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.
- checksums.yaml +7 -0
- data/ext/mlx/CMakeLists.txt +7 -0
- data/ext/mlx/Makefile +273 -0
- data/ext/mlx/extconf.rb +94 -0
- data/ext/mlx/mkmf.log +44 -0
- data/ext/mlx/native.bundle +0 -0
- data/ext/mlx/native.bundle.dSYM/Contents/Info.plist +20 -0
- data/ext/mlx/native.bundle.dSYM/Contents/Resources/DWARF/native.bundle +0 -0
- data/ext/mlx/native.bundle.dSYM/Contents/Resources/Relocations/aarch64/native.bundle.yml +5 -0
- data/ext/mlx/native.cpp +8027 -0
- data/ext/mlx/native.o +0 -0
- data/lib/mlx/core.rb +1678 -0
- data/lib/mlx/distributed_utils/common.rb +116 -0
- data/lib/mlx/distributed_utils/config.rb +600 -0
- data/lib/mlx/distributed_utils/launch.rb +490 -0
- data/lib/mlx/extension.rb +24 -0
- data/lib/mlx/nn/base.rb +388 -0
- data/lib/mlx/nn/init.rb +140 -0
- data/lib/mlx/nn/layers/activations.rb +336 -0
- data/lib/mlx/nn/layers/base.rb +6 -0
- data/lib/mlx/nn/layers/containers.rb +20 -0
- data/lib/mlx/nn/layers/convolution.rb +120 -0
- data/lib/mlx/nn/layers/convolution_transpose.rb +114 -0
- data/lib/mlx/nn/layers/distributed.rb +309 -0
- data/lib/mlx/nn/layers/dropout.rb +75 -0
- data/lib/mlx/nn/layers/embedding.rb +28 -0
- data/lib/mlx/nn/layers/linear.rb +79 -0
- data/lib/mlx/nn/layers/normalization.rb +216 -0
- data/lib/mlx/nn/layers/pooling.rb +167 -0
- data/lib/mlx/nn/layers/positional_encoding.rb +126 -0
- data/lib/mlx/nn/layers/quantized.rb +215 -0
- data/lib/mlx/nn/layers/recurrent.rb +135 -0
- data/lib/mlx/nn/layers/transformer.rb +330 -0
- data/lib/mlx/nn/layers/upsample.rb +97 -0
- data/lib/mlx/nn/layers.rb +18 -0
- data/lib/mlx/nn/losses.rb +251 -0
- data/lib/mlx/nn/utils.rb +167 -0
- data/lib/mlx/nn.rb +12 -0
- data/lib/mlx/optimizers/optimizers.rb +808 -0
- data/lib/mlx/optimizers/schedulers.rb +62 -0
- data/lib/mlx/optimizers.rb +9 -0
- data/lib/mlx/utils.rb +171 -0
- data/lib/mlx/version +1 -0
- data/lib/mlx/version.rb +5 -0
- data/lib/mlx.rb +64 -0
- data/mlx/.clang-format +87 -0
- data/mlx/.git +1 -0
- data/mlx/.github/ISSUE_TEMPLATE/bug_report.md +28 -0
- data/mlx/.github/actions/build-cuda-release/action.yml +31 -0
- data/mlx/.github/actions/build-docs/action.yml +38 -0
- data/mlx/.github/actions/build-linux/action.yml +38 -0
- data/mlx/.github/actions/build-linux-release/action.yml +42 -0
- data/mlx/.github/actions/build-macos/action.yml +80 -0
- data/mlx/.github/actions/build-macos-release/action.yml +36 -0
- data/mlx/.github/actions/build-windows/action.yml +26 -0
- data/mlx/.github/actions/setup-linux/action.yml +93 -0
- data/mlx/.github/actions/setup-macos/action.yml +24 -0
- data/mlx/.github/actions/setup-windows/action.yml +42 -0
- data/mlx/.github/actions/test-linux/action.yml +69 -0
- data/mlx/.github/actions/test-windows/action.yml +20 -0
- data/mlx/.github/dependabot.yml +6 -0
- data/mlx/.github/pull_request_template.md +12 -0
- data/mlx/.github/scripts/build-sanitizer-tests.sh +48 -0
- data/mlx/.github/scripts/setup+build-cpp-linux-fedora-container.sh +27 -0
- data/mlx/.github/workflows/build_and_test.yml +152 -0
- data/mlx/.github/workflows/documentation.yml +28 -0
- data/mlx/.github/workflows/nightly.yml +104 -0
- data/mlx/.github/workflows/release.yml +256 -0
- data/mlx/.gitignore +81 -0
- data/mlx/.pre-commit-config.yaml +27 -0
- data/mlx/ACKNOWLEDGMENTS.md +268 -0
- data/mlx/CITATION.cff +24 -0
- data/mlx/CMakeLists.txt +437 -0
- data/mlx/CODE_OF_CONDUCT.md +132 -0
- data/mlx/CONTRIBUTING.md +38 -0
- data/mlx/LICENSE +21 -0
- data/mlx/MANIFEST.in +6 -0
- data/mlx/README.md +121 -0
- data/mlx/benchmarks/cpp/CMakeLists.txt +11 -0
- data/mlx/benchmarks/cpp/autograd.cpp +39 -0
- data/mlx/benchmarks/cpp/compare_devices.cpp +27 -0
- data/mlx/benchmarks/cpp/irregular_strides.cpp +201 -0
- data/mlx/benchmarks/cpp/single_ops.cpp +288 -0
- data/mlx/benchmarks/cpp/time_utils.h +39 -0
- data/mlx/benchmarks/numpy/single_ops.py +39 -0
- data/mlx/benchmarks/numpy/time_utils.py +20 -0
- data/mlx/benchmarks/python/batch_matmul_bench.py +62 -0
- data/mlx/benchmarks/python/blas/bench_gemm.py +191 -0
- data/mlx/benchmarks/python/blas/bench_gemv.py +220 -0
- data/mlx/benchmarks/python/comparative/README.md +15 -0
- data/mlx/benchmarks/python/comparative/bench_mlx.py +519 -0
- data/mlx/benchmarks/python/comparative/bench_torch.py +482 -0
- data/mlx/benchmarks/python/comparative/compare.py +284 -0
- data/mlx/benchmarks/python/compile_bench.py +107 -0
- data/mlx/benchmarks/python/conv1d_bench.py +123 -0
- data/mlx/benchmarks/python/conv2d_bench_cpu.py +127 -0
- data/mlx/benchmarks/python/conv2d_train_bench_cpu.py +143 -0
- data/mlx/benchmarks/python/conv2d_transpose_bench_cpu.py +129 -0
- data/mlx/benchmarks/python/conv3d_bench_cpu.py +110 -0
- data/mlx/benchmarks/python/conv3d_train_bench_cpu.py +143 -0
- data/mlx/benchmarks/python/conv3d_transpose_bench_cpu.py +116 -0
- data/mlx/benchmarks/python/conv_bench.py +135 -0
- data/mlx/benchmarks/python/conv_transpose_bench.py +135 -0
- data/mlx/benchmarks/python/conv_unaligned_bench.py +107 -0
- data/mlx/benchmarks/python/distributed_bench.py +66 -0
- data/mlx/benchmarks/python/einsum_bench.py +84 -0
- data/mlx/benchmarks/python/fft_bench.py +118 -0
- data/mlx/benchmarks/python/gather_bench.py +52 -0
- data/mlx/benchmarks/python/gather_mm_bench.py +74 -0
- data/mlx/benchmarks/python/gather_qmm_bench.py +84 -0
- data/mlx/benchmarks/python/hadamard_bench.py +70 -0
- data/mlx/benchmarks/python/large_gemm_bench.py +119 -0
- data/mlx/benchmarks/python/layer_norm_bench.py +82 -0
- data/mlx/benchmarks/python/masked_scatter.py +212 -0
- data/mlx/benchmarks/python/rms_norm_bench.py +63 -0
- data/mlx/benchmarks/python/rope_bench.py +35 -0
- data/mlx/benchmarks/python/scatter_bench.py +96 -0
- data/mlx/benchmarks/python/sdpa_bench.py +223 -0
- data/mlx/benchmarks/python/sdpa_vector_bench.py +95 -0
- data/mlx/benchmarks/python/single_ops.py +132 -0
- data/mlx/benchmarks/python/synchronize_bench.py +55 -0
- data/mlx/benchmarks/python/time_utils.py +38 -0
- data/mlx/cmake/FindCUDNN.cmake +177 -0
- data/mlx/cmake/FindNCCL.cmake +54 -0
- data/mlx/cmake/Findnvpl.cmake +3 -0
- data/mlx/cmake/extension.cmake +50 -0
- data/mlx/docs/.clang-format +2 -0
- data/mlx/docs/.gitignore +3 -0
- data/mlx/docs/.nojekyll +0 -0
- data/mlx/docs/Doxyfile +51 -0
- data/mlx/docs/Makefile +18 -0
- data/mlx/docs/README.md +54 -0
- data/mlx/docs/index.html +1 -0
- data/mlx/docs/requirements.txt +5 -0
- data/mlx/docs/src/_static/distributed/m3-ultra-mesh-broken.png +0 -0
- data/mlx/docs/src/_static/distributed/m3-ultra-mesh.png +0 -0
- data/mlx/docs/src/_static/metal_debugger/capture.png +0 -0
- data/mlx/docs/src/_static/metal_debugger/schema.png +0 -0
- data/mlx/docs/src/_static/mlx_logo.png +0 -0
- data/mlx/docs/src/_static/mlx_logo_dark.png +0 -0
- data/mlx/docs/src/_static/tp_inference/all-to-sharded-linear.png +0 -0
- data/mlx/docs/src/_static/tp_inference/column-row-tp.png +0 -0
- data/mlx/docs/src/_static/tp_inference/llama-transformer.png +0 -0
- data/mlx/docs/src/_static/tp_inference/sharded-to-all-linear.png +0 -0
- data/mlx/docs/src/_templates/module-base-class.rst +33 -0
- data/mlx/docs/src/_templates/nn-module-template.rst +20 -0
- data/mlx/docs/src/_templates/optimizers-template.rst +20 -0
- data/mlx/docs/src/conf.py +99 -0
- data/mlx/docs/src/cpp/ops.rst +7 -0
- data/mlx/docs/src/dev/custom_metal_kernels.rst +445 -0
- data/mlx/docs/src/dev/extensions.rst +811 -0
- data/mlx/docs/src/dev/metal_debugger.rst +68 -0
- data/mlx/docs/src/dev/metal_logging.rst +40 -0
- data/mlx/docs/src/dev/mlx_in_cpp.rst +121 -0
- data/mlx/docs/src/examples/data_parallelism.rst +91 -0
- data/mlx/docs/src/examples/linear_regression.rst +77 -0
- data/mlx/docs/src/examples/llama-inference.rst +382 -0
- data/mlx/docs/src/examples/mlp.rst +134 -0
- data/mlx/docs/src/examples/tensor_parallelism.rst +239 -0
- data/mlx/docs/src/index.rst +96 -0
- data/mlx/docs/src/install.rst +340 -0
- data/mlx/docs/src/python/array.rst +65 -0
- data/mlx/docs/src/python/cuda.rst +9 -0
- data/mlx/docs/src/python/data_types.rst +78 -0
- data/mlx/docs/src/python/devices_and_streams.rst +21 -0
- data/mlx/docs/src/python/distributed.rst +22 -0
- data/mlx/docs/src/python/export.rst +14 -0
- data/mlx/docs/src/python/fast.rst +16 -0
- data/mlx/docs/src/python/fft.rst +24 -0
- data/mlx/docs/src/python/linalg.rst +27 -0
- data/mlx/docs/src/python/memory_management.rst +16 -0
- data/mlx/docs/src/python/metal.rst +12 -0
- data/mlx/docs/src/python/nn/distributed.rst +30 -0
- data/mlx/docs/src/python/nn/functions.rst +40 -0
- data/mlx/docs/src/python/nn/init.rst +45 -0
- data/mlx/docs/src/python/nn/layers.rst +74 -0
- data/mlx/docs/src/python/nn/losses.rst +25 -0
- data/mlx/docs/src/python/nn/module.rst +38 -0
- data/mlx/docs/src/python/nn.rst +186 -0
- data/mlx/docs/src/python/ops.rst +184 -0
- data/mlx/docs/src/python/optimizers/common_optimizers.rst +22 -0
- data/mlx/docs/src/python/optimizers/optimizer.rst +23 -0
- data/mlx/docs/src/python/optimizers/schedulers.rst +15 -0
- data/mlx/docs/src/python/optimizers.rst +78 -0
- data/mlx/docs/src/python/random.rst +48 -0
- data/mlx/docs/src/python/transforms.rst +22 -0
- data/mlx/docs/src/python/tree_utils.rst +23 -0
- data/mlx/docs/src/usage/compile.rst +516 -0
- data/mlx/docs/src/usage/distributed.rst +572 -0
- data/mlx/docs/src/usage/export.rst +288 -0
- data/mlx/docs/src/usage/function_transforms.rst +191 -0
- data/mlx/docs/src/usage/indexing.rst +194 -0
- data/mlx/docs/src/usage/launching_distributed.rst +234 -0
- data/mlx/docs/src/usage/lazy_evaluation.rst +144 -0
- data/mlx/docs/src/usage/numpy.rst +124 -0
- data/mlx/docs/src/usage/quick_start.rst +67 -0
- data/mlx/docs/src/usage/saving_and_loading.rst +81 -0
- data/mlx/docs/src/usage/unified_memory.rst +78 -0
- data/mlx/docs/src/usage/using_streams.rst +18 -0
- data/mlx/examples/cmake_project/CMakeLists.txt +22 -0
- data/mlx/examples/cmake_project/README.md +26 -0
- data/mlx/examples/cmake_project/example.cpp +14 -0
- data/mlx/examples/cpp/CMakeLists.txt +12 -0
- data/mlx/examples/cpp/distributed.cpp +22 -0
- data/mlx/examples/cpp/linear_regression.cpp +54 -0
- data/mlx/examples/cpp/logistic_regression.cpp +54 -0
- data/mlx/examples/cpp/metal_capture.cpp +31 -0
- data/mlx/examples/cpp/timer.h +20 -0
- data/mlx/examples/cpp/tutorial.cpp +99 -0
- data/mlx/examples/export/CMakeLists.txt +22 -0
- data/mlx/examples/export/README.md +49 -0
- data/mlx/examples/export/eval_mlp.cpp +25 -0
- data/mlx/examples/export/eval_mlp.py +52 -0
- data/mlx/examples/export/train_mlp.cpp +35 -0
- data/mlx/examples/export/train_mlp.py +76 -0
- data/mlx/examples/extensions/CMakeLists.txt +78 -0
- data/mlx/examples/extensions/README.md +24 -0
- data/mlx/examples/extensions/axpby/axpby.cpp +306 -0
- data/mlx/examples/extensions/axpby/axpby.h +90 -0
- data/mlx/examples/extensions/axpby/axpby.metal +47 -0
- data/mlx/examples/extensions/bindings.cpp +39 -0
- data/mlx/examples/extensions/mlx_sample_extensions/__init__.py +5 -0
- data/mlx/examples/extensions/pyproject.toml +8 -0
- data/mlx/examples/extensions/requirements.txt +4 -0
- data/mlx/examples/extensions/setup.py +18 -0
- data/mlx/examples/extensions/test.py +12 -0
- data/mlx/examples/python/linear_regression.py +46 -0
- data/mlx/examples/python/logistic_regression.py +49 -0
- data/mlx/examples/python/qqmm.py +117 -0
- data/mlx/mlx/3rdparty/.clang-format +2 -0
- data/mlx/mlx/3rdparty/pocketfft.h +3581 -0
- data/mlx/mlx/CMakeLists.txt +107 -0
- data/mlx/mlx/allocator.h +75 -0
- data/mlx/mlx/api.h +29 -0
- data/mlx/mlx/array.cpp +354 -0
- data/mlx/mlx/array.h +647 -0
- data/mlx/mlx/backend/common/CMakeLists.txt +9 -0
- data/mlx/mlx/backend/common/binary.h +97 -0
- data/mlx/mlx/backend/common/broadcasting.cpp +24 -0
- data/mlx/mlx/backend/common/broadcasting.h +11 -0
- data/mlx/mlx/backend/common/buffer_cache.h +158 -0
- data/mlx/mlx/backend/common/common.cpp +305 -0
- data/mlx/mlx/backend/common/compiled.cpp +243 -0
- data/mlx/mlx/backend/common/compiled.h +77 -0
- data/mlx/mlx/backend/common/copy.h +50 -0
- data/mlx/mlx/backend/common/hadamard.h +109 -0
- data/mlx/mlx/backend/common/load.cpp +57 -0
- data/mlx/mlx/backend/common/matmul.h +67 -0
- data/mlx/mlx/backend/common/reduce.cpp +154 -0
- data/mlx/mlx/backend/common/reduce.h +59 -0
- data/mlx/mlx/backend/common/slicing.cpp +71 -0
- data/mlx/mlx/backend/common/slicing.h +20 -0
- data/mlx/mlx/backend/common/ternary.h +85 -0
- data/mlx/mlx/backend/common/unary.h +29 -0
- data/mlx/mlx/backend/common/utils.cpp +231 -0
- data/mlx/mlx/backend/common/utils.h +205 -0
- data/mlx/mlx/backend/cpu/CMakeLists.txt +88 -0
- data/mlx/mlx/backend/cpu/arange.h +28 -0
- data/mlx/mlx/backend/cpu/arg_reduce.cpp +124 -0
- data/mlx/mlx/backend/cpu/binary.cpp +269 -0
- data/mlx/mlx/backend/cpu/binary.h +517 -0
- data/mlx/mlx/backend/cpu/binary_ops.h +98 -0
- data/mlx/mlx/backend/cpu/binary_two.h +166 -0
- data/mlx/mlx/backend/cpu/cholesky.cpp +85 -0
- data/mlx/mlx/backend/cpu/compiled.cpp +357 -0
- data/mlx/mlx/backend/cpu/compiled_preamble.h +12 -0
- data/mlx/mlx/backend/cpu/conv.cpp +1351 -0
- data/mlx/mlx/backend/cpu/copy.cpp +386 -0
- data/mlx/mlx/backend/cpu/copy.h +36 -0
- data/mlx/mlx/backend/cpu/device_info.cpp +113 -0
- data/mlx/mlx/backend/cpu/device_info.h +28 -0
- data/mlx/mlx/backend/cpu/distributed.cpp +103 -0
- data/mlx/mlx/backend/cpu/eig.cpp +281 -0
- data/mlx/mlx/backend/cpu/eigh.cpp +241 -0
- data/mlx/mlx/backend/cpu/encoder.cpp +16 -0
- data/mlx/mlx/backend/cpu/encoder.h +67 -0
- data/mlx/mlx/backend/cpu/eval.cpp +40 -0
- data/mlx/mlx/backend/cpu/eval.h +12 -0
- data/mlx/mlx/backend/cpu/fft.cpp +120 -0
- data/mlx/mlx/backend/cpu/gemm.h +26 -0
- data/mlx/mlx/backend/cpu/gemms/bnns.cpp +214 -0
- data/mlx/mlx/backend/cpu/gemms/cblas.cpp +134 -0
- data/mlx/mlx/backend/cpu/gemms/simd_bf16.cpp +45 -0
- data/mlx/mlx/backend/cpu/gemms/simd_fp16.cpp +45 -0
- data/mlx/mlx/backend/cpu/gemms/simd_gemm.h +139 -0
- data/mlx/mlx/backend/cpu/hadamard.cpp +121 -0
- data/mlx/mlx/backend/cpu/indexing.cpp +854 -0
- data/mlx/mlx/backend/cpu/inverse.cpp +160 -0
- data/mlx/mlx/backend/cpu/jit_compiler.cpp +166 -0
- data/mlx/mlx/backend/cpu/jit_compiler.h +20 -0
- data/mlx/mlx/backend/cpu/lapack.h +80 -0
- data/mlx/mlx/backend/cpu/logsumexp.cpp +139 -0
- data/mlx/mlx/backend/cpu/luf.cpp +120 -0
- data/mlx/mlx/backend/cpu/make_compiled_preamble.ps1 +38 -0
- data/mlx/mlx/backend/cpu/make_compiled_preamble.sh +41 -0
- data/mlx/mlx/backend/cpu/masked_mm.cpp +608 -0
- data/mlx/mlx/backend/cpu/matmul.cpp +166 -0
- data/mlx/mlx/backend/cpu/primitives.cpp +478 -0
- data/mlx/mlx/backend/cpu/qrf.cpp +147 -0
- data/mlx/mlx/backend/cpu/quantized.cpp +1370 -0
- data/mlx/mlx/backend/cpu/reduce.cpp +587 -0
- data/mlx/mlx/backend/cpu/scan.cpp +338 -0
- data/mlx/mlx/backend/cpu/select.cpp +95 -0
- data/mlx/mlx/backend/cpu/simd/accelerate_fp16_simd.h +56 -0
- data/mlx/mlx/backend/cpu/simd/accelerate_simd.h +329 -0
- data/mlx/mlx/backend/cpu/simd/base_simd.h +319 -0
- data/mlx/mlx/backend/cpu/simd/math.h +193 -0
- data/mlx/mlx/backend/cpu/simd/neon_fp16_simd.h +212 -0
- data/mlx/mlx/backend/cpu/simd/simd.h +4 -0
- data/mlx/mlx/backend/cpu/simd/type.h +11 -0
- data/mlx/mlx/backend/cpu/slicing.h +21 -0
- data/mlx/mlx/backend/cpu/softmax.cpp +170 -0
- data/mlx/mlx/backend/cpu/sort.cpp +481 -0
- data/mlx/mlx/backend/cpu/svd.cpp +289 -0
- data/mlx/mlx/backend/cpu/ternary.h +154 -0
- data/mlx/mlx/backend/cpu/threefry.cpp +31 -0
- data/mlx/mlx/backend/cpu/threefry.h +21 -0
- data/mlx/mlx/backend/cpu/unary.cpp +238 -0
- data/mlx/mlx/backend/cpu/unary.h +281 -0
- data/mlx/mlx/backend/cpu/unary_ops.h +175 -0
- data/mlx/mlx/backend/cuda/CMakeLists.txt +265 -0
- data/mlx/mlx/backend/cuda/allocator.cpp +451 -0
- data/mlx/mlx/backend/cuda/allocator.h +94 -0
- data/mlx/mlx/backend/cuda/arange.cu +68 -0
- data/mlx/mlx/backend/cuda/arg_reduce.cu +189 -0
- data/mlx/mlx/backend/cuda/bin2h.cmake +150 -0
- data/mlx/mlx/backend/cuda/binary/CMakeLists.txt +21 -0
- data/mlx/mlx/backend/cuda/binary/add.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/arctan2.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/binary.cuh +383 -0
- data/mlx/mlx/backend/cuda/binary/bitwise_binary.cu +27 -0
- data/mlx/mlx/backend/cuda/binary/divide.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/equal.cu +15 -0
- data/mlx/mlx/backend/cuda/binary/greater.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/greater_equal.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/less.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/less_equal.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/log_add_exp.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/logical_and.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/logical_or.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/maximum.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/minimum.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/multiply.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/not_equal.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/power.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/remainder.cu +7 -0
- data/mlx/mlx/backend/cuda/binary/subtract.cu +7 -0
- data/mlx/mlx/backend/cuda/binary_two.cu +412 -0
- data/mlx/mlx/backend/cuda/compiled.cpp +357 -0
- data/mlx/mlx/backend/cuda/conv/conv.h +126 -0
- data/mlx/mlx/backend/cuda/conv/gemm_conv.cu +217 -0
- data/mlx/mlx/backend/cuda/conv/gemm_grouped_conv.cu +231 -0
- data/mlx/mlx/backend/cuda/conv.cpp +403 -0
- data/mlx/mlx/backend/cuda/copy/copy.cuh +55 -0
- data/mlx/mlx/backend/cuda/copy/copy_contiguous.cu +88 -0
- data/mlx/mlx/backend/cuda/copy/copy_general.cu +171 -0
- data/mlx/mlx/backend/cuda/copy/copy_general_dynamic.cu +118 -0
- data/mlx/mlx/backend/cuda/copy/copy_general_input.cu +229 -0
- data/mlx/mlx/backend/cuda/copy.cu +132 -0
- data/mlx/mlx/backend/cuda/cublas_utils.cpp +222 -0
- data/mlx/mlx/backend/cuda/cublas_utils.h +95 -0
- data/mlx/mlx/backend/cuda/cuda.h +21 -0
- data/mlx/mlx/backend/cuda/cuda_utils.h +90 -0
- data/mlx/mlx/backend/cuda/cudnn_utils.cpp +133 -0
- data/mlx/mlx/backend/cuda/cudnn_utils.h +187 -0
- data/mlx/mlx/backend/cuda/custom_kernel.cpp +379 -0
- data/mlx/mlx/backend/cuda/cutlass_utils.cuh +46 -0
- data/mlx/mlx/backend/cuda/delayload.cpp +80 -0
- data/mlx/mlx/backend/cuda/device/atomic_ops.cuh +63 -0
- data/mlx/mlx/backend/cuda/device/binary_ops.cuh +300 -0
- data/mlx/mlx/backend/cuda/device/cast_op.cuh +118 -0
- data/mlx/mlx/backend/cuda/device/complex.cuh +60 -0
- data/mlx/mlx/backend/cuda/device/config.h +12 -0
- data/mlx/mlx/backend/cuda/device/fp16_math.cuh +96 -0
- data/mlx/mlx/backend/cuda/device/gather.cuh +53 -0
- data/mlx/mlx/backend/cuda/device/gather_axis.cuh +65 -0
- data/mlx/mlx/backend/cuda/device/indexing.cuh +30 -0
- data/mlx/mlx/backend/cuda/device/scatter.cuh +68 -0
- data/mlx/mlx/backend/cuda/device/scatter_axis.cuh +67 -0
- data/mlx/mlx/backend/cuda/device/scatter_ops.cuh +44 -0
- data/mlx/mlx/backend/cuda/device/ternary_ops.cuh +13 -0
- data/mlx/mlx/backend/cuda/device/unary_ops.cuh +350 -0
- data/mlx/mlx/backend/cuda/device/utils.cuh +464 -0
- data/mlx/mlx/backend/cuda/device.cpp +522 -0
- data/mlx/mlx/backend/cuda/device.h +195 -0
- data/mlx/mlx/backend/cuda/device_info.cpp +232 -0
- data/mlx/mlx/backend/cuda/distributed.cu +121 -0
- data/mlx/mlx/backend/cuda/eval.cpp +66 -0
- data/mlx/mlx/backend/cuda/event.cu +415 -0
- data/mlx/mlx/backend/cuda/event.h +79 -0
- data/mlx/mlx/backend/cuda/fence.cpp +42 -0
- data/mlx/mlx/backend/cuda/gemms/cublas_gemm.cpp +233 -0
- data/mlx/mlx/backend/cuda/gemms/cublas_gemm.h +114 -0
- data/mlx/mlx/backend/cuda/gemms/cublas_gemm_batched_12_0.cpp +77 -0
- data/mlx/mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu +329 -0
- data/mlx/mlx/backend/cuda/gemms/gemv.cu +327 -0
- data/mlx/mlx/backend/cuda/gemms/gemv.h +34 -0
- data/mlx/mlx/backend/cuda/gemms/grouped_gemm.h +25 -0
- data/mlx/mlx/backend/cuda/gemms/grouped_gemm_unaligned.cu +358 -0
- data/mlx/mlx/backend/cuda/indexing.cpp +434 -0
- data/mlx/mlx/backend/cuda/jit_module.cpp +443 -0
- data/mlx/mlx/backend/cuda/jit_module.h +120 -0
- data/mlx/mlx/backend/cuda/kernel_utils.cu +52 -0
- data/mlx/mlx/backend/cuda/kernel_utils.cuh +148 -0
- data/mlx/mlx/backend/cuda/layer_norm.cu +417 -0
- data/mlx/mlx/backend/cuda/load.cpp +60 -0
- data/mlx/mlx/backend/cuda/logsumexp.cu +161 -0
- data/mlx/mlx/backend/cuda/lru_cache.h +190 -0
- data/mlx/mlx/backend/cuda/matmul.cpp +373 -0
- data/mlx/mlx/backend/cuda/no_cuda.cpp +47 -0
- data/mlx/mlx/backend/cuda/primitives.cpp +46 -0
- data/mlx/mlx/backend/cuda/quantized/affine_quantize.cu +329 -0
- data/mlx/mlx/backend/cuda/quantized/convert_fp8.cu +19 -0
- data/mlx/mlx/backend/cuda/quantized/cublas_qqmm.cpp +206 -0
- data/mlx/mlx/backend/cuda/quantized/cublas_qqmm.h +88 -0
- data/mlx/mlx/backend/cuda/quantized/cuda_fp4.h +100 -0
- data/mlx/mlx/backend/cuda/quantized/fp_quantize.cu +496 -0
- data/mlx/mlx/backend/cuda/quantized/mxfp8_quantize.cuh +32 -0
- data/mlx/mlx/backend/cuda/quantized/no_qqmm_impl.cpp +26 -0
- data/mlx/mlx/backend/cuda/quantized/nvfp4_quantize.cuh +334 -0
- data/mlx/mlx/backend/cuda/quantized/qmv.cu +304 -0
- data/mlx/mlx/backend/cuda/quantized/qmv.h +21 -0
- data/mlx/mlx/backend/cuda/quantized/qqmm.cpp +158 -0
- data/mlx/mlx/backend/cuda/quantized/qqmm_impl.cpp +50 -0
- data/mlx/mlx/backend/cuda/quantized/qqmm_impl.h +26 -0
- data/mlx/mlx/backend/cuda/quantized/qqmm_utils.cu +227 -0
- data/mlx/mlx/backend/cuda/quantized/qqmm_utils.h +30 -0
- data/mlx/mlx/backend/cuda/quantized/quantized.cpp +85 -0
- data/mlx/mlx/backend/cuda/quantized/quantized.h +53 -0
- data/mlx/mlx/backend/cuda/quantized/quantized_utils.cuh +88 -0
- data/mlx/mlx/backend/cuda/quantized/quantized_utils.h +50 -0
- data/mlx/mlx/backend/cuda/random.cu +202 -0
- data/mlx/mlx/backend/cuda/reduce/all_reduce.cu +159 -0
- data/mlx/mlx/backend/cuda/reduce/col_reduce.cu +510 -0
- data/mlx/mlx/backend/cuda/reduce/init_reduce.cu +50 -0
- data/mlx/mlx/backend/cuda/reduce/reduce.cuh +71 -0
- data/mlx/mlx/backend/cuda/reduce/reduce_ops.cuh +211 -0
- data/mlx/mlx/backend/cuda/reduce/reduce_utils.cuh +145 -0
- data/mlx/mlx/backend/cuda/reduce/row_reduce.cu +361 -0
- data/mlx/mlx/backend/cuda/reduce.cu +73 -0
- data/mlx/mlx/backend/cuda/rms_norm.cu +536 -0
- data/mlx/mlx/backend/cuda/rope.cu +429 -0
- data/mlx/mlx/backend/cuda/scaled_dot_product_attention.cpp +681 -0
- data/mlx/mlx/backend/cuda/scaled_dot_product_attention.cu +796 -0
- data/mlx/mlx/backend/cuda/scan.cu +468 -0
- data/mlx/mlx/backend/cuda/slicing.cpp +111 -0
- data/mlx/mlx/backend/cuda/softmax.cu +162 -0
- data/mlx/mlx/backend/cuda/sort.cu +1076 -0
- data/mlx/mlx/backend/cuda/steel/defines.cuh +9 -0
- data/mlx/mlx/backend/cuda/steel/gemm.cuh +101 -0
- data/mlx/mlx/backend/cuda/steel/mma.cuh +117 -0
- data/mlx/mlx/backend/cuda/steel/tiles.cuh +450 -0
- data/mlx/mlx/backend/cuda/steel/utils.cuh +89 -0
- data/mlx/mlx/backend/cuda/ternary.cu +271 -0
- data/mlx/mlx/backend/cuda/unary/CMakeLists.txt +34 -0
- data/mlx/mlx/backend/cuda/unary/abs.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/arccos.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/arccosh.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/arcsin.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/arcsinh.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/arctan.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/arctanh.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/bitwise_invert.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/ceil.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/conjugate.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/cos.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/cosh.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/erf.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/erf_inv.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/exp.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/expm1.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/floor.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/imag.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/log.cu +21 -0
- data/mlx/mlx/backend/cuda/unary/log1p.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/logical_not.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/negative.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/real.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/round.cu +18 -0
- data/mlx/mlx/backend/cuda/unary/sigmoid.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/sign.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/sin.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/sinh.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/sqrt.cu +15 -0
- data/mlx/mlx/backend/cuda/unary/square.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/tan.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/tanh.cu +7 -0
- data/mlx/mlx/backend/cuda/unary/unary.cuh +224 -0
- data/mlx/mlx/backend/cuda/utils.cpp +116 -0
- data/mlx/mlx/backend/cuda/utils.h +49 -0
- data/mlx/mlx/backend/cuda/vector_types.cuh +48 -0
- data/mlx/mlx/backend/cuda/worker.cpp +79 -0
- data/mlx/mlx/backend/cuda/worker.h +55 -0
- data/mlx/mlx/backend/gpu/CMakeLists.txt +5 -0
- data/mlx/mlx/backend/gpu/copy.cpp +89 -0
- data/mlx/mlx/backend/gpu/copy.h +57 -0
- data/mlx/mlx/backend/gpu/device_info.h +36 -0
- data/mlx/mlx/backend/gpu/eval.h +18 -0
- data/mlx/mlx/backend/gpu/primitives.cpp +307 -0
- data/mlx/mlx/backend/gpu/slicing.cpp +44 -0
- data/mlx/mlx/backend/gpu/slicing.h +36 -0
- data/mlx/mlx/backend/metal/CMakeLists.txt +144 -0
- data/mlx/mlx/backend/metal/allocator.cpp +279 -0
- data/mlx/mlx/backend/metal/allocator.h +79 -0
- data/mlx/mlx/backend/metal/binary.cpp +257 -0
- data/mlx/mlx/backend/metal/binary.h +33 -0
- data/mlx/mlx/backend/metal/compiled.cpp +471 -0
- data/mlx/mlx/backend/metal/conv.cpp +1118 -0
- data/mlx/mlx/backend/metal/copy.cpp +235 -0
- data/mlx/mlx/backend/metal/custom_kernel.cpp +430 -0
- data/mlx/mlx/backend/metal/device.cpp +816 -0
- data/mlx/mlx/backend/metal/device.h +289 -0
- data/mlx/mlx/backend/metal/device_info.cpp +58 -0
- data/mlx/mlx/backend/metal/distributed.cpp +38 -0
- data/mlx/mlx/backend/metal/eval.cpp +97 -0
- data/mlx/mlx/backend/metal/event.cpp +62 -0
- data/mlx/mlx/backend/metal/fence.cpp +162 -0
- data/mlx/mlx/backend/metal/fft.cpp +807 -0
- data/mlx/mlx/backend/metal/hadamard.cpp +198 -0
- data/mlx/mlx/backend/metal/indexing.cpp +727 -0
- data/mlx/mlx/backend/metal/jit/includes.h +58 -0
- data/mlx/mlx/backend/metal/jit/indexing.h +76 -0
- data/mlx/mlx/backend/metal/jit_kernels.cpp +1118 -0
- data/mlx/mlx/backend/metal/kernels/CMakeLists.txt +193 -0
- data/mlx/mlx/backend/metal/kernels/arange.h +9 -0
- data/mlx/mlx/backend/metal/kernels/arange.metal +20 -0
- data/mlx/mlx/backend/metal/kernels/arg_reduce.metal +182 -0
- data/mlx/mlx/backend/metal/kernels/atomic.h +345 -0
- data/mlx/mlx/backend/metal/kernels/bf16.h +16 -0
- data/mlx/mlx/backend/metal/kernels/bf16_math.h +380 -0
- data/mlx/mlx/backend/metal/kernels/binary.h +199 -0
- data/mlx/mlx/backend/metal/kernels/binary.metal +109 -0
- data/mlx/mlx/backend/metal/kernels/binary_ops.h +330 -0
- data/mlx/mlx/backend/metal/kernels/binary_two.h +244 -0
- data/mlx/mlx/backend/metal/kernels/binary_two.metal +54 -0
- data/mlx/mlx/backend/metal/kernels/cexpf.h +134 -0
- data/mlx/mlx/backend/metal/kernels/complex.h +173 -0
- data/mlx/mlx/backend/metal/kernels/conv.metal +701 -0
- data/mlx/mlx/backend/metal/kernels/copy.h +276 -0
- data/mlx/mlx/backend/metal/kernels/copy.metal +75 -0
- data/mlx/mlx/backend/metal/kernels/defines.h +24 -0
- data/mlx/mlx/backend/metal/kernels/erf.h +69 -0
- data/mlx/mlx/backend/metal/kernels/expm1f.h +90 -0
- data/mlx/mlx/backend/metal/kernels/fence.metal +52 -0
- data/mlx/mlx/backend/metal/kernels/fft/radix.h +328 -0
- data/mlx/mlx/backend/metal/kernels/fft/readwrite.h +624 -0
- data/mlx/mlx/backend/metal/kernels/fft.h +486 -0
- data/mlx/mlx/backend/metal/kernels/fft.metal +67 -0
- data/mlx/mlx/backend/metal/kernels/fp4.h +48 -0
- data/mlx/mlx/backend/metal/kernels/fp8.h +80 -0
- data/mlx/mlx/backend/metal/kernels/fp_quantized.h +1850 -0
- data/mlx/mlx/backend/metal/kernels/fp_quantized.metal +153 -0
- data/mlx/mlx/backend/metal/kernels/fp_quantized_nax.h +1044 -0
- data/mlx/mlx/backend/metal/kernels/fp_quantized_nax.metal +79 -0
- data/mlx/mlx/backend/metal/kernels/gemv.metal +868 -0
- data/mlx/mlx/backend/metal/kernels/gemv_masked.h +827 -0
- data/mlx/mlx/backend/metal/kernels/gemv_masked.metal +76 -0
- data/mlx/mlx/backend/metal/kernels/hadamard.h +182 -0
- data/mlx/mlx/backend/metal/kernels/indexing/gather.h +51 -0
- data/mlx/mlx/backend/metal/kernels/indexing/gather_axis.h +44 -0
- data/mlx/mlx/backend/metal/kernels/indexing/gather_front.h +24 -0
- data/mlx/mlx/backend/metal/kernels/indexing/indexing.h +23 -0
- data/mlx/mlx/backend/metal/kernels/indexing/masked_scatter.h +41 -0
- data/mlx/mlx/backend/metal/kernels/indexing/scatter.h +59 -0
- data/mlx/mlx/backend/metal/kernels/indexing/scatter_axis.h +52 -0
- data/mlx/mlx/backend/metal/kernels/layer_norm.metal +433 -0
- data/mlx/mlx/backend/metal/kernels/logging.h +26 -0
- data/mlx/mlx/backend/metal/kernels/logsumexp.h +140 -0
- data/mlx/mlx/backend/metal/kernels/logsumexp.metal +18 -0
- data/mlx/mlx/backend/metal/kernels/quantized.h +2508 -0
- data/mlx/mlx/backend/metal/kernels/quantized.metal +144 -0
- data/mlx/mlx/backend/metal/kernels/quantized_nax.h +1705 -0
- data/mlx/mlx/backend/metal/kernels/quantized_nax.metal +106 -0
- data/mlx/mlx/backend/metal/kernels/quantized_utils.h +90 -0
- data/mlx/mlx/backend/metal/kernels/random.metal +103 -0
- data/mlx/mlx/backend/metal/kernels/reduce.h +5 -0
- data/mlx/mlx/backend/metal/kernels/reduce.metal +169 -0
- data/mlx/mlx/backend/metal/kernels/reduce_utils.h +6 -0
- data/mlx/mlx/backend/metal/kernels/reduction/ops.h +275 -0
- data/mlx/mlx/backend/metal/kernels/reduction/reduce_all.h +66 -0
- data/mlx/mlx/backend/metal/kernels/reduction/reduce_col.h +398 -0
- data/mlx/mlx/backend/metal/kernels/reduction/reduce_init.h +8 -0
- data/mlx/mlx/backend/metal/kernels/reduction/reduce_row.h +369 -0
- data/mlx/mlx/backend/metal/kernels/rms_norm.metal +391 -0
- data/mlx/mlx/backend/metal/kernels/rope.metal +229 -0
- data/mlx/mlx/backend/metal/kernels/scaled_dot_product_attention.metal +44 -0
- data/mlx/mlx/backend/metal/kernels/scan.h +514 -0
- data/mlx/mlx/backend/metal/kernels/scan.metal +109 -0
- data/mlx/mlx/backend/metal/kernels/sdpa_vector.h +394 -0
- data/mlx/mlx/backend/metal/kernels/softmax.h +190 -0
- data/mlx/mlx/backend/metal/kernels/softmax.metal +24 -0
- data/mlx/mlx/backend/metal/kernels/sort.h +719 -0
- data/mlx/mlx/backend/metal/kernels/sort.metal +80 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/attn.h +296 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.h +471 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.metal +27 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.h +481 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.metal +28 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/loader.h +264 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/mma.h +750 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/nax.h +1076 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/params.h +44 -0
- data/mlx/mlx/backend/metal/kernels/steel/attn/transforms.h +71 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/conv.h +13 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv.h +176 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv.metal +56 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv_general.h +225 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/kernels/steel_conv_general.metal +47 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/loader.h +6 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_l.h +451 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_n.h +319 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/loaders/loader_general.h +381 -0
- data/mlx/mlx/backend/metal/kernels/steel/conv/params.h +62 -0
- data/mlx/mlx/backend/metal/kernels/steel/defines.h +7 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/gemm.h +295 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/gemm_nax.h +157 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused.h +346 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused.metal +34 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.h +219 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.metal +30 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather.h +459 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather.metal +59 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.h +143 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_gather_nax.metal +37 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_masked.h +719 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_masked.metal +76 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_segmented.h +266 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_segmented.metal +43 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk.h +227 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk.metal +76 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk_nax.h +152 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_splitk_nax.metal +30 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/loader.h +137 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/mma.h +1146 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/nax.h +1084 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/params.h +65 -0
- data/mlx/mlx/backend/metal/kernels/steel/gemm/transforms.h +72 -0
- data/mlx/mlx/backend/metal/kernels/steel/utils/integral_constant.h +134 -0
- data/mlx/mlx/backend/metal/kernels/steel/utils/type_traits.h +55 -0
- data/mlx/mlx/backend/metal/kernels/steel/utils.h +42 -0
- data/mlx/mlx/backend/metal/kernels/ternary.h +145 -0
- data/mlx/mlx/backend/metal/kernels/ternary.metal +48 -0
- data/mlx/mlx/backend/metal/kernels/ternary_ops.h +10 -0
- data/mlx/mlx/backend/metal/kernels/unary.h +63 -0
- data/mlx/mlx/backend/metal/kernels/unary.metal +115 -0
- data/mlx/mlx/backend/metal/kernels/unary_ops.h +454 -0
- data/mlx/mlx/backend/metal/kernels/utils.h +445 -0
- data/mlx/mlx/backend/metal/kernels.h +375 -0
- data/mlx/mlx/backend/metal/logsumexp.cpp +95 -0
- data/mlx/mlx/backend/metal/make_compiled_preamble.sh +120 -0
- data/mlx/mlx/backend/metal/matmul.cpp +2572 -0
- data/mlx/mlx/backend/metal/matmul.h +144 -0
- data/mlx/mlx/backend/metal/metal.cpp +50 -0
- data/mlx/mlx/backend/metal/metal.h +25 -0
- data/mlx/mlx/backend/metal/no_metal.cpp +42 -0
- data/mlx/mlx/backend/metal/nojit_kernels.cpp +414 -0
- data/mlx/mlx/backend/metal/normalization.cpp +433 -0
- data/mlx/mlx/backend/metal/primitives.cpp +242 -0
- data/mlx/mlx/backend/metal/quantized.cpp +1651 -0
- data/mlx/mlx/backend/metal/reduce.cpp +1038 -0
- data/mlx/mlx/backend/metal/reduce.h +41 -0
- data/mlx/mlx/backend/metal/resident.cpp +100 -0
- data/mlx/mlx/backend/metal/resident.h +32 -0
- data/mlx/mlx/backend/metal/rope.cpp +165 -0
- data/mlx/mlx/backend/metal/scaled_dot_product_attention.cpp +798 -0
- data/mlx/mlx/backend/metal/scan.cpp +145 -0
- data/mlx/mlx/backend/metal/scan.h +17 -0
- data/mlx/mlx/backend/metal/slicing.cpp +99 -0
- data/mlx/mlx/backend/metal/softmax.cpp +87 -0
- data/mlx/mlx/backend/metal/sort.cpp +368 -0
- data/mlx/mlx/backend/metal/ternary.cpp +160 -0
- data/mlx/mlx/backend/metal/ternary.h +21 -0
- data/mlx/mlx/backend/metal/unary.cpp +161 -0
- data/mlx/mlx/backend/metal/unary.h +21 -0
- data/mlx/mlx/backend/metal/utils.cpp +77 -0
- data/mlx/mlx/backend/metal/utils.h +99 -0
- data/mlx/mlx/backend/no_cpu/CMakeLists.txt +7 -0
- data/mlx/mlx/backend/no_cpu/compiled.cpp +24 -0
- data/mlx/mlx/backend/no_cpu/device_info.cpp +22 -0
- data/mlx/mlx/backend/no_cpu/primitives.cpp +146 -0
- data/mlx/mlx/backend/no_gpu/CMakeLists.txt +8 -0
- data/mlx/mlx/backend/no_gpu/allocator.cpp +134 -0
- data/mlx/mlx/backend/no_gpu/apple_memory.h +16 -0
- data/mlx/mlx/backend/no_gpu/device_info.cpp +22 -0
- data/mlx/mlx/backend/no_gpu/eval.cpp +24 -0
- data/mlx/mlx/backend/no_gpu/event.cpp +53 -0
- data/mlx/mlx/backend/no_gpu/fence.cpp +54 -0
- data/mlx/mlx/backend/no_gpu/linux_memory.h +22 -0
- data/mlx/mlx/backend/no_gpu/primitives.cpp +185 -0
- data/mlx/mlx/compile.cpp +1243 -0
- data/mlx/mlx/compile.h +45 -0
- data/mlx/mlx/compile_impl.h +70 -0
- data/mlx/mlx/device.cpp +72 -0
- data/mlx/mlx/device.h +56 -0
- data/mlx/mlx/distributed/CMakeLists.txt +14 -0
- data/mlx/mlx/distributed/distributed.cpp +197 -0
- data/mlx/mlx/distributed/distributed.h +61 -0
- data/mlx/mlx/distributed/distributed_impl.h +59 -0
- data/mlx/mlx/distributed/jaccl/CMakeLists.txt +12 -0
- data/mlx/mlx/distributed/jaccl/jaccl.cpp +178 -0
- data/mlx/mlx/distributed/jaccl/jaccl.h +12 -0
- data/mlx/mlx/distributed/jaccl/mesh.cpp +451 -0
- data/mlx/mlx/distributed/jaccl/mesh.h +122 -0
- data/mlx/mlx/distributed/jaccl/no_jaccl.cpp +20 -0
- data/mlx/mlx/distributed/jaccl/ring.cpp +692 -0
- data/mlx/mlx/distributed/jaccl/ring.h +178 -0
- data/mlx/mlx/distributed/jaccl/utils.cpp +329 -0
- data/mlx/mlx/distributed/jaccl/utils.h +342 -0
- data/mlx/mlx/distributed/mpi/CMakeLists.txt +5 -0
- data/mlx/mlx/distributed/mpi/mpi.cpp +501 -0
- data/mlx/mlx/distributed/mpi/mpi.h +12 -0
- data/mlx/mlx/distributed/mpi/mpi_declarations.h +28 -0
- data/mlx/mlx/distributed/mpi/no_mpi.cpp +20 -0
- data/mlx/mlx/distributed/nccl/CMakeLists.txt +26 -0
- data/mlx/mlx/distributed/nccl/nccl.cpp +443 -0
- data/mlx/mlx/distributed/nccl/nccl.h +12 -0
- data/mlx/mlx/distributed/nccl/nccl_stub/CMakeLists.txt +1 -0
- data/mlx/mlx/distributed/nccl/nccl_stub/nccl_stubs.cpp +54 -0
- data/mlx/mlx/distributed/nccl/no_nccl.cpp +20 -0
- data/mlx/mlx/distributed/ops.cpp +186 -0
- data/mlx/mlx/distributed/ops.h +57 -0
- data/mlx/mlx/distributed/primitives.cpp +95 -0
- data/mlx/mlx/distributed/primitives.h +156 -0
- data/mlx/mlx/distributed/reduction_ops.h +38 -0
- data/mlx/mlx/distributed/ring/CMakeLists.txt +5 -0
- data/mlx/mlx/distributed/ring/no_ring.cpp +20 -0
- data/mlx/mlx/distributed/ring/ring.cpp +870 -0
- data/mlx/mlx/distributed/ring/ring.h +12 -0
- data/mlx/mlx/distributed/utils.cpp +206 -0
- data/mlx/mlx/distributed/utils.h +67 -0
- data/mlx/mlx/dtype.cpp +197 -0
- data/mlx/mlx/dtype.h +116 -0
- data/mlx/mlx/dtype_utils.cpp +42 -0
- data/mlx/mlx/dtype_utils.h +119 -0
- data/mlx/mlx/einsum.cpp +941 -0
- data/mlx/mlx/einsum.h +23 -0
- data/mlx/mlx/event.h +58 -0
- data/mlx/mlx/export.cpp +1130 -0
- data/mlx/mlx/export.h +137 -0
- data/mlx/mlx/export_impl.h +99 -0
- data/mlx/mlx/fast.cpp +941 -0
- data/mlx/mlx/fast.h +103 -0
- data/mlx/mlx/fast_primitives.h +427 -0
- data/mlx/mlx/fence.h +39 -0
- data/mlx/mlx/fft.cpp +262 -0
- data/mlx/mlx/fft.h +159 -0
- data/mlx/mlx/graph_utils.cpp +175 -0
- data/mlx/mlx/graph_utils.h +67 -0
- data/mlx/mlx/io/CMakeLists.txt +25 -0
- data/mlx/mlx/io/gguf.cpp +470 -0
- data/mlx/mlx/io/gguf.h +20 -0
- data/mlx/mlx/io/gguf_quants.cpp +164 -0
- data/mlx/mlx/io/load.cpp +397 -0
- data/mlx/mlx/io/load.h +175 -0
- data/mlx/mlx/io/no_gguf.cpp +20 -0
- data/mlx/mlx/io/no_safetensors.cpp +37 -0
- data/mlx/mlx/io/safetensors.cpp +234 -0
- data/mlx/mlx/io.h +61 -0
- data/mlx/mlx/linalg.cpp +708 -0
- data/mlx/mlx/linalg.h +115 -0
- data/mlx/mlx/memory.h +80 -0
- data/mlx/mlx/mlx.h +25 -0
- data/mlx/mlx/ops.cpp +6094 -0
- data/mlx/mlx/ops.h +1610 -0
- data/mlx/mlx/primitives.cpp +5850 -0
- data/mlx/mlx/primitives.h +2525 -0
- data/mlx/mlx/random.cpp +492 -0
- data/mlx/mlx/random.h +283 -0
- data/mlx/mlx/scheduler.cpp +73 -0
- data/mlx/mlx/scheduler.h +189 -0
- data/mlx/mlx/small_vector.h +540 -0
- data/mlx/mlx/stream.h +42 -0
- data/mlx/mlx/threadpool.h +133 -0
- data/mlx/mlx/transforms.cpp +1065 -0
- data/mlx/mlx/transforms.h +231 -0
- data/mlx/mlx/transforms_impl.h +88 -0
- data/mlx/mlx/types/bf16.h +187 -0
- data/mlx/mlx/types/complex.h +113 -0
- data/mlx/mlx/types/fp16.h +234 -0
- data/mlx/mlx/types/half_types.h +58 -0
- data/mlx/mlx/types/limits.h +70 -0
- data/mlx/mlx/utils.cpp +302 -0
- data/mlx/mlx/utils.h +174 -0
- data/mlx/mlx/version.cpp +11 -0
- data/mlx/mlx/version.h +22 -0
- data/mlx/mlx.pc.in +52 -0
- data/mlx/pyproject.toml +7 -0
- data/mlx/python/mlx/__main__.py +27 -0
- data/mlx/python/mlx/_distributed_utils/common.py +135 -0
- data/mlx/python/mlx/_distributed_utils/config.py +631 -0
- data/mlx/python/mlx/_distributed_utils/launch.py +570 -0
- data/mlx/python/mlx/_reprlib_fix.py +16 -0
- data/mlx/python/mlx/_stub_patterns.txt +36 -0
- data/mlx/python/mlx/extension.py +88 -0
- data/mlx/python/mlx/nn/__init__.py +5 -0
- data/mlx/python/mlx/nn/init.py +441 -0
- data/mlx/python/mlx/nn/layers/__init__.py +105 -0
- data/mlx/python/mlx/nn/layers/activations.py +661 -0
- data/mlx/python/mlx/nn/layers/base.py +675 -0
- data/mlx/python/mlx/nn/layers/containers.py +24 -0
- data/mlx/python/mlx/nn/layers/convolution.py +232 -0
- data/mlx/python/mlx/nn/layers/convolution_transpose.py +242 -0
- data/mlx/python/mlx/nn/layers/distributed.py +601 -0
- data/mlx/python/mlx/nn/layers/dropout.py +137 -0
- data/mlx/python/mlx/nn/layers/embedding.py +53 -0
- data/mlx/python/mlx/nn/layers/linear.py +180 -0
- data/mlx/python/mlx/nn/layers/normalization.py +363 -0
- data/mlx/python/mlx/nn/layers/pooling.py +398 -0
- data/mlx/python/mlx/nn/layers/positional_encoding.py +162 -0
- data/mlx/python/mlx/nn/layers/quantized.py +426 -0
- data/mlx/python/mlx/nn/layers/recurrent.py +289 -0
- data/mlx/python/mlx/nn/layers/transformer.py +354 -0
- data/mlx/python/mlx/nn/layers/upsample.py +277 -0
- data/mlx/python/mlx/nn/losses.py +610 -0
- data/mlx/python/mlx/nn/utils.py +165 -0
- data/mlx/python/mlx/optimizers/__init__.py +4 -0
- data/mlx/python/mlx/optimizers/optimizers.py +976 -0
- data/mlx/python/mlx/optimizers/schedulers.py +158 -0
- data/mlx/python/mlx/py.typed +1 -0
- data/mlx/python/mlx/utils.py +325 -0
- data/mlx/python/src/CMakeLists.txt +96 -0
- data/mlx/python/src/array.cpp +1525 -0
- data/mlx/python/src/buffer.h +124 -0
- data/mlx/python/src/constants.cpp +15 -0
- data/mlx/python/src/convert.cpp +504 -0
- data/mlx/python/src/convert.h +50 -0
- data/mlx/python/src/cuda.cpp +19 -0
- data/mlx/python/src/device.cpp +98 -0
- data/mlx/python/src/distributed.cpp +352 -0
- data/mlx/python/src/export.cpp +356 -0
- data/mlx/python/src/fast.cpp +627 -0
- data/mlx/python/src/fft.cpp +514 -0
- data/mlx/python/src/indexing.cpp +1016 -0
- data/mlx/python/src/indexing.h +41 -0
- data/mlx/python/src/linalg.cpp +663 -0
- data/mlx/python/src/load.cpp +531 -0
- data/mlx/python/src/load.h +51 -0
- data/mlx/python/src/memory.cpp +125 -0
- data/mlx/python/src/metal.cpp +98 -0
- data/mlx/python/src/mlx.cpp +51 -0
- data/mlx/python/src/mlx_func.cpp +116 -0
- data/mlx/python/src/mlx_func.h +31 -0
- data/mlx/python/src/ops.cpp +5545 -0
- data/mlx/python/src/random.cpp +516 -0
- data/mlx/python/src/small_vector.h +76 -0
- data/mlx/python/src/stream.cpp +147 -0
- data/mlx/python/src/transforms.cpp +1542 -0
- data/mlx/python/src/trees.cpp +311 -0
- data/mlx/python/src/trees.h +62 -0
- data/mlx/python/src/utils.cpp +98 -0
- data/mlx/python/src/utils.h +78 -0
- data/mlx/python/tests/__main__.py +5 -0
- data/mlx/python/tests/cuda_skip.py +62 -0
- data/mlx/python/tests/mlx_distributed_tests.py +314 -0
- data/mlx/python/tests/mlx_tests.py +116 -0
- data/mlx/python/tests/mpi_test_distributed.py +142 -0
- data/mlx/python/tests/nccl_test_distributed.py +52 -0
- data/mlx/python/tests/ring_test_distributed.py +131 -0
- data/mlx/python/tests/test_array.py +2139 -0
- data/mlx/python/tests/test_autograd.py +880 -0
- data/mlx/python/tests/test_bf16.py +196 -0
- data/mlx/python/tests/test_blas.py +1429 -0
- data/mlx/python/tests/test_compile.py +1277 -0
- data/mlx/python/tests/test_constants.py +41 -0
- data/mlx/python/tests/test_conv.py +1198 -0
- data/mlx/python/tests/test_conv_transpose.py +810 -0
- data/mlx/python/tests/test_device.py +150 -0
- data/mlx/python/tests/test_double.py +306 -0
- data/mlx/python/tests/test_einsum.py +363 -0
- data/mlx/python/tests/test_eval.py +200 -0
- data/mlx/python/tests/test_export_import.py +614 -0
- data/mlx/python/tests/test_fast.py +923 -0
- data/mlx/python/tests/test_fast_sdpa.py +647 -0
- data/mlx/python/tests/test_fft.py +323 -0
- data/mlx/python/tests/test_graph.py +37 -0
- data/mlx/python/tests/test_init.py +139 -0
- data/mlx/python/tests/test_linalg.py +621 -0
- data/mlx/python/tests/test_load.py +447 -0
- data/mlx/python/tests/test_losses.py +427 -0
- data/mlx/python/tests/test_memory.py +77 -0
- data/mlx/python/tests/test_nn.py +1986 -0
- data/mlx/python/tests/test_ops.py +3261 -0
- data/mlx/python/tests/test_optimizers.py +584 -0
- data/mlx/python/tests/test_quantized.py +1160 -0
- data/mlx/python/tests/test_random.py +392 -0
- data/mlx/python/tests/test_reduce.py +223 -0
- data/mlx/python/tests/test_tree.py +96 -0
- data/mlx/python/tests/test_upsample.py +100 -0
- data/mlx/python/tests/test_vmap.py +860 -0
- data/mlx/setup.py +315 -0
- data/mlx/tests/CMakeLists.txt +44 -0
- data/mlx/tests/allocator_tests.cpp +41 -0
- data/mlx/tests/arg_reduce_tests.cpp +204 -0
- data/mlx/tests/array_tests.cpp +663 -0
- data/mlx/tests/autograd_tests.cpp +1399 -0
- data/mlx/tests/blas_tests.cpp +110 -0
- data/mlx/tests/compile_tests.cpp +818 -0
- data/mlx/tests/creations_tests.cpp +239 -0
- data/mlx/tests/custom_vjp_tests.cpp +55 -0
- data/mlx/tests/device_tests.cpp +35 -0
- data/mlx/tests/einsum_tests.cpp +85 -0
- data/mlx/tests/eval_tests.cpp +93 -0
- data/mlx/tests/export_import_tests.cpp +164 -0
- data/mlx/tests/fft_tests.cpp +366 -0
- data/mlx/tests/gpu_tests.cpp +523 -0
- data/mlx/tests/linalg_tests.cpp +639 -0
- data/mlx/tests/load_tests.cpp +270 -0
- data/mlx/tests/ops_tests.cpp +4159 -0
- data/mlx/tests/random_tests.cpp +716 -0
- data/mlx/tests/scheduler_tests.cpp +121 -0
- data/mlx/tests/tests.cpp +26 -0
- data/mlx/tests/utils_tests.cpp +67 -0
- data/mlx/tests/vmap_tests.cpp +547 -0
- metadata +958 -0
|
@@ -0,0 +1,631 @@
|
|
|
1
|
+
# Copyright © 2025 Apple Inc.
|
|
2
|
+
|
|
3
|
+
import argparse
|
|
4
|
+
import json
|
|
5
|
+
import shlex
|
|
6
|
+
import sys
|
|
7
|
+
import threading
|
|
8
|
+
from collections import defaultdict
|
|
9
|
+
from dataclasses import dataclass
|
|
10
|
+
from subprocess import DEVNULL, run
|
|
11
|
+
from typing import Optional
|
|
12
|
+
|
|
13
|
+
import mlx.core as mx
|
|
14
|
+
|
|
15
|
+
from .common import (
|
|
16
|
+
Host,
|
|
17
|
+
Hostfile,
|
|
18
|
+
OptionalBoolAction,
|
|
19
|
+
log,
|
|
20
|
+
log_error,
|
|
21
|
+
log_warning,
|
|
22
|
+
)
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+
@dataclass
|
|
26
|
+
class SSHInfo:
|
|
27
|
+
can_ssh: bool
|
|
28
|
+
has_sudo: bool
|
|
29
|
+
|
|
30
|
+
def __bool__(self):
|
|
31
|
+
return self.can_ssh
|
|
32
|
+
|
|
33
|
+
|
|
34
|
+
@dataclass
|
|
35
|
+
class ThunderboltPort:
|
|
36
|
+
iface: str
|
|
37
|
+
uuid: str
|
|
38
|
+
connected_to: Optional[str]
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
@dataclass
|
|
42
|
+
class ThunderboltHost:
|
|
43
|
+
name: str
|
|
44
|
+
ports: list[ThunderboltPort]
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
def add_ips(hosts, verbose=False):
|
|
48
|
+
# Get the ips for each host
|
|
49
|
+
for h in hosts:
|
|
50
|
+
log(verbose, "Getting the ip from", h.ssh_hostname)
|
|
51
|
+
ip = run(
|
|
52
|
+
["ssh", h.ssh_hostname, "ipconfig", "getifaddr", "en0"],
|
|
53
|
+
capture_output=True,
|
|
54
|
+
text=True,
|
|
55
|
+
).stdout.strip()
|
|
56
|
+
if ip != "":
|
|
57
|
+
h.ips.append(ip)
|
|
58
|
+
continue
|
|
59
|
+
|
|
60
|
+
ip = run(
|
|
61
|
+
["ssh", h.ssh_hostname, "ipconfig", "getifaddr", "en1"],
|
|
62
|
+
capture_output=True,
|
|
63
|
+
text=True,
|
|
64
|
+
).stdout.strip()
|
|
65
|
+
if ip != "":
|
|
66
|
+
h.ips.append(ip)
|
|
67
|
+
continue
|
|
68
|
+
|
|
69
|
+
log_warning("Could not extract ip for", h.ssh_hostname)
|
|
70
|
+
|
|
71
|
+
|
|
72
|
+
def save_hostfile(args, hostfile):
|
|
73
|
+
if args.output_hostfile:
|
|
74
|
+
with open(args.output_hostfile, "w") as f:
|
|
75
|
+
json.dump(hostfile.to_json(), f, indent=4)
|
|
76
|
+
else:
|
|
77
|
+
print("Hostfile")
|
|
78
|
+
print("========")
|
|
79
|
+
print(json.dumps(hostfile.to_json(), indent=4))
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
def check_rdma(hosts, verbose=False, strict=True):
|
|
83
|
+
# Check whether the hosts are capable of RDMA over thunderbolt
|
|
84
|
+
log_f = log_warning if not strict else log_error
|
|
85
|
+
failed = False
|
|
86
|
+
for h in hosts:
|
|
87
|
+
log(verbose, "Checking that", h.ssh_hostname, "supports RDMA")
|
|
88
|
+
rdma_devs = (
|
|
89
|
+
run(["ssh", h.ssh_hostname, "ibv_devices"], capture_output=True, text=True)
|
|
90
|
+
.stdout.strip()
|
|
91
|
+
.split()
|
|
92
|
+
)
|
|
93
|
+
rdma_devs = [d for d in rdma_devs if d.startswith("rdma_")]
|
|
94
|
+
if not rdma_devs:
|
|
95
|
+
log_f(h.ssh_hostname, "does not seem to have RDMA enabled")
|
|
96
|
+
failed = True
|
|
97
|
+
|
|
98
|
+
if failed:
|
|
99
|
+
log_f()
|
|
100
|
+
log_f("Some of the hosts don't have RDMA enabled or they don't support RDMA.")
|
|
101
|
+
log_f()
|
|
102
|
+
log_f("See https://ml-explore.github.io/mlx/build/html/usage/distributed.html")
|
|
103
|
+
log_f("for instructions on how to enable RDMA.")
|
|
104
|
+
|
|
105
|
+
if failed and strict:
|
|
106
|
+
sys.exit(1)
|
|
107
|
+
|
|
108
|
+
return not failed
|
|
109
|
+
|
|
110
|
+
|
|
111
|
+
def can_auto_setup(hosts, sshinfo, auto_setup=False):
|
|
112
|
+
has_sudo = all(info.has_sudo for info in sshinfo)
|
|
113
|
+
if not has_sudo and auto_setup:
|
|
114
|
+
log_warning(
|
|
115
|
+
"Automatic setup requested but the following hosts do not have passwordless sudo"
|
|
116
|
+
)
|
|
117
|
+
for h, i in zip(hosts, sshinfo):
|
|
118
|
+
if not i.has_sudo:
|
|
119
|
+
log_warning(" - ", h.ssh_hostname)
|
|
120
|
+
return has_sudo
|
|
121
|
+
|
|
122
|
+
|
|
123
|
+
class IPConfigurator:
|
|
124
|
+
def __init__(self, hosts, tb_hosts, uuid_reverse_index):
|
|
125
|
+
assigned = set()
|
|
126
|
+
ips = defaultdict(list)
|
|
127
|
+
ip0 = 0
|
|
128
|
+
ip1 = 0
|
|
129
|
+
for src_node, h in enumerate(tb_hosts):
|
|
130
|
+
for src_port, p in enumerate(h.ports):
|
|
131
|
+
if not p.connected_to:
|
|
132
|
+
continue
|
|
133
|
+
if p.connected_to not in uuid_reverse_index:
|
|
134
|
+
continue
|
|
135
|
+
if (src_node, src_port) in assigned:
|
|
136
|
+
continue
|
|
137
|
+
|
|
138
|
+
dst_node, dst_port = uuid_reverse_index[p.connected_to]
|
|
139
|
+
|
|
140
|
+
ip_src = f"192.168.{ip0}.{ip1 + 1}"
|
|
141
|
+
ip_dst = f"192.168.{ip0}.{ip1 + 2}"
|
|
142
|
+
iface_src = p.iface
|
|
143
|
+
iface_dst = tb_hosts[dst_node].ports[dst_port].iface
|
|
144
|
+
|
|
145
|
+
ips[src_node, dst_node].append((iface_src, ip_src))
|
|
146
|
+
ips[dst_node, src_node].append((iface_dst, ip_dst))
|
|
147
|
+
|
|
148
|
+
assigned.add((src_node, src_port))
|
|
149
|
+
assigned.add((dst_node, dst_port))
|
|
150
|
+
|
|
151
|
+
ip1 += 4
|
|
152
|
+
if ip1 > 255:
|
|
153
|
+
ip0 += 1
|
|
154
|
+
ip1 = 0
|
|
155
|
+
if ip0 > 255:
|
|
156
|
+
raise ValueError("Ran out of available local IPs")
|
|
157
|
+
|
|
158
|
+
self.ips = ips
|
|
159
|
+
self.hosts = hosts
|
|
160
|
+
self.tb_hosts = tb_hosts
|
|
161
|
+
|
|
162
|
+
def setup(self, verbose=False, auto_setup=False):
|
|
163
|
+
netmask = "255.255.255.252"
|
|
164
|
+
for i, (h, th) in enumerate(zip(self.hosts, self.tb_hosts)):
|
|
165
|
+
command = ""
|
|
166
|
+
command += "sudo ifconfig bridge0 down\n"
|
|
167
|
+
for j in range(len(self.hosts)):
|
|
168
|
+
if i == j or (i, j) not in self.ips:
|
|
169
|
+
continue
|
|
170
|
+
for (iface, ip), (_, peer) in zip(self.ips[i, j], self.ips[j, i]):
|
|
171
|
+
command += f"sudo ifconfig {iface} inet {ip} netmask {netmask}\n"
|
|
172
|
+
command += f"sudo route change {peer} -interface {iface}\n"
|
|
173
|
+
if auto_setup:
|
|
174
|
+
print(f"Running auto setup for {h.ssh_hostname}")
|
|
175
|
+
command = command.strip().replace("\n", " ; ")
|
|
176
|
+
command = ["ssh", h.ssh_hostname, command]
|
|
177
|
+
log(verbose, shlex.join(command))
|
|
178
|
+
run(command)
|
|
179
|
+
else:
|
|
180
|
+
msg = f"Setup for {h.ssh_hostname}"
|
|
181
|
+
print(msg)
|
|
182
|
+
print("=" * len(msg))
|
|
183
|
+
print(command)
|
|
184
|
+
input("Enter to continue")
|
|
185
|
+
print()
|
|
186
|
+
|
|
187
|
+
|
|
188
|
+
def parse_hardware_ports(ports_string):
|
|
189
|
+
ports = {}
|
|
190
|
+
port_name = None
|
|
191
|
+
for l in ports_string.decode("utf-8").split("\n"):
|
|
192
|
+
if l.startswith("Hardware Port:"):
|
|
193
|
+
port_name = l.strip()[15:]
|
|
194
|
+
elif l.startswith("Device:"):
|
|
195
|
+
ports[port_name] = l.strip()[8:]
|
|
196
|
+
port_name = None
|
|
197
|
+
return ports
|
|
198
|
+
|
|
199
|
+
|
|
200
|
+
def extract_connectivity(hosts, verbose):
|
|
201
|
+
# Extract the current connectivity from the remote hosts
|
|
202
|
+
thunderbolt_connections = []
|
|
203
|
+
for h in hosts:
|
|
204
|
+
log(verbose, "Getting connectivity from", h.ssh_hostname)
|
|
205
|
+
thunderbolt_connections.append(
|
|
206
|
+
json.loads(
|
|
207
|
+
run(
|
|
208
|
+
[
|
|
209
|
+
"ssh",
|
|
210
|
+
h.ssh_hostname,
|
|
211
|
+
"system_profiler",
|
|
212
|
+
"SPThunderboltDataType",
|
|
213
|
+
"-json",
|
|
214
|
+
],
|
|
215
|
+
capture_output=True,
|
|
216
|
+
).stdout
|
|
217
|
+
)
|
|
218
|
+
)
|
|
219
|
+
interface_maps = []
|
|
220
|
+
for h in hosts:
|
|
221
|
+
log(verbose, "Getting interface names from", h.ssh_hostname)
|
|
222
|
+
interface_maps.append(
|
|
223
|
+
parse_hardware_ports(
|
|
224
|
+
run(
|
|
225
|
+
[
|
|
226
|
+
"ssh",
|
|
227
|
+
h.ssh_hostname,
|
|
228
|
+
"networksetup",
|
|
229
|
+
"-listallhardwareports",
|
|
230
|
+
],
|
|
231
|
+
capture_output=True,
|
|
232
|
+
).stdout
|
|
233
|
+
)
|
|
234
|
+
)
|
|
235
|
+
|
|
236
|
+
# Parse the connectivity into some simple dataclasses
|
|
237
|
+
tb_hosts = []
|
|
238
|
+
for c, iface_map in zip(thunderbolt_connections, interface_maps):
|
|
239
|
+
name = ""
|
|
240
|
+
ports = []
|
|
241
|
+
for t in c["SPThunderboltDataType"]:
|
|
242
|
+
uuid = t.get("domain_uuid_key")
|
|
243
|
+
if uuid is None:
|
|
244
|
+
continue
|
|
245
|
+
name = t["device_name_key"]
|
|
246
|
+
tag = t["receptacle_1_tag"]["receptacle_id_key"]
|
|
247
|
+
items = t.get("_items", [])
|
|
248
|
+
connected_items = [item for item in items if "domain_uuid_key" in item]
|
|
249
|
+
connected_to = (
|
|
250
|
+
connected_items[0]["domain_uuid_key"] if connected_items else None
|
|
251
|
+
)
|
|
252
|
+
iface = iface_map[f"Thunderbolt {tag}"]
|
|
253
|
+
ports.append(ThunderboltPort(iface, uuid, connected_to))
|
|
254
|
+
tb_hosts.append(ThunderboltHost(name, sorted(ports, key=lambda x: x.iface)))
|
|
255
|
+
|
|
256
|
+
# Create a reverse index to be able to map uuids to (host, port) quickly
|
|
257
|
+
uuid_reverse_index = {}
|
|
258
|
+
for i, h in enumerate(tb_hosts):
|
|
259
|
+
for j, p in enumerate(h.ports):
|
|
260
|
+
uuid_reverse_index[p.uuid] = (i, j)
|
|
261
|
+
|
|
262
|
+
return tb_hosts, uuid_reverse_index
|
|
263
|
+
|
|
264
|
+
|
|
265
|
+
def make_connectivity_matrix(tb_hosts, uuid_reverse_index):
|
|
266
|
+
connectivity = []
|
|
267
|
+
for i, h in enumerate(tb_hosts):
|
|
268
|
+
c = [0] * len(tb_hosts)
|
|
269
|
+
for p in h.ports:
|
|
270
|
+
if p.connected_to in uuid_reverse_index:
|
|
271
|
+
j, _ = uuid_reverse_index[p.connected_to]
|
|
272
|
+
c[j] += 1
|
|
273
|
+
connectivity.append(c)
|
|
274
|
+
return connectivity
|
|
275
|
+
|
|
276
|
+
|
|
277
|
+
def tb_connectivity_to_dot(hosts, tb_hosts, uuid_reverse_index):
|
|
278
|
+
# Make ids per node
|
|
279
|
+
names = []
|
|
280
|
+
for i in range(len(tb_hosts)):
|
|
281
|
+
n = ""
|
|
282
|
+
j = i
|
|
283
|
+
while True:
|
|
284
|
+
n += chr(97 + j % 26)
|
|
285
|
+
j //= 26
|
|
286
|
+
if j == 0:
|
|
287
|
+
break
|
|
288
|
+
names.append(n)
|
|
289
|
+
|
|
290
|
+
print("graph G {")
|
|
291
|
+
print(" node [shape=rectangle];")
|
|
292
|
+
for i, h in enumerate(hosts):
|
|
293
|
+
print(f' {names[i]} [label="{h.ssh_hostname}"];')
|
|
294
|
+
for i, h in enumerate(tb_hosts):
|
|
295
|
+
for p in h.ports:
|
|
296
|
+
if not p.connected_to:
|
|
297
|
+
continue
|
|
298
|
+
dst = uuid_reverse_index[p.connected_to]
|
|
299
|
+
if dst[0] < i:
|
|
300
|
+
continue
|
|
301
|
+
print(f" {names[i]} -- {names[dst[0]]}", end="")
|
|
302
|
+
print(f' [label="{p.iface}/{tb_hosts[dst[0]].ports[dst[1]].iface}"]')
|
|
303
|
+
print("}")
|
|
304
|
+
|
|
305
|
+
|
|
306
|
+
def extract_rings(connectivity):
|
|
307
|
+
rings = []
|
|
308
|
+
existing_rings = set()
|
|
309
|
+
num_nodes = len(connectivity)
|
|
310
|
+
|
|
311
|
+
def dfs(start_node, node, path, visited):
|
|
312
|
+
path.append(node)
|
|
313
|
+
visited.add(node)
|
|
314
|
+
for j in range(num_nodes):
|
|
315
|
+
if connectivity[node][j] <= 0:
|
|
316
|
+
continue
|
|
317
|
+
if j == start_node:
|
|
318
|
+
yield path[:]
|
|
319
|
+
if j not in visited:
|
|
320
|
+
yield from dfs(start_node, j, path, visited)
|
|
321
|
+
path.pop()
|
|
322
|
+
visited.remove(node)
|
|
323
|
+
|
|
324
|
+
for start in range(num_nodes):
|
|
325
|
+
for r in dfs(start, start, [], set()):
|
|
326
|
+
cnt = min(connectivity[r[i]][r[(i + 1) % len(r)]] for i in range(len(r)))
|
|
327
|
+
rkey = tuple(sorted(r))
|
|
328
|
+
if rkey not in existing_rings:
|
|
329
|
+
rings.append((r, cnt))
|
|
330
|
+
existing_rings.add(rkey)
|
|
331
|
+
|
|
332
|
+
return sorted(rings, key=lambda x: -len(x[0]))
|
|
333
|
+
|
|
334
|
+
|
|
335
|
+
def check_valid_mesh(hosts, connectivity, strict=True):
|
|
336
|
+
num_nodes = len(connectivity)
|
|
337
|
+
for i in range(num_nodes):
|
|
338
|
+
for j in range(num_nodes):
|
|
339
|
+
if i == j:
|
|
340
|
+
continue
|
|
341
|
+
if connectivity[i][j] <= 0:
|
|
342
|
+
if strict:
|
|
343
|
+
log_error(
|
|
344
|
+
f"Incomplete mesh, {hosts[i].ssh_hostname} is not connected to {hosts[j].ssh_hostname}"
|
|
345
|
+
)
|
|
346
|
+
log_error()
|
|
347
|
+
log_error("Try passing --dot to visualize the connectivity")
|
|
348
|
+
sys.exit(1)
|
|
349
|
+
else:
|
|
350
|
+
return False
|
|
351
|
+
return True
|
|
352
|
+
|
|
353
|
+
|
|
354
|
+
def check_valid_ring(hosts, rings, strict=True):
|
|
355
|
+
has_ring = len(rings) > 0 and len(rings[0][0]) == len(hosts)
|
|
356
|
+
if strict and not has_ring:
|
|
357
|
+
log_error("Could not find a full ring.")
|
|
358
|
+
log_error()
|
|
359
|
+
log_error("Try passing --dot to visualize the connectivity")
|
|
360
|
+
if len(rings) > 0:
|
|
361
|
+
log_error("Rings found:")
|
|
362
|
+
for r in rings:
|
|
363
|
+
log_error(f" - {','.join(hosts[i].ssh_hostname for i in r)}")
|
|
364
|
+
sys.exit(1)
|
|
365
|
+
return has_ring
|
|
366
|
+
|
|
367
|
+
|
|
368
|
+
def check_ssh_connections(hosts):
|
|
369
|
+
results = [None] * len(hosts)
|
|
370
|
+
|
|
371
|
+
def _check(hostname, i):
|
|
372
|
+
info = SSHInfo(False, False)
|
|
373
|
+
results[i] = info
|
|
374
|
+
|
|
375
|
+
# Check for ssh
|
|
376
|
+
result = run(
|
|
377
|
+
[
|
|
378
|
+
"ssh",
|
|
379
|
+
"-o",
|
|
380
|
+
"BatchMode=yes",
|
|
381
|
+
"-o",
|
|
382
|
+
"ConnectTimeout=5",
|
|
383
|
+
hostname,
|
|
384
|
+
"echo",
|
|
385
|
+
"success",
|
|
386
|
+
],
|
|
387
|
+
stdout=DEVNULL,
|
|
388
|
+
stderr=DEVNULL,
|
|
389
|
+
)
|
|
390
|
+
info.can_ssh = result.returncode == 0
|
|
391
|
+
if not info.can_ssh:
|
|
392
|
+
return
|
|
393
|
+
|
|
394
|
+
# Check for sudo
|
|
395
|
+
result = run(
|
|
396
|
+
[
|
|
397
|
+
"ssh",
|
|
398
|
+
"-o",
|
|
399
|
+
"BatchMode=yes",
|
|
400
|
+
"-o",
|
|
401
|
+
"ConnectTimeout=5",
|
|
402
|
+
hostname,
|
|
403
|
+
"sudo",
|
|
404
|
+
"ls",
|
|
405
|
+
],
|
|
406
|
+
stdout=DEVNULL,
|
|
407
|
+
stderr=DEVNULL,
|
|
408
|
+
)
|
|
409
|
+
info.has_sudo = result.returncode == 0
|
|
410
|
+
|
|
411
|
+
threads = [
|
|
412
|
+
threading.Thread(target=_check, args=(h.ssh_hostname, i))
|
|
413
|
+
for i, h in enumerate(hosts)
|
|
414
|
+
]
|
|
415
|
+
for t in threads:
|
|
416
|
+
t.start()
|
|
417
|
+
for t in threads:
|
|
418
|
+
t.join()
|
|
419
|
+
|
|
420
|
+
if not all(results):
|
|
421
|
+
log_error("Could not ssh to the following hosts:")
|
|
422
|
+
for i, h in enumerate(hosts):
|
|
423
|
+
if not results[i]:
|
|
424
|
+
log_error(" - ", h.ssh_hostname)
|
|
425
|
+
log_error()
|
|
426
|
+
log_error("Maybe they are not set-up for password-less ssh?")
|
|
427
|
+
sys.exit(1)
|
|
428
|
+
|
|
429
|
+
return results
|
|
430
|
+
|
|
431
|
+
|
|
432
|
+
def prepare_ethernet_hostfile(args, hosts):
|
|
433
|
+
log(args.verbose, f"Preparing an ethernet hostfile")
|
|
434
|
+
add_ips(hosts, args.verbose)
|
|
435
|
+
|
|
436
|
+
hostfile = Hostfile(
|
|
437
|
+
[Host(i, h.ssh_hostname, h.ips, []) for i, h in enumerate(hosts)], "", args.env
|
|
438
|
+
)
|
|
439
|
+
|
|
440
|
+
save_hostfile(args, hostfile)
|
|
441
|
+
|
|
442
|
+
|
|
443
|
+
def configure_ring(args, hosts, ips, ring, sshinfo):
|
|
444
|
+
log(args.verbose, "Prepare a ring hostfile")
|
|
445
|
+
ring, count = ring
|
|
446
|
+
ring_hosts = []
|
|
447
|
+
for i, node in enumerate(ring):
|
|
448
|
+
h = hosts[node]
|
|
449
|
+
peer = ring[i - 1]
|
|
450
|
+
ring_hosts.append(
|
|
451
|
+
Host(
|
|
452
|
+
i, h.ssh_hostname, [ips.ips[node, peer][c][1] for c in range(count)], []
|
|
453
|
+
)
|
|
454
|
+
)
|
|
455
|
+
hostfile = Hostfile(ring_hosts, "ring", args.env)
|
|
456
|
+
|
|
457
|
+
has_sudo = can_auto_setup(hosts, sshinfo, args.auto_setup)
|
|
458
|
+
ips.setup(verbose=args.verbose, auto_setup=args.auto_setup and has_sudo)
|
|
459
|
+
|
|
460
|
+
save_hostfile(args, hostfile)
|
|
461
|
+
|
|
462
|
+
|
|
463
|
+
def configure_jaccl(args, hosts, ips, sshinfo):
|
|
464
|
+
log(args.verbose, "Prepare a jaccl hostfile")
|
|
465
|
+
add_ips(hosts, args.verbose)
|
|
466
|
+
|
|
467
|
+
jaccl_hosts = []
|
|
468
|
+
for i, h in enumerate(hosts):
|
|
469
|
+
rdma = []
|
|
470
|
+
for j in range(len(hosts)):
|
|
471
|
+
if i == j:
|
|
472
|
+
rdma.append(None)
|
|
473
|
+
else:
|
|
474
|
+
rdma.append(f"rdma_{ips.ips[i, j][0][0]}")
|
|
475
|
+
jaccl_hosts.append(Host(i, h.ssh_hostname, h.ips, rdma))
|
|
476
|
+
hostfile = Hostfile(jaccl_hosts, "jaccl", args.env)
|
|
477
|
+
|
|
478
|
+
has_sudo = can_auto_setup(hosts, sshinfo, args.auto_setup)
|
|
479
|
+
ips.setup(verbose=args.verbose, auto_setup=args.auto_setup and has_sudo)
|
|
480
|
+
|
|
481
|
+
save_hostfile(args, hostfile)
|
|
482
|
+
|
|
483
|
+
|
|
484
|
+
def configure_jaccl_ring(args, hosts, ips, ring, sshinfo):
|
|
485
|
+
log(args.verbose, "Prepare a jaccl-ring hostfile")
|
|
486
|
+
add_ips(hosts, args.verbose)
|
|
487
|
+
|
|
488
|
+
jaccl_hosts = []
|
|
489
|
+
num_nodes = len(hosts)
|
|
490
|
+
ring, count = ring
|
|
491
|
+
for i, node in enumerate(ring):
|
|
492
|
+
h = hosts[node]
|
|
493
|
+
peer_left = ring[i - 1]
|
|
494
|
+
peer_right = ring[(i + 1) % num_nodes]
|
|
495
|
+
rdmas = []
|
|
496
|
+
for j in range(len(hosts)):
|
|
497
|
+
if j not in (peer_left, peer_right):
|
|
498
|
+
rdmas.append(None)
|
|
499
|
+
else:
|
|
500
|
+
rdma = []
|
|
501
|
+
for c in range(count):
|
|
502
|
+
rdma.append(f"rdma_{ips.ips[i, j][c][0]}")
|
|
503
|
+
rdmas.append(rdma[0] if count == 1 else rdma)
|
|
504
|
+
jaccl_hosts.append(Host(i, h.ssh_hostname, h.ips, rdmas))
|
|
505
|
+
hostfile = Hostfile(jaccl_hosts, "jaccl-ring", args.env)
|
|
506
|
+
|
|
507
|
+
has_sudo = can_auto_setup(hosts, sshinfo, args.auto_setup)
|
|
508
|
+
ips.setup(verbose=args.verbose, auto_setup=args.auto_setup and has_sudo)
|
|
509
|
+
|
|
510
|
+
save_hostfile(args, hostfile)
|
|
511
|
+
|
|
512
|
+
|
|
513
|
+
def prepare_tb_hostfile(args, hosts, sshinfo):
|
|
514
|
+
log(args.verbose, f"Preparing for communication over thunderbolt")
|
|
515
|
+
tb_hosts, uuid_reverse_index = extract_connectivity(hosts, args.verbose)
|
|
516
|
+
|
|
517
|
+
if args.dot:
|
|
518
|
+
tb_connectivity_to_dot(hosts, tb_hosts, uuid_reverse_index)
|
|
519
|
+
return
|
|
520
|
+
|
|
521
|
+
ips = IPConfigurator(hosts, tb_hosts, uuid_reverse_index)
|
|
522
|
+
connectivity = make_connectivity_matrix(tb_hosts, uuid_reverse_index)
|
|
523
|
+
|
|
524
|
+
if args.backend is None:
|
|
525
|
+
rings = extract_rings(connectivity)
|
|
526
|
+
has_mesh = check_valid_mesh(hosts, connectivity, False)
|
|
527
|
+
has_ring = check_valid_ring(hosts, rings, False)
|
|
528
|
+
has_rdma = check_rdma(hosts, args.verbose, False)
|
|
529
|
+
|
|
530
|
+
if not has_ring and not has_mesh:
|
|
531
|
+
log_error("Neither thunderbolt mesh nor ring found.")
|
|
532
|
+
log_error("Perhaps run with --dot to generate a plot of the connectivity.")
|
|
533
|
+
sys.exit(1)
|
|
534
|
+
|
|
535
|
+
elif has_rdma and has_mesh:
|
|
536
|
+
configure_jaccl(args, hosts, ips, sshinfo)
|
|
537
|
+
|
|
538
|
+
elif has_rdma and has_ring:
|
|
539
|
+
configure_jaccl_ring(args, hosts, ips, rings[0], sshinfo)
|
|
540
|
+
|
|
541
|
+
elif has_ring:
|
|
542
|
+
configure_ring(args, hosts, ips, rings[0], sshinfo)
|
|
543
|
+
|
|
544
|
+
else:
|
|
545
|
+
log_error("RDMA is not available and ring is not found.")
|
|
546
|
+
log_error("Perhaps run with --dot to generate a plot of the connectivity.")
|
|
547
|
+
sys.exit(1)
|
|
548
|
+
|
|
549
|
+
elif args.backend == "ring":
|
|
550
|
+
rings = extract_rings(connectivity)
|
|
551
|
+
check_valid_ring(hosts, rings)
|
|
552
|
+
configure_ring(args, hosts, ips, rings[0], sshinfo)
|
|
553
|
+
|
|
554
|
+
elif args.backend == "jaccl":
|
|
555
|
+
check_valid_mesh(hosts, connectivity)
|
|
556
|
+
check_rdma(hosts, args.verbose)
|
|
557
|
+
configure_jaccl(args, hosts, ips, sshinfo)
|
|
558
|
+
|
|
559
|
+
elif args.backend == "jaccl-ring":
|
|
560
|
+
rings = extract_rings(connectivity)
|
|
561
|
+
check_valid_ring(hosts, rings)
|
|
562
|
+
check_rdma(hosts, args.verbose)
|
|
563
|
+
configure_jaccl_ring(args, hosts, ips, rings[0], sshinfo)
|
|
564
|
+
|
|
565
|
+
|
|
566
|
+
def main():
|
|
567
|
+
parser = argparse.ArgumentParser(
|
|
568
|
+
description="Configure remote machines for use with MLX distributed"
|
|
569
|
+
)
|
|
570
|
+
parser.add_argument(
|
|
571
|
+
"--verbose", action="store_true", help="Print debug messages in stdout"
|
|
572
|
+
)
|
|
573
|
+
parser.add_argument(
|
|
574
|
+
"--hosts", default="127.0.0.1", help="A comma separated list of hosts"
|
|
575
|
+
)
|
|
576
|
+
parser.add_argument("--hostfile", help="The file containing the hosts")
|
|
577
|
+
parser.add_argument(
|
|
578
|
+
"--over",
|
|
579
|
+
choices=["thunderbolt", "ethernet"],
|
|
580
|
+
default="thunderbolt",
|
|
581
|
+
help="What type of connectivity to configure",
|
|
582
|
+
required=True,
|
|
583
|
+
)
|
|
584
|
+
parser.add_argument(
|
|
585
|
+
"--output-hostfile", help="If provided, save the hostfile to this path"
|
|
586
|
+
)
|
|
587
|
+
parser.add_argument(
|
|
588
|
+
"--auto-setup",
|
|
589
|
+
"--no-auto-setup",
|
|
590
|
+
action=OptionalBoolAction,
|
|
591
|
+
nargs=0,
|
|
592
|
+
dest="auto_setup",
|
|
593
|
+
default=None,
|
|
594
|
+
)
|
|
595
|
+
parser.add_argument(
|
|
596
|
+
"--dot", action="store_true", help="Output the topology in DOT format and exit"
|
|
597
|
+
)
|
|
598
|
+
parser.add_argument(
|
|
599
|
+
"--backend",
|
|
600
|
+
choices=["ring", "jaccl", "jaccl-ring"],
|
|
601
|
+
default=None,
|
|
602
|
+
help="Which distributed backend to configure",
|
|
603
|
+
)
|
|
604
|
+
parser.add_argument(
|
|
605
|
+
"--env",
|
|
606
|
+
action="append",
|
|
607
|
+
default=[],
|
|
608
|
+
help="Set environment variables for the jobs",
|
|
609
|
+
)
|
|
610
|
+
args = parser.parse_args()
|
|
611
|
+
|
|
612
|
+
if args.hostfile is not None:
|
|
613
|
+
hosts = Hostfile.from_file(args.hostfile).hosts
|
|
614
|
+
else:
|
|
615
|
+
hosts = Hostfile.from_list(args.hosts).hosts
|
|
616
|
+
|
|
617
|
+
# Check that we can ssh
|
|
618
|
+
log(
|
|
619
|
+
args.verbose,
|
|
620
|
+
f"Checking for ssh access for {', '.join(h.ssh_hostname for h in hosts)}",
|
|
621
|
+
)
|
|
622
|
+
sshinfo = check_ssh_connections(hosts)
|
|
623
|
+
|
|
624
|
+
# Prepare a hostfile for communication over ethernet using the ips of the
|
|
625
|
+
# provided hostnames
|
|
626
|
+
if args.over == "ethernet":
|
|
627
|
+
prepare_ethernet_hostfile(args, hosts)
|
|
628
|
+
|
|
629
|
+
# Configure the macs for communication over thunderbolt, both via RDMA and IP
|
|
630
|
+
else:
|
|
631
|
+
prepare_tb_hostfile(args, hosts, sshinfo)
|