or c == "_" for c in key):
134
+ log_warning(
135
+ f"'{e}' is an invalid environment variable so it is ignored"
136
+ )
137
+ continue
138
+ script += f"export {key}={value}; "
139
+
140
+ # Make the temporary files
141
+ for env_name, content in files.items():
142
+ script += "fname=$(mktemp); "
143
+ script += f"echo {shlex.quote(content)} >$fname; "
144
+ script += f"export {env_name}=$fname; "
145
+
146
+ # Finally add the rank
147
+ script += f"export MLX_RANK={rank}; "
148
+
149
+ # Replace the process with the script
150
+ script += f"cmd=({' '.join(map(shlex.quote, command))}); "
151
+ script += 'exec "${cmd[@]}"'
152
+
153
+ return script
154
+
155
+ @staticmethod
156
+ def make_kill_script(pidfile):
157
+ script = ""
158
+ script += f"pid=$(cat {pidfile}); "
159
+ script += "if ps -p $pid >/dev/null; then "
160
+ script += " kill $pid; "
161
+ script += " echo 1; "
162
+ script += "else "
163
+ script += " echo 0; "
164
+ script += "fi; "
165
+ script += f"rm {pidfile}"
166
+
167
+ return script
168
+
169
+
170
+ def _launch_with_io(command_class, arguments, verbose):
171
+ stop = False
172
+ exit_codes = [(None, None)] * len(arguments)
173
+
174
+ def _thread_fn(rank, *args, **kwargs):
175
+ stdin_queue = kwargs.pop("stdin_queue")
176
+ stdout_queue = kwargs.pop("stdout_queue")
177
+ stderr_queue = kwargs.pop("stderr_queue")
178
+
179
+ command = command_class(rank, *args, **kwargs)
180
+ p = command.process
181
+ os.set_blocking(p.stdout.fileno(), False)
182
+ os.set_blocking(p.stderr.fileno(), False)
183
+ os.set_blocking(p.stdin.fileno(), False)
184
+
185
+ to_read = [p.stdout.fileno(), p.stderr.fileno()]
186
+ to_write = [p.stdin.fileno()]
187
+
188
+ stdin_buffer = b""
189
+ while p.poll() is None:
190
+ try:
191
+ stdin_buffer += stdin_queue.get_nowait()
192
+ except QueueEmpty:
193
+ pass
194
+ rlist, wlist, _ = select(to_read, to_write, [], 1.0)
195
+ for fd in rlist:
196
+ is_stdout = fd == p.stdout.fileno()
197
+ msg = os.read(fd, 8192).decode(errors="ignore")
198
+ msg = command.preprocess_output(msg, is_stdout)
199
+ if is_stdout:
200
+ stdout_queue.put(msg.encode())
201
+ else:
202
+ stderr_queue.put(msg.encode())
203
+ for fd in wlist:
204
+ if len(stdin_buffer) > 0:
205
+ n = os.write(fd, stdin_buffer)
206
+ stdin_buffer = stdin_buffer[n:]
207
+ if stop:
208
+ command.terminate()
209
+ break
210
+ exit_codes[rank] = command.exit_status
211
+
212
+ if exit_codes[rank][1]:
213
+ log_warning(f"Node with rank {rank} was killed")
214
+ elif exit_codes[rank][0] != 0:
215
+ log_warning(f"Node with rank {rank} exited with code {exit_codes[rank][0]}")
216
+ else:
217
+ log(verbose, f"Node with rank {rank} completed")
218
+
219
+ stdin_queues = []
220
+ stdout_queues = []
221
+ stderr_queues = []
222
+ threads = []
223
+ for i, (args, kwargs) in enumerate(arguments):
224
+ stdin_queues.append(Queue())
225
+ stdout_queues.append(Queue())
226
+ stderr_queues.append(Queue())
227
+ t = threading.Thread(
228
+ target=_thread_fn,
229
+ args=args,
230
+ kwargs=kwargs
231
+ | {
232
+ "stdin_queue": stdin_queues[-1],
233
+ "stdout_queue": stdout_queues[-1],
234
+ "stderr_queue": stderr_queues[-1],
235
+ },
236
+ )
237
+ t.start()
238
+ threads.append(t)
239
+
240
+ os.set_blocking(sys.stdin.fileno(), False)
241
+ os.set_blocking(sys.stdout.fileno(), True)
242
+ os.set_blocking(sys.stderr.fileno(), True)
243
+ while not stop or any(not q.empty() for q in chain(stdout_queues, stderr_queues)):
244
+ # Broadcast user input to the jobs
245
+ rlist, _, _ = select([sys.stdin.fileno()], [], [], 0.1)
246
+ for fd in rlist:
247
+ stdin_buffer = os.read(fd, 8192)
248
+ for q in stdin_queues:
249
+ q.put(stdin_buffer)
250
+
251
+ # Gather job output
252
+ for q in stdout_queues:
253
+ try:
254
+ while not q.empty():
255
+ sys.stdout.buffer.write(q.get_nowait())
256
+ except QueueEmpty:
257
+ pass
258
+ for q in stderr_queues:
259
+ try:
260
+ while not q.empty():
261
+ sys.stderr.buffer.write(q.get_nowait())
262
+ except QueueEmpty:
263
+ pass
264
+ sys.stdout.buffer.flush()
265
+ sys.stderr.buffer.flush()
266
+
267
+ # Check if all are running and terminate otherwise
268
+ if any(t.is_alive() for t in threads):
269
+ for i, t in enumerate(threads):
270
+ if not t.is_alive():
271
+ if exit_codes[i][0] != 0:
272
+ stop = True
273
+ break
274
+ else:
275
+ break
276
+
277
+ # Wait for the jobs to finish
278
+ for t in threads:
279
+ t.join()
280
+
281
+ # Process any remaining outputs
282
+ for q in stdout_queues:
283
+ while not q.empty():
284
+ sys.stdout.buffer.write(q.get())
285
+ for q in stderr_queues:
286
+ while not q.empty():
287
+ sys.stderr.buffer.write(q.get())
288
+ sys.stdout.buffer.flush()
289
+ sys.stderr.buffer.flush()
290
+
291
+
292
+ def launch_ring(parser, hosts, args, command):
293
+ if any(len(h.ips) == 0 for h in hosts):
294
+ parser.error(
295
+ "The ring backend requires IPs to be provided instead of hostnames"
296
+ )
297
+
298
+ port = args.starting_port
299
+ ring_hosts = []
300
+ for h in hosts:
301
+ node = []
302
+ for ip in h.ips:
303
+ for i in range(args.connections_per_ip):
304
+ node.append(f"{ip}:{port}")
305
+ port += 1
306
+ ring_hosts.append(node)
307
+ hostfile = json.dumps(ring_hosts) if len(ring_hosts) > 1 else ""
308
+
309
+ files = {"MLX_HOSTFILE": hostfile}
310
+ env = args.env
311
+ if args.verbose:
312
+ env.append("MLX_RING_VERBOSE=1")
313
+ cwd = args.cwd
314
+
315
+ log(args.verbose, "Running", shlex.join(command))
316
+
317
+ _launch_with_io(
318
+ RemoteProcess,
319
+ [
320
+ ((rank, h.ssh_hostname, args.python, cwd, files, env, command), {})
321
+ for rank, h in enumerate(hosts)
322
+ ],
323
+ args.verbose,
324
+ )
325
+
326
+
327
+ def launch_nccl(parser, hosts, args, command):
328
+ if not hosts[0].ips:
329
+ raise ValueError("Rank 0 should have an IP reachable from all other ranks")
330
+
331
+ master_host = hosts[0].ips[0]
332
+ master_port = args.nccl_port
333
+ world_size = len(hosts)
334
+
335
+ env = args.env
336
+ cwd = args.cwd
337
+ if args.verbose:
338
+ env.append("NCCL_DEBUG=INFO")
339
+ env.append(f"NCCL_HOST_IP={master_host}")
340
+ env.append(f"NCCL_PORT={master_port}")
341
+ env.append(f"MLX_WORLD_SIZE={world_size}")
342
+
343
+ log(args.verbose, "Running", shlex.join(command))
344
+
345
+ _launch_with_io(
346
+ RemoteProcess,
347
+ [
348
+ (
349
+ (
350
+ rank,
351
+ h.ssh_hostname,
352
+ args.python,
353
+ cwd,
354
+ {},
355
+ env + [f"CUDA_VISIBLE_DEVICES={rank % args.repeat_hosts}"],
356
+ command,
357
+ ),
358
+ {},
359
+ )
360
+ for rank, h in enumerate(hosts)
361
+ ],
362
+ args.verbose,
363
+ )
364
+
365
+
366
+ def launch_jaccl(parser, hosts, args, command):
367
+ if not hosts[0].ips:
368
+ raise ValueError("Rank 0 should have an IP reachable from all other ranks")
369
+
370
+ jaccl_ring = args.backend == "jaccl-ring"
371
+ have_rdmas = all(len(h.rdma) == len(hosts) for h in hosts)
372
+ have_nulls = all(h.rdma[i] is None for i, h in enumerate(hosts))
373
+ if not have_rdmas or not have_nulls:
374
+ raise ValueError("Malformed hostfile for jaccl backend")
375
+
376
+ coordinator = hosts[0].ips[0]
377
+ env = args.env
378
+ cwd = args.cwd
379
+ env.append(f"MLX_JACCL_COORDINATOR={coordinator}:{args.starting_port}")
380
+ if jaccl_ring:
381
+ env.append("MLX_JACCL_RING=1")
382
+ files = {"MLX_IBV_DEVICES": json.dumps([h.rdma for h in hosts])}
383
+
384
+ log(args.verbose, "Running", shlex.join(command))
385
+
386
+ _launch_with_io(
387
+ RemoteProcess,
388
+ [
389
+ ((rank, h.ssh_hostname, args.python, cwd, files, env, command), {})
390
+ for rank, h in enumerate(hosts)
391
+ ],
392
+ args.verbose,
393
+ )
394
+
395
+
396
+ def get_mpi_libname():
397
+ try:
398
+ ompi_info = run(["which", "ompi_info"], check=True, capture_output=True)
399
+ ompi_info = ompi_info.stdout.strip().decode()
400
+
401
+ if platform.system() == "Darwin":
402
+ otool_output = run(
403
+ ["otool", "-L", ompi_info], check=True, capture_output=True
404
+ )
405
+ else:
406
+ otool_output = run(["ldd", ompi_info], check=True, capture_output=True)
407
+ otool_output = otool_output.stdout.decode()
408
+
409
+ # StopIteration if not found
410
+ libmpi_line = next(
411
+ filter(lambda line: "libmpi" in line, otool_output.splitlines())
412
+ )
413
+ return libmpi_line.strip().split()[0].removeprefix("@rpath/")
414
+ except:
415
+ return None
416
+
417
+
418
+ def launch_mpi(parser, hosts, args, command):
419
+ mpirun = run(["which", "mpirun"], check=True, capture_output=True)
420
+ mpirun = mpirun.stdout.strip().decode()
421
+
422
+ # Compatibility with homebrew and pip installs
423
+ mpi_libname = get_mpi_libname()
424
+ if mpi_libname is not None:
425
+ dyld = Path(mpirun).parent.parent / "lib"
426
+ args.env = [
427
+ f"DYLD_LIBRARY_PATH={str(dyld)}",
428
+ f"MLX_MPI_LIBNAME={mpi_libname}",
429
+ ] + args.env
430
+
431
+ log(args.verbose, f"Using '{mpirun}'")
432
+ with tempfile.NamedTemporaryFile(mode="w") as f:
433
+ hosts = Counter((h.ssh_hostname for h in hosts))
434
+ for h, n in hosts.items():
435
+ print(f"{h} slots={n}", file=f)
436
+ f.flush()
437
+
438
+ cmd = [
439
+ mpirun,
440
+ "--output",
441
+ ":raw", # do not line buffer output
442
+ "--hostfile",
443
+ f.name,
444
+ *(["-cwd", args.cwd] if args.cwd else []),
445
+ *sum((["-x", e] for e in args.env), []),
446
+ *sum([shlex.split(arg) for arg in args.mpi_arg], []),
447
+ "--",
448
+ *command,
449
+ ]
450
+ log(args.verbose, "Running", " ".join(cmd))
451
+ try:
452
+ run(cmd)
453
+ except KeyboardInterrupt:
454
+ pass
455
+
456
+
457
+ def main():
458
+ parser = argparse.ArgumentParser(description="Launch an MLX distributed program")
459
+ parser.add_argument(
460
+ "--print-python",
461
+ action="store_true",
462
+ help="Print the path to the current python executable and exit",
463
+ )
464
+ parser.add_argument(
465
+ "--verbose", action="store_true", help="Print debug messages in stdout"
466
+ )
467
+ parser.add_argument(
468
+ "--hosts", default="127.0.0.1", help="A comma separated list of hosts"
469
+ )
470
+ parser.add_argument(
471
+ "--repeat-hosts",
472
+ "-n",
473
+ type=positive_number,
474
+ default=1,
475
+ help="Repeat each host a given number of times",
476
+ )
477
+ parser.add_argument("--hostfile", help="The file containing the hosts")
478
+ parser.add_argument(
479
+ "--backend",
480
+ help="Which distributed backend to launch",
481
+ )
482
+ parser.add_argument(
483
+ "--env",
484
+ action="append",
485
+ default=[],
486
+ help="Set environment variables for the jobs",
487
+ )
488
+ parser.add_argument(
489
+ "--mpi-arg",
490
+ action="append",
491
+ default=[],
492
+ help="Arguments to pass directly to mpirun",
493
+ )
494
+ parser.add_argument(
495
+ "--connections-per-ip",
496
+ default=1,
497
+ type=int,
498
+ help="How many connections per ip to use for the ring backend",
499
+ )
500
+ parser.add_argument(
501
+ "--starting-port",
502
+ "-p",
503
+ type=int,
504
+ default=32323,
505
+ help="For the ring backend listen on this port increasing by 1 per rank and IP",
506
+ )
507
+ parser.add_argument(
508
+ "--cwd", help="Set the working directory on each node to the provided one"
509
+ )
510
+ parser.add_argument(
511
+ "--nccl-port",
512
+ type=int,
513
+ default=12345,
514
+ help="The port to use for the NCCL communication (only for nccl backend)",
515
+ )
516
+ parser.add_argument(
517
+ "--no-verify-script",
518
+ action="store_false",
519
+ dest="verify_script",
520
+ help="Do not verify that the script exists",
521
+ )
522
+ parser.add_argument(
523
+ "--python", default=sys.executable, help="Use this python on the remote hosts"
524
+ )
525
+
526
+ args, rest = parser.parse_known_args()
527
+
528
+ if args.print_python:
529
+ print(args.python)
530
+ return
531
+
532
+ if len(rest) == 0:
533
+ parser.error("No script is provided")
534
+ if rest[0] == "--":
535
+ rest.pop(0)
536
+
537
+ # Try to extract a list of hosts and corresponding ips
538
+ if args.hostfile is not None:
539
+ hostfile = Hostfile.from_file(args.hostfile)
540
+ else:
541
+ hostfile = Hostfile.from_list(args.hosts, args.repeat_hosts)
542
+
543
+ # Extract extra arguments from the hostfile
544
+ if hostfile.backend != "" and args.backend is None:
545
+ args.backend = hostfile.backend
546
+ if args.backend is None:
547
+ args.backend = "nccl" if mx.cuda.is_available() else "ring"
548
+ args.env = hostfile.envs + args.env
549
+
550
+ # Check if the script is a file and convert it to a full path
551
+ if (script := Path(rest[0])).exists() and script.is_file():
552
+ rest[0:1] = [args.python, str(script.resolve())]
553
+ elif (command := shutil.which(rest[0])) is not None:
554
+ rest[0] = command
555
+ elif args.verify_script:
556
+ raise ValueError(f"Invalid script or command {rest[0]}")
557
+
558
+ # Launch
559
+ if args.backend == "ring":
560
+ launch_ring(parser, hostfile.hosts, args, rest)
561
+ elif args.backend == "mpi":
562
+ launch_mpi(parser, hostfile.hosts, args, rest)
563
+ elif args.backend == "nccl":
564
+ launch_nccl(parser, hostfile.hosts, args, rest)
565
+ elif args.backend == "jaccl" or args.backend == "jaccl-ring":
566
+ launch_jaccl(parser, hostfile.hosts, args, rest)
567
+ else:
568
+ parser.error(
569
+ "The backend should be one of {'ring', 'mpi', 'nccl', 'jaccl', 'jaccl-ring'}"
570
+ )
@@ -0,0 +1,16 @@
1
+ # Copyright © 2023 Apple Inc.
2
+
3
+ import array
4
+ import reprlib
5
+
6
+ _old_repr_array = reprlib.Repr.repr_array
7
+
8
+
9
+ def repr_array(self, x, maxlevel):
10
+ if isinstance(x, array.array):
11
+ return _old_repr_array(self, x, maxlevel)
12
+ else:
13
+ return self.repr_instance(x, maxlevel)
14
+
15
+
16
+ reprlib.Repr.repr_array = repr_array
@@ -0,0 +1,36 @@
1
+ mlx.core.__prefix__:
2
+ from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union, ParamSpec, TypeVar
3
+ import sys
4
+ if sys.version_info >= (3, 10):
5
+ from typing import TypeAlias
6
+ else:
7
+ from typing_extensions import TypeAlias
8
+ P = ParamSpec("P")
9
+ R = TypeVar("R")
10
+
11
+ mlx.core.__suffix__:
12
+ from typing import Union
13
+ scalar: TypeAlias = Union[int, float, bool]
14
+ list_or_scalar: TypeAlias = Union[scalar, list["list_or_scalar"]]
15
+ bool_: Dtype = ...
16
+
17
+ mlx.core.distributed.__prefix__:
18
+ from mlx.core import array, Dtype, Device, Stream, scalar
19
+ from mlx.core.distributed import Group
20
+ from typing import Sequence, Optional, Union
21
+
22
+ mlx.core.fast.__prefix__:
23
+ from mlx.core import array, Dtype, Device, Stream, scalar
24
+ from typing import Sequence, Optional, Union
25
+
26
+ mlx.core.linalg.__prefix__:
27
+ from mlx.core import array, Dtype, Device, Stream, scalar
28
+ from typing import Sequence, Optional, Tuple, Union
29
+
30
+ mlx.core.metal.__prefix__:
31
+ from mlx.core import array, Dtype, Device, Stream, scalar
32
+ from typing import Sequence, Optional, Union
33
+
34
+ mlx.core.random.__prefix__:
35
+ from mlx.core import array, Dtype, Device, Stream, scalar, float32, int32
36
+ from typing import Sequence, Optional, Union
@@ -0,0 +1,88 @@
1
+ # Copyright © 2023 Apple Inc.
2
+
3
+ import os
4
+ import re
5
+ import subprocess
6
+ import sys
7
+ from pathlib import Path
8
+
9
+ from setuptools import Extension
10
+ from setuptools.command.build_ext import build_ext
11
+
12
+ import mlx
13
+
14
+ _MLX_PATH = str(mlx.__path__[0])
15
+
16
+
17
+ # A CMakeExtension needs a sourcedir instead of a file list.
18
+ class CMakeExtension(Extension):
19
+ def __init__(self, name: str, sourcedir: str = "") -> None:
20
+ super().__init__(name, sources=[])
21
+ self.sourcedir = os.fspath(Path(sourcedir).resolve())
22
+
23
+
24
+ class CMakeBuild(build_ext):
25
+ def build_extension(self, ext: CMakeExtension) -> None:
26
+ # Must be in this form due to bug in .resolve() only fixed in Python 3.10+
27
+ ext_fullpath = Path.cwd() / self.get_ext_fullpath(ext.name) # type: ignore[no-untyped-call]
28
+ extdir = ext_fullpath.parent.resolve()
29
+
30
+ debug = int(os.environ.get("DEBUG", 0)) if self.debug is None else self.debug
31
+ cfg = "Debug" if debug else "Release"
32
+
33
+ # Set Python_EXECUTABLE instead if you use PYBIND11_FINDPYTHON
34
+ # EXAMPLE_VERSION_INFO shows you how to pass a value into the C++ code
35
+ # from Python.
36
+ cmake_args = [
37
+ f"-DCMAKE_LIBRARY_OUTPUT_DIRECTORY={extdir}{os.sep}",
38
+ f"-DCMAKE_BUILD_TYPE={cfg}",
39
+ "-DBUILD_SHARED_LIBS=ON",
40
+ ]
41
+ build_args = []
42
+ # Adding CMake arguments set as environment variable
43
+ # (needed e.g. to build for ARM OSx on conda-forge)
44
+ if "CMAKE_ARGS" in os.environ:
45
+ cmake_args += [item for item in os.environ["CMAKE_ARGS"].split(" ") if item]
46
+
47
+ if sys.platform.startswith("darwin"):
48
+ # Cross-compile support for macOS - respect ARCHFLAGS if set
49
+ archs = re.findall(r"-arch (\S+)", os.environ.get("ARCHFLAGS", ""))
50
+ if archs:
51
+ cmake_args += ["-DCMAKE_OSX_ARCHITECTURES={}".format(";".join(archs))]
52
+
53
+ # Set CMAKE_BUILD_PARALLEL_LEVEL to control the parallel build level
54
+ # across all generators.
55
+ if "CMAKE_BUILD_PARALLEL_LEVEL" not in os.environ:
56
+ build_args += [f"-j{os.cpu_count()}"]
57
+
58
+ build_temp = Path(self.build_temp) / ext.name
59
+ if not build_temp.exists():
60
+ build_temp.mkdir(parents=True)
61
+
62
+ # Make sure cmake can find MLX
63
+ os.environ["MLX_DIR"] = _MLX_PATH
64
+
65
+ subprocess.run(
66
+ ["cmake", ext.sourcedir, *cmake_args], cwd=build_temp, check=True
67
+ )
68
+ subprocess.run(
69
+ ["cmake", "--build", ".", *build_args], cwd=build_temp, check=True
70
+ )
71
+
72
+ def run(self) -> None:
73
+ super().run()
74
+
75
+ # Based on https://github.com/pypa/setuptools/blob/main/setuptools/command/build_ext.py#L102
76
+ if self.inplace:
77
+ for ext in self.extensions:
78
+ if isinstance(ext, CMakeExtension):
79
+ # Resolve inplace package dir
80
+ build_py = self.get_finalized_command("build_py")
81
+ inplace_file, regular_file = self._get_inplace_equivalent(
82
+ build_py, ext
83
+ )
84
+
85
+ inplace_dir = str(Path(inplace_file).parent.resolve())
86
+ regular_dir = str(Path(regular_file).parent.resolve())
87
+
88
+ self.copy_tree(regular_dir, inplace_dir)