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 +1,600 @@
1
+ # frozen_string_literal: true
2
+
3
+ require "json"
4
+ require "open3"
5
+ require "optparse"
6
+ require "set"
7
+ require "shellwords"
8
+
9
+ module MLX
10
+ module DistributedUtils
11
+ SSHInfo = Struct.new(:can_ssh, :has_sudo, keyword_init: true) do
12
+ def to_bool
13
+ can_ssh
14
+ end
15
+ end
16
+ ThunderboltPort = Struct.new(:iface, :uuid, :connected_to, keyword_init: true)
17
+ ThunderboltHost = Struct.new(:name, :ports, keyword_init: true)
18
+
19
+ class IPConfigurator
20
+ attr_reader :ips, :hosts, :tb_hosts
21
+
22
+ def initialize(hosts = [], tb_hosts = [], uuid_reverse_index = {})
23
+ assigned = Set.new
24
+ ip_map = Hash.new { |h, key| h[key] = [] }
25
+ ip0 = 0
26
+ ip1 = 0
27
+
28
+ tb_hosts.each_with_index do |host, src_node|
29
+ host.ports.each_with_index do |port, src_port|
30
+ next if port.connected_to.nil?
31
+ next unless uuid_reverse_index.key?(port.connected_to)
32
+ next if assigned.include?([src_node, src_port])
33
+
34
+ dst_node, dst_port = uuid_reverse_index[port.connected_to]
35
+ ip_src = "192.168.#{ip0}.#{ip1 + 1}"
36
+ ip_dst = "192.168.#{ip0}.#{ip1 + 2}"
37
+ iface_src = port.iface
38
+ iface_dst = tb_hosts[dst_node].ports[dst_port].iface
39
+
40
+ ip_map[[src_node, dst_node]] << [iface_src, ip_src]
41
+ ip_map[[dst_node, src_node]] << [iface_dst, ip_dst]
42
+
43
+ assigned.add([src_node, src_port])
44
+ assigned.add([dst_node, dst_port])
45
+
46
+ ip1 += 4
47
+ if ip1 > 255
48
+ ip0 += 1
49
+ ip1 = 0
50
+ end
51
+ raise ArgumentError, "Ran out of available local IPs" if ip0 > 255
52
+ end
53
+ end
54
+
55
+ @ips = ip_map
56
+ @hosts = hosts
57
+ @tb_hosts = tb_hosts
58
+ end
59
+
60
+ def setup(verbose: false, auto_setup: false, runner: nil, input: $stdin, output: $stdout)
61
+ netmask = "255.255.255.252"
62
+ @hosts.each_with_index do |host, i|
63
+ command = +""
64
+ command << "sudo ifconfig bridge0 down\n"
65
+ @hosts.length.times do |j|
66
+ next if i == j
67
+ next unless @ips.key?([i, j])
68
+
69
+ @ips[[i, j]].zip(@ips[[j, i]]).each do |(iface, ip), (_peer_iface, peer_ip)|
70
+ command << "sudo ifconfig #{iface} inet #{ip} netmask #{netmask}\n"
71
+ command << "sudo route change #{peer_ip} -interface #{iface}\n"
72
+ end
73
+ end
74
+
75
+ if auto_setup
76
+ output.puts("Running auto setup for #{host.ssh_hostname}")
77
+ ssh_cmd = ["ssh", host.ssh_hostname, command.strip.gsub("\n", " ; ")]
78
+ MLX::DistributedUtils.log(verbose, Shellwords.join(ssh_cmd))
79
+ MLX::DistributedUtils.__send__(:run_command, ssh_cmd, runner: runner)
80
+ else
81
+ title = "Setup for #{host.ssh_hostname}"
82
+ output.puts(title)
83
+ output.puts("=" * title.length)
84
+ output.puts(command)
85
+ output.print("Enter to continue")
86
+ input.gets
87
+ end
88
+ output.puts
89
+ end
90
+ end
91
+ end
92
+
93
+ class << self
94
+ def add_ips(hosts, verbose: false, runner: nil)
95
+ hosts.each do |host|
96
+ log(verbose, "Getting the ip from", host.ssh_hostname)
97
+ ip = read_ip(host.ssh_hostname, "en0", runner: runner)
98
+ ip = read_ip(host.ssh_hostname, "en1", runner: runner) if ip.empty?
99
+
100
+ if ip.empty?
101
+ log_warning("Could not extract ip for", host.ssh_hostname)
102
+ else
103
+ host.ips << ip
104
+ end
105
+ end
106
+ end
107
+
108
+ def save_hostfile(args, hostfile)
109
+ if cfg_arg(args, :output_hostfile, nil)
110
+ File.write(cfg_arg(args, :output_hostfile, nil), JSON.pretty_generate(hostfile.to_json_obj))
111
+ else
112
+ puts "Hostfile"
113
+ puts "========"
114
+ puts JSON.pretty_generate(hostfile.to_json_obj)
115
+ end
116
+ end
117
+
118
+ def check_rdma(hosts, verbose: false, strict: true, runner: nil)
119
+ failed = false
120
+ hosts.each do |host|
121
+ log(verbose, "Checking that", host.ssh_hostname, "supports RDMA")
122
+ result = run_command(["ssh", host.ssh_hostname, "ibv_devices"], runner: runner)
123
+ devices = stdout_for(result).split.map { |entry| entry.strip }
124
+ rdma = devices.select { |entry| entry.start_with?("rdma_") }
125
+ next unless rdma.empty?
126
+
127
+ (strict ? method(:log_error) : method(:log_warning)).call(
128
+ host.ssh_hostname,
129
+ "does not seem to have RDMA enabled"
130
+ )
131
+ failed = true
132
+ end
133
+
134
+ if failed
135
+ logger = strict ? method(:log_error) : method(:log_warning)
136
+ logger.call
137
+ logger.call("Some of the hosts don't have RDMA enabled or they don't support RDMA.")
138
+ logger.call
139
+ logger.call("See https://ml-explore.github.io/mlx/build/html/usage/distributed.html")
140
+ logger.call("for instructions on how to enable RDMA.")
141
+ end
142
+
143
+ raise ArgumentError, "RDMA validation failed for one or more hosts" if failed && strict
144
+
145
+ !failed
146
+ end
147
+
148
+ def can_auto_setup(hosts, sshinfo, auto_setup: false)
149
+ has_sudo = sshinfo.all?(&:has_sudo)
150
+ if !has_sudo && auto_setup
151
+ log_warning("Automatic setup requested but some hosts do not have passwordless sudo")
152
+ hosts.zip(sshinfo).each do |host, info|
153
+ log_warning(" - #{host.ssh_hostname}") unless info.has_sudo
154
+ end
155
+ end
156
+ has_sudo
157
+ end
158
+
159
+ def parse_hardware_ports(ports_string)
160
+ ports = {}
161
+ port_name = nil
162
+ ports_string.to_s.split("\n").each do |line|
163
+ if line.start_with?("Hardware Port:")
164
+ port_name = line.strip[15..]
165
+ elsif line.start_with?("Device:")
166
+ ports[port_name] = line.strip[8..]
167
+ port_name = nil
168
+ end
169
+ end
170
+ ports
171
+ end
172
+
173
+ def extract_connectivity(hosts, verbose, runner: nil)
174
+ thunderbolt_connections = []
175
+ hosts.each do |host|
176
+ log(verbose, "Getting connectivity from", host.ssh_hostname)
177
+ result = run_command(
178
+ ["ssh", host.ssh_hostname, "system_profiler", "SPThunderboltDataType", "-json"],
179
+ runner: runner
180
+ )
181
+ thunderbolt_connections << JSON.parse(stdout_for(result))
182
+ end
183
+
184
+ interface_maps = []
185
+ hosts.each do |host|
186
+ log(verbose, "Getting interface names from", host.ssh_hostname)
187
+ result = run_command(
188
+ ["ssh", host.ssh_hostname, "networksetup", "-listallhardwareports"],
189
+ runner: runner
190
+ )
191
+ interface_maps << parse_hardware_ports(stdout_for(result))
192
+ end
193
+
194
+ tb_hosts = []
195
+ thunderbolt_connections.zip(interface_maps).each do |connection, iface_map|
196
+ name = ""
197
+ ports = []
198
+ Array(connection["SPThunderboltDataType"]).each do |entry|
199
+ uuid = entry["domain_uuid_key"]
200
+ next if uuid.nil?
201
+
202
+ name = entry["device_name_key"].to_s
203
+ tag = entry.dig("receptacle_1_tag", "receptacle_id_key")
204
+ items = Array(entry["_items"])
205
+ connected = items.find { |item| item.key?("domain_uuid_key") }
206
+ connected_to = connected.nil? ? nil : connected["domain_uuid_key"]
207
+ iface = iface_map["Thunderbolt #{tag}"]
208
+ ports << ThunderboltPort.new(iface: iface, uuid: uuid, connected_to: connected_to)
209
+ end
210
+ tb_hosts << ThunderboltHost.new(name: name, ports: ports.sort_by { |port| port.iface.to_s })
211
+ end
212
+
213
+ uuid_reverse_index = {}
214
+ tb_hosts.each_with_index do |host, i|
215
+ host.ports.each_with_index do |port, j|
216
+ uuid_reverse_index[port.uuid] = [i, j]
217
+ end
218
+ end
219
+
220
+ [tb_hosts, uuid_reverse_index]
221
+ end
222
+
223
+ def make_connectivity_matrix(tb_hosts, uuid_reverse_index)
224
+ connectivity = []
225
+ tb_hosts.each_with_index do |host, i|
226
+ row = Array.new(tb_hosts.length, 0)
227
+ host.ports.each do |port|
228
+ next unless uuid_reverse_index.key?(port.connected_to)
229
+
230
+ j, = uuid_reverse_index[port.connected_to]
231
+ row[j] += 1
232
+ end
233
+ connectivity[i] = row
234
+ end
235
+ connectivity
236
+ end
237
+
238
+ def tb_connectivity_to_dot(hosts, tb_hosts, uuid_reverse_index, io: $stdout)
239
+ names = []
240
+ tb_hosts.length.times do |i|
241
+ value = +""
242
+ j = i
243
+ loop do
244
+ value << (97 + (j % 26)).chr
245
+ j /= 26
246
+ break if j.zero?
247
+ end
248
+ names << value
249
+ end
250
+
251
+ io.puts("graph G {")
252
+ io.puts(" node [shape=rectangle];")
253
+ hosts.each_with_index do |host, i|
254
+ io.puts(" #{names[i]} [label=\"#{host.ssh_hostname}\"];")
255
+ end
256
+ tb_hosts.each_with_index do |host, i|
257
+ host.ports.each do |port|
258
+ next if port.connected_to.nil?
259
+
260
+ dst = uuid_reverse_index[port.connected_to]
261
+ next if dst.nil? || dst[0] < i
262
+
263
+ io.puts(" #{names[i]} -- #{names[dst[0]]} [label=\"#{port.iface}/#{tb_hosts[dst[0]].ports[dst[1]].iface}\"]")
264
+ end
265
+ end
266
+ io.puts("}")
267
+ end
268
+
269
+ def extract_rings(connectivity)
270
+ rings = []
271
+ existing = Set.new
272
+ num_nodes = connectivity.length
273
+
274
+ dfs = lambda do |start_node, node, path, visited, &blk|
275
+ path << node
276
+ visited.add(node)
277
+ num_nodes.times do |j|
278
+ next unless connectivity[node][j] > 0
279
+
280
+ if j == start_node
281
+ blk.call(path.dup)
282
+ elsif !visited.include?(j)
283
+ dfs.call(start_node, j, path, visited, &blk)
284
+ end
285
+ end
286
+ path.pop
287
+ visited.delete(node)
288
+ end
289
+
290
+ num_nodes.times do |start|
291
+ dfs.call(start, start, [], Set.new) do |ring|
292
+ count = ring.length.times.map do |i|
293
+ connectivity[ring[i]][ring[(i + 1) % ring.length]]
294
+ end.min
295
+ key = ring.sort
296
+ next if existing.include?(key)
297
+
298
+ rings << [ring, count]
299
+ existing.add(key)
300
+ end
301
+ end
302
+
303
+ rings.sort_by { |ring, _count| -ring.length }
304
+ end
305
+
306
+ def check_valid_mesh(hosts, connectivity, strict: true)
307
+ num_nodes = connectivity.length
308
+ num_nodes.times do |i|
309
+ num_nodes.times do |j|
310
+ next if i == j
311
+ next if connectivity[i][j] > 0
312
+
313
+ if strict
314
+ raise ArgumentError,
315
+ "Incomplete mesh, #{hosts[i].ssh_hostname} is not connected to #{hosts[j].ssh_hostname}"
316
+ else
317
+ return false
318
+ end
319
+ end
320
+ end
321
+ true
322
+ end
323
+
324
+ def check_valid_ring(hosts, rings, strict: true)
325
+ has_ring = !rings.empty? && rings[0][0].length == hosts.length
326
+ raise ArgumentError, "Could not find a full ring." if strict && !has_ring
327
+
328
+ has_ring
329
+ end
330
+
331
+ def check_ssh_connections(hosts, runner: nil)
332
+ infos = hosts.map do |host|
333
+ ssh_result = run_command(
334
+ [
335
+ "ssh",
336
+ "-o",
337
+ "BatchMode=yes",
338
+ "-o",
339
+ "ConnectTimeout=5",
340
+ host.ssh_hostname,
341
+ "echo",
342
+ "success"
343
+ ],
344
+ runner: runner
345
+ )
346
+ can_ssh = success_for(ssh_result)
347
+ has_sudo = false
348
+
349
+ if can_ssh
350
+ sudo_result = run_command(
351
+ [
352
+ "ssh",
353
+ "-o",
354
+ "BatchMode=yes",
355
+ "-o",
356
+ "ConnectTimeout=5",
357
+ host.ssh_hostname,
358
+ "sudo",
359
+ "ls"
360
+ ],
361
+ runner: runner
362
+ )
363
+ has_sudo = success_for(sudo_result)
364
+ end
365
+
366
+ SSHInfo.new(can_ssh: can_ssh, has_sudo: has_sudo)
367
+ end
368
+
369
+ unless infos.all?(&:can_ssh)
370
+ bad_hosts = hosts.zip(infos).filter_map { |host, info| host.ssh_hostname unless info.can_ssh }
371
+ raise ArgumentError, "Could not ssh to the following hosts: #{bad_hosts.join(', ')}"
372
+ end
373
+
374
+ infos
375
+ end
376
+
377
+ def prepare_ethernet_hostfile(args, hosts, runner: nil)
378
+ log(cfg_arg(args, :verbose, false), "Preparing an ethernet hostfile")
379
+ add_ips(hosts, verbose: cfg_arg(args, :verbose, false), runner: runner)
380
+ hostfile_hosts = hosts.each_with_index.map do |host, i|
381
+ Host.new(rank: i, ssh_hostname: host.ssh_hostname, ips: host.ips, rdma: [])
382
+ end
383
+ hostfile = Hostfile.new(hostfile_hosts, "", cfg_arg(args, :env, []))
384
+ save_hostfile(args, hostfile)
385
+ end
386
+
387
+ def configure_ring(args, hosts, ips, ring, sshinfo, runner: nil)
388
+ log(cfg_arg(args, :verbose, false), "Prepare a ring hostfile")
389
+ ring_nodes, count = ring
390
+ ring_hosts = []
391
+ ring_nodes.each_with_index do |node, i|
392
+ host = hosts[node]
393
+ peer = ring_nodes[i - 1]
394
+ peer_ips = count.times.map { |c| ips.ips[[node, peer]][c][1] }
395
+ ring_hosts << Host.new(rank: i, ssh_hostname: host.ssh_hostname, ips: peer_ips, rdma: [])
396
+ end
397
+ hostfile = Hostfile.new(ring_hosts, "ring", cfg_arg(args, :env, []))
398
+
399
+ has_sudo = can_auto_setup(hosts, sshinfo, auto_setup: cfg_arg(args, :auto_setup, false))
400
+ ips.setup(
401
+ verbose: cfg_arg(args, :verbose, false),
402
+ auto_setup: cfg_arg(args, :auto_setup, false) && has_sudo,
403
+ runner: runner
404
+ )
405
+ save_hostfile(args, hostfile)
406
+ end
407
+
408
+ def configure_jaccl(args, hosts, ips, sshinfo, runner: nil)
409
+ log(cfg_arg(args, :verbose, false), "Prepare a jaccl hostfile")
410
+ add_ips(hosts, verbose: cfg_arg(args, :verbose, false), runner: runner)
411
+
412
+ jaccl_hosts = hosts.each_with_index.map do |host, i|
413
+ rdma = hosts.each_index.map do |j|
414
+ if i == j
415
+ nil
416
+ else
417
+ "rdma_#{ips.ips[[i, j]][0][0]}"
418
+ end
419
+ end
420
+ Host.new(rank: i, ssh_hostname: host.ssh_hostname, ips: host.ips, rdma: rdma)
421
+ end
422
+ hostfile = Hostfile.new(jaccl_hosts, "jaccl", cfg_arg(args, :env, []))
423
+
424
+ has_sudo = can_auto_setup(hosts, sshinfo, auto_setup: cfg_arg(args, :auto_setup, false))
425
+ ips.setup(
426
+ verbose: cfg_arg(args, :verbose, false),
427
+ auto_setup: cfg_arg(args, :auto_setup, false) && has_sudo,
428
+ runner: runner
429
+ )
430
+ save_hostfile(args, hostfile)
431
+ end
432
+
433
+ def configure_jaccl_ring(args, hosts, ips, ring, sshinfo, runner: nil)
434
+ log(cfg_arg(args, :verbose, false), "Prepare a jaccl-ring hostfile")
435
+ add_ips(hosts, verbose: cfg_arg(args, :verbose, false), runner: runner)
436
+
437
+ ring_nodes, count = ring
438
+ num_nodes = hosts.length
439
+ jaccl_hosts = []
440
+ ring_nodes.each_with_index do |node, i|
441
+ host = hosts[node]
442
+ peer_left = ring_nodes[i - 1]
443
+ peer_right = ring_nodes[(i + 1) % num_nodes]
444
+ rdmas = hosts.each_index.map do |j|
445
+ if j != peer_left && j != peer_right
446
+ nil
447
+ else
448
+ rdma_values = count.times.map { |c| "rdma_#{ips.ips[[i, j]][c][0]}" }
449
+ count == 1 ? rdma_values[0] : rdma_values
450
+ end
451
+ end
452
+ jaccl_hosts << Host.new(rank: i, ssh_hostname: host.ssh_hostname, ips: host.ips, rdma: rdmas)
453
+ end
454
+ hostfile = Hostfile.new(jaccl_hosts, "jaccl-ring", cfg_arg(args, :env, []))
455
+
456
+ has_sudo = can_auto_setup(hosts, sshinfo, auto_setup: cfg_arg(args, :auto_setup, false))
457
+ ips.setup(
458
+ verbose: cfg_arg(args, :verbose, false),
459
+ auto_setup: cfg_arg(args, :auto_setup, false) && has_sudo,
460
+ runner: runner
461
+ )
462
+ save_hostfile(args, hostfile)
463
+ end
464
+
465
+ def prepare_tb_hostfile(args, hosts, sshinfo, runner: nil)
466
+ log(cfg_arg(args, :verbose, false), "Preparing for communication over thunderbolt")
467
+ tb_hosts, uuid_reverse_index = extract_connectivity(hosts, cfg_arg(args, :verbose, false), runner: runner)
468
+
469
+ if cfg_arg(args, :dot, false)
470
+ tb_connectivity_to_dot(hosts, tb_hosts, uuid_reverse_index)
471
+ return
472
+ end
473
+
474
+ ips = IPConfigurator.new(hosts, tb_hosts, uuid_reverse_index)
475
+ connectivity = make_connectivity_matrix(tb_hosts, uuid_reverse_index)
476
+ backend = cfg_arg(args, :backend, nil)
477
+
478
+ if backend.nil?
479
+ rings = extract_rings(connectivity)
480
+ has_mesh = check_valid_mesh(hosts, connectivity, strict: false)
481
+ has_ring = check_valid_ring(hosts, rings, strict: false)
482
+ has_rdma = check_rdma(hosts, cfg_arg(args, :verbose, false), strict: false, runner: runner)
483
+
484
+ if !has_ring && !has_mesh
485
+ raise ArgumentError, "Neither thunderbolt mesh nor ring found."
486
+ elsif has_rdma && has_mesh
487
+ configure_jaccl(args, hosts, ips, sshinfo, runner: runner)
488
+ elsif has_rdma && has_ring
489
+ configure_jaccl_ring(args, hosts, ips, rings[0], sshinfo, runner: runner)
490
+ elsif has_ring
491
+ configure_ring(args, hosts, ips, rings[0], sshinfo, runner: runner)
492
+ else
493
+ raise ArgumentError, "RDMA is not available and ring is not found."
494
+ end
495
+ elsif backend == "ring"
496
+ rings = extract_rings(connectivity)
497
+ check_valid_ring(hosts, rings)
498
+ configure_ring(args, hosts, ips, rings[0], sshinfo, runner: runner)
499
+ elsif backend == "jaccl"
500
+ check_valid_mesh(hosts, connectivity)
501
+ check_rdma(hosts, cfg_arg(args, :verbose, false), runner: runner)
502
+ configure_jaccl(args, hosts, ips, sshinfo, runner: runner)
503
+ elsif backend == "jaccl-ring"
504
+ rings = extract_rings(connectivity)
505
+ check_valid_ring(hosts, rings)
506
+ check_rdma(hosts, cfg_arg(args, :verbose, false), runner: runner)
507
+ configure_jaccl_ring(args, hosts, ips, rings[0], sshinfo, runner: runner)
508
+ end
509
+ end
510
+
511
+ def config_main(argv = ARGV, runner: nil)
512
+ opts = {
513
+ verbose: false,
514
+ hosts: "127.0.0.1",
515
+ hostfile: nil,
516
+ over: "thunderbolt",
517
+ output_hostfile: nil,
518
+ auto_setup: nil,
519
+ dot: false,
520
+ backend: nil,
521
+ env: []
522
+ }
523
+
524
+ parser = OptionParser.new do |o|
525
+ o.on("--verbose") { opts[:verbose] = true }
526
+ o.on("--hosts HOSTS") { |v| opts[:hosts] = v }
527
+ o.on("--hostfile FILE") { |v| opts[:hostfile] = v }
528
+ o.on("--over TYPE") { |v| opts[:over] = v }
529
+ o.on("--output-hostfile FILE") { |v| opts[:output_hostfile] = v }
530
+ o.on("--auto-setup") { opts[:auto_setup] = true }
531
+ o.on("--no-auto-setup") { opts[:auto_setup] = false }
532
+ o.on("--dot") { opts[:dot] = true }
533
+ o.on("--backend BACKEND") { |v| opts[:backend] = v }
534
+ o.on("--env ENV") { |v| opts[:env] << v }
535
+ end
536
+ parser.parse!(argv.dup)
537
+
538
+ hosts = if opts[:hostfile]
539
+ Hostfile.from_file(opts[:hostfile]).hosts
540
+ else
541
+ Hostfile.from_list(opts[:hosts]).hosts
542
+ end
543
+
544
+ log(opts[:verbose], "Checking for ssh access for #{hosts.map(&:ssh_hostname).join(', ')}")
545
+ sshinfo = check_ssh_connections(hosts, runner: runner)
546
+
547
+ args = opts
548
+ if opts[:over] == "ethernet"
549
+ prepare_ethernet_hostfile(args, hosts, runner: runner)
550
+ else
551
+ prepare_tb_hostfile(args, hosts, sshinfo, runner: runner)
552
+ end
553
+ 0
554
+ end
555
+ private :config_main
556
+
557
+ private
558
+
559
+ def cfg_arg(args, key, default)
560
+ if args.respond_to?(key)
561
+ value = args.public_send(key)
562
+ value.nil? ? default : value
563
+ elsif args.respond_to?(:[])
564
+ value = args[key]
565
+ value.nil? ? default : value
566
+ else
567
+ default
568
+ end
569
+ end
570
+
571
+ def read_ip(hostname, iface, runner: nil)
572
+ result = run_command(["ssh", hostname, "ipconfig", "getifaddr", iface], runner: runner)
573
+ stdout_for(result).strip
574
+ end
575
+
576
+ def run_command(cmd, runner: nil)
577
+ return runner.call(cmd) unless runner.nil?
578
+
579
+ stdout, stderr, status = Open3.capture3(*cmd)
580
+ Struct.new(:stdout, :stderr, :status, keyword_init: true).new(stdout: stdout, stderr: stderr, status: status)
581
+ end
582
+
583
+ def stdout_for(result)
584
+ if result.respond_to?(:stdout)
585
+ result.stdout.to_s
586
+ else
587
+ ""
588
+ end
589
+ end
590
+
591
+ def success_for(result)
592
+ status = result.respond_to?(:status) ? result.status : result
593
+ return status.success? if status.respond_to?(:success?)
594
+ return status.exitstatus.zero? if status.respond_to?(:exitstatus)
595
+
596
+ true
597
+ end
598
+ end
599
+ end
600
+ end