update circle

.circleci/config.yml

@@ -7,18 +7,6 @@ parameters:
   nightly_build:
     type: boolean
     default: false
-  weekly_build:
-    type: boolean
-    default: false
-  test_release:
-    type: boolean
-    default: false
-  linux_release:
-    type: boolean
-    default: false
-  cuda_release:
-    type: boolean
-    default: false
 
 jobs:
   build_documentation:
@@ -301,9 +289,9 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      extra_env:
+      build_env:
         type: string
-        default: "DEV_RELEASE=1"
+        default: ""
     docker:
       - image: ubuntu:20.04
     steps:
@@ -332,17 +320,22 @@ jobs:
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> pip install . -v
+            << parameters.build_env >> pip install . -v
             pip install typing_extensions
             python setup.py generate_stubs
-            << parameters.extra_env >> python -m build --wheel
+            << parameters.build_env >> python -m build --wheel
             auditwheel show dist/*
             auditwheel repair dist/* --plat manylinux_2_31_x86_64
-      - run:
-          name: Upload package
-          command: |
-            source env/bin/activate
-            twine upload wheelhouse/*
+            << parameters.build_env >> MLX_BUILD_COMMON=1 \
+              python -m build --wheel --outdir wheelhouse
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload packages
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*
       - store_artifacts:
           path: wheelhouse/
 
@@ -351,9 +344,9 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      extra_env:
+      build_env:
         type: string
-        default: "DEV_RELEASE=1"
+        default: ""
     machine:
       image: linux-cuda-12:default
       resource_class: gpu.nvidia.small.gen2
@@ -364,25 +357,29 @@ jobs:
           command: |
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install zip
             python -m venv env
             source env/bin/activate
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> \
+            << parameters.build_env >> \
               CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               pip install ".[dev]" -v
             python setup.py generate_stubs
-            << parameters.extra_env >> \
+            << parameters.build_env >> \
               CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               python -m build --wheel
             bash python/scripts/repair_cuda.sh
-      - run:
-          name: Upload package
-          command: |
-            source env/bin/activate
-            twine upload wheelhouse/*.whl
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload package
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
 
@@ -394,8 +391,6 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
           matrix:
@@ -409,8 +404,6 @@ workflows:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
           filters:
@@ -501,7 +494,17 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
+              build_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              build_env: ["PYPI_RELEASE=1"]
 
   prb:
     when:
@@ -580,99 +583,21 @@ workflows:
               - macosx_deployment_target: "15.0"
                 xcode_version: "15.0.0"
                 python_version: "3.13"
-  weekly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.weekly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
       - build_linux_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
-  cuda_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.cuda_release >>
-    jobs:
       - build_cuda_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]

docs/src/install.rst

@@ -38,8 +38,16 @@ and SM 7.0 (Volta) and up. To install MLX with CUDA support, run:
 
 .. code-block:: shell
 
-    pip install mlx-cuda
+    pip install "mlx[cuda]"
 
+CPU only (Linux)
+^^^^^^^^^^^^^^^^
+
+For a CPU-only version of MLX that runs on Linux use:
+
+.. code-block:: shell
+
+    pip install "mlx[cpu]"
 
 Troubleshooting
 ^^^^^^^^^^^^^^^

mlx/distributed/ring/ring.cpp

@@ -22,20 +22,78 @@
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/distributed/distributed.h"
 #include "mlx/distributed/distributed_impl.h"
-#include "mlx/dtype_utils.h"
 #include "mlx/threadpool.h"
 
 #ifndef SOL_TCP
 #define SOL_TCP IPPROTO_TCP
 #endif
 
+#define SWITCH_TYPE(x, ...)  \
+  switch ((x).dtype()) {     \
+    case bool_: {            \
+      using T = bool;        \
+      __VA_ARGS__;           \
+    } break;                 \
+    case int8: {             \
+      using T = int8_t;      \
+      __VA_ARGS__;           \
+    } break;                 \
+    case int16: {            \
+      using T = int16_t;     \
+      __VA_ARGS__;           \
+    } break;                 \
+    case int32: {            \
+      using T = int32_t;     \
+      __VA_ARGS__;           \
+    } break;                 \
+    case int64: {            \
+      using T = int64_t;     \
+      __VA_ARGS__;           \
+    } break;                 \
+    case uint8: {            \
+      using T = uint8_t;     \
+      __VA_ARGS__;           \
+    } break;                 \
+    case uint16: {           \
+      using T = uint16_t;    \
+      __VA_ARGS__;           \
+    } break;                 \
+    case uint32: {           \
+      using T = uint32_t;    \
+      __VA_ARGS__;           \
+    } break;                 \
+    case uint64: {           \
+      using T = uint64_t;    \
+      __VA_ARGS__;           \
+    } break;                 \
+    case bfloat16: {         \
+      using T = bfloat16_t;  \
+      __VA_ARGS__;           \
+    } break;                 \
+    case float16: {          \
+      using T = float16_t;   \
+      __VA_ARGS__;           \
+    } break;                 \
+    case float32: {          \
+      using T = float;       \
+      __VA_ARGS__;           \
+    } break;                 \
+    case float64: {          \
+      using T = double;      \
+      __VA_ARGS__;           \
+    } break;                 \
+    case complex64: {        \
+      using T = complex64_t; \
+      __VA_ARGS__;           \
+    } break;                 \
+  }
+
 namespace mlx::core::distributed::ring {
 
 constexpr const size_t ALL_SUM_SIZE = 8 * 1024 * 1024;
 constexpr const size_t ALL_SUM_BUFFERS = 2;
 constexpr const int CONN_ATTEMPTS = 5;
 constexpr const int CONN_WAIT = 1000;
-constexpr const int INIT_TIMEOUT = 20000;
 
 using GroupImpl = mlx::core::distributed::detail::GroupImpl;
 using json = nlohmann::json;
@@ -445,7 +503,6 @@ std::vector<int> make_connections(
 
   return sockets;
 }
-
 template <typename T>
 struct SumOp {
   void operator()(const T* input, T* output, size_t N) {
@@ -493,27 +550,19 @@ class RingGroup : public GroupImpl {
     size_ = nodes.size();
     int connect_to = (rank_ + 1) % size_;
 
-    // Initialize the ring by making all the connections
-    log_info(verbose_, "Rank", rank_, "accepting");
-    log_info(verbose_, "Rank", rank_, "connecting to", connect_to);
-    auto sl = std::async(std::launch::async, accept_connections, nodes[rank_]);
-    auto sr = std::async(
-        std::launch::async, make_connections, nodes[connect_to], verbose);
-    std::future_status status_sl, status_sr;
-    for (int i = 0; i < 10; i++) {
-      status_sl = sl.wait_for(std::chrono::milliseconds(INIT_TIMEOUT / 10));
-      status_sr = sl.wait_for(std::chrono::milliseconds(INIT_TIMEOUT / 10));
-      if (status_sl == std::future_status::ready &&
-          status_sr == std::future_status::ready) {
-        break;
-      }
+    // We define the connection order by having the rank_ == size_ - 1 connect
+    // first and accept after.
+    if (rank_ < connect_to) {
+      log_info(verbose_, "Rank", rank_, "accepting");
+      sockets_left_ = std::move(accept_connections(nodes[rank_]));
+      log_info(verbose_, "Rank", rank_, "connecting to", connect_to);
+      sockets_right_ = std::move(make_connections(nodes[connect_to], verbose));
+    } else {
+      log_info(verbose_, "Rank", rank_, "connecting to", connect_to);
+      sockets_right_ = std::move(make_connections(nodes[connect_to], verbose));
+      log_info(verbose_, "Rank", rank_, "accepting");
+      sockets_left_ = std::move(accept_connections(nodes[rank_]));
     }
-    if (status_sl != std::future_status::ready ||
-        status_sr != std::future_status::ready) {
-      throw std::runtime_error("[ring] Ring initialization timed out");
-    }
-    sockets_left_ = std::move(sl.get());
-    sockets_right_ = std::move(sr.get());
 
     // Failure if we couldn't make right or left sockets
     if (sockets_right_.empty()) {
@@ -579,24 +628,18 @@ class RingGroup : public GroupImpl {
   }
 
   void all_sum(const array& input, array& output, Stream stream) override {
-    dispatch_all_types(output.dtype(), [&](auto type_tag) {
-      using T = MLX_GET_TYPE(type_tag);
-      all_reduce<T, SumOp<T>>(input, output, stream, SumOp<T>());
-    });
+    SWITCH_TYPE(
+        output, all_reduce<T, SumOp<T>>(input, output, stream, SumOp<T>()));
   }
 
   void all_max(const array& input, array& output, Stream stream) override {
-    dispatch_all_types(output.dtype(), [&](auto type_tag) {
-      using T = MLX_GET_TYPE(type_tag);
-      all_reduce<T, MaxOp<T>>(input, output, stream, MaxOp<T>());
-    });
+    SWITCH_TYPE(
+        output, all_reduce<T, MaxOp<T>>(input, output, stream, MaxOp<T>()));
   }
 
   void all_min(const array& input, array& output, Stream stream) override {
-    dispatch_all_types(output.dtype(), [&](auto type_tag) {
-      using T = MLX_GET_TYPE(type_tag);
-      all_reduce<T, MinOp<T>>(input, output, stream, MinOp<T>());
-    });
+    SWITCH_TYPE(
+        output, all_reduce<T, MinOp<T>>(input, output, stream, MinOp<T>()));
   }
 
   std::shared_ptr<GroupImpl> split(int color, int key = -1) override {

python/scripts/repair_cuda.sh

@@ -3,15 +3,19 @@
 auditwheel repair dist/* \
   --plat manylinux_2_35_x86_64 \
   --exclude libcublas* \
-  --exclude libnvrtc*
+  --exclude libnvrtc* \
+  -w wheel_tmp
 
-cd wheelhouse
+
+mkdir wheelhouse
+cd wheel_tmp
 repaired_wheel=$(find . -name "*.whl" -print -quit)
 unzip -q "${repaired_wheel}"
+rm "${repaired_wheel}"
 core_so=$(find mlx -name "core*.so" -print -quit)
 rpath=$(patchelf --print-rpath "${core_so}")
 rpath=$rpath:\$ORIGIN/../nvidia/cublas/lib:\$ORIGIN/../nvidia/cuda_nvrtc/lib
 patchelf --force-rpath --set-rpath "$rpath" "$core_so"
 
 # Re-zip the repaired wheel
-zip -r -q "${repaired_wheel}" .
+zip -r -q "../wheelhouse/${repaired_wheel}" .

python/tests/mpi_test_distributed.py

@@ -4,7 +4,6 @@ import unittest
 
 import mlx.core as mx
 import mlx_distributed_tests
-import mlx_tests
 
 
 class TestMPIDistributed(mlx_distributed_tests.MLXDistributedCommonTestCase):
@@ -151,4 +150,4 @@ class TestMPIDistributed(mlx_distributed_tests.MLXDistributedCommonTestCase):
 
 
 if __name__ == "__main__":
-    mlx_tests.MLXTestRunner()
+    unittest.main()

python/tests/ring_test_distributed.py

@@ -4,7 +4,6 @@ import unittest
 
 import mlx.core as mx
 import mlx_distributed_tests
-import mlx_tests
 
 
 class TestRingDistributed(mlx_distributed_tests.MLXDistributedCommonTestCase):

setup.py

@@ -5,10 +5,11 @@ import os
 import platform
 import re
 import subprocess
+from functools import partial
 from pathlib import Path
 from subprocess import run
 
-from setuptools import Command, Extension, find_namespace_packages, setup
+from setuptools import Command, Extension, setup
 from setuptools.command.build_ext import build_ext
 
 
@@ -165,19 +166,28 @@ with open(Path(__file__).parent / "README.md", encoding="utf-8") as f:
 # The information here can also be placed in setup.cfg - better separation of
 # logic and declaration, and simpler if you include description/version in a file.
 if __name__ == "__main__":
-    packages = find_namespace_packages(
-        where="python", exclude=["src", "tests", "tests.*"]
-    )
     package_dir = {"": "python"}
     package_data = {"mlx": ["lib/*", "include/*", "share/*"], "mlx.core": ["*.pyi"]}
-    install_requires = []
+    packages = [
+        "mlx",
+        "mlx.nn",
+        "mlx.nn.layers",
+        "mlx.optimizers",
+    ]
+
+    is_release = "PYPI_RELEASE" in os.environ
+    build_macos = platform.system() == "Darwin"
     build_cuda = "MLX_BUILD_CUDA=ON" in os.environ.get("CMAKE_ARGS", "")
+    build_common = "MLX_BUILD_COMMON" in os.environ
+
+    install_requires = []
     if build_cuda:
         install_requires = ["nvidia-cublas-cu12", "nvidia-cuda-nvrtc-cu12"]
+    version = get_version()
 
-    setup(
-        name="mlx-cuda" if build_cuda else "mlx",
-        version=get_version(),
+    _setup = partial(
+        setup,
+        version=version,
         author="MLX Contributors",
         author_email="mlx@group.apple.com",
         description="A framework for machine learning on Apple silicon.",
@@ -185,29 +195,56 @@ if __name__ == "__main__":
         long_description_content_type="text/markdown",
         license="MIT",
         url="https://github.com/ml-explore/mlx",
-        packages=packages,
         package_dir=package_dir,
         package_data=package_data,
-        include_package_data=True,
-        install_requires=install_requires,
-        extras_require={
-            "dev": [
-                "nanobind==2.4.0",
-                "numpy",
-                "pre-commit",
-                "setuptools>=42",
-                "torch",
-                "typing_extensions",
-            ],
-        },
-        entry_points={
-            "console_scripts": [
-                "mlx.launch = mlx.distributed_run:main",
-                "mlx.distributed_config = mlx.distributed_run:distributed_config",
-            ]
-        },
-        ext_modules=[CMakeExtension("mlx.core")],
-        cmdclass={"build_ext": CMakeBuild, "generate_stubs": GenerateStubs},
         zip_safe=False,
         python_requires=">=3.9",
+        install_requires=install_requires,
     )
+
+    extras = {
+        "dev": [
+            "nanobind==2.4.0",
+            "numpy",
+            "pre-commit",
+            "setuptools>=42",
+            "torch",
+            "typing_extensions",
+        ],
+    }
+    entry_points = {
+        "console_scripts": [
+            "mlx.launch = mlx.distributed_run:main",
+            "mlx.distributed_config = mlx.distributed_run:distributed_config",
+        ]
+    }
+
+    if not is_release or build_macos:
+        _setup(
+            name="mlx",
+            include_package_data=True,
+            packages=packages,
+            extras_require=extras,
+            entry_points=entry_points,
+            ext_modules=[CMakeExtension("mlx.core")],
+            cmdclass={"build_ext": CMakeBuild, "generate_stubs": GenerateStubs},
+        )
+    elif build_common:
+        extras["cpu"] = [f"mlx-cpu=={version}"]
+        extras["cuda"] = [f"mlx-cuda=={version}"]
+        _setup(
+            name="mlx",
+            packages=["mlx"],
+            extras_require=extras,
+            entry_points=entry_points,
+            exclude_package_data=package_data,
+        )
+    else:
+        _setup(
+            name="mlx-cuda" if build_cuda else "mlx-cpu",
+            include_package_data=True,
+            packages=packages,
+            extras_require=extras,
+            ext_modules=[CMakeExtension("mlx.core")],
+            cmdclass={"build_ext": CMakeBuild, "generate_stubs": GenerateStubs},
+        )