🔄 Auto-Sync: Apple Upstream Repository

.github/scripts/build-linux-cuda-cmake.sh

@@ -16,6 +16,14 @@ rm -rf build
 mkdir -p build
 pushd build
 cmake -DMLX_BUILD_METAL=OFF -DMLX_BUILD_CUDA=ON -DMLX_C_BUILD_EXAMPLES=OFF .. -G Ninja
+
+# Verify the local mlx patch is actually in place: without it the examples below
+# crash on exit, which is easy to misread as an unrelated CUDA failure.
+git -C _deps/mlx-src apply --reverse --check "$PWD/../cmake/mlx.patch" || {
+  echo "error: cmake/mlx.patch is not applied to _deps/mlx-src"
+  exit 1
+}
+
 ninja
 ./example1 --device gpu
 ./tutorial --device gpu

.gitmodules

@@ -0,0 +1,6 @@
+[submodule "submodules/mlx"]
+	path = Source/Cmlx/mlx
+	url = https://github.com/ml-explore/mlx
+[submodule "submodules/mlx-c"]
+	path = Source/Cmlx/mlx-c
+	url = https://github.com/ml-explore/mlx-c

CMakeLists.txt

@@ -14,7 +14,9 @@ set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 # ----------------------------- Configuration -----------------------------
 # note: mirrors a subset of MLX options exactly (1:1 mapping)
 
-option(MLX_BUILD_EXAMPLES "Build examples for mlx" ON)
+# note: MLX_BUILD_EXAMPLES is owned by mlx itself (see below), so the Swift
+# examples use their own option name to avoid colliding with mlx's C++ examples
+option(MLX_SWIFT_BUILD_EXAMPLES "Build Swift examples for mlx-swift" ON)
 option(MLX_BUILD_METAL "Build metal backend" ON)
 option(MLX_BUILD_CUDA "Build cuda backend" OFF)
 
@@ -26,21 +28,61 @@ if(POLICY CMP0135)
   cmake_policy(SET CMP0135 NEW)
 endif()
 
+# mlx
+
+# note: patches are applied via cmake/apply-patch.cmake so that they are
+# idempotent (the patch step re-runs on every update step) while still failing
+# the build loudly if a patch does not apply -- a silently unpatched dependency
+# is very hard to diagnose (it typically shows up as a runtime failure).
+set(apply_patch ${CMAKE_COMMAND} -DREPO=<SOURCE_DIR>)
+set(apply_patch_script -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/apply-patch.cmake)
+
+# two local patches for CUDA:
+#
+# * https://github.com/ml-explore/mlx/pull/4480: leak the global command encoder
+#   map rather than synchronizing on process shutdown
+# * make ~CudaHandle non-throwing: cu::Worker is detached and holds the last
+#   reference to itself, so ~Worker -> ~CudaEvent -> cudaEventDestroy runs on
+#   the worker thread at an arbitrary point during exit, racing the CUDA
+#   runtime's own teardown. A failing destroy there terminated the process.
+set(mlx_patch
+    ${apply_patch} -DPATCH=${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch
+    ${apply_patch_script})
+
+# note: this must be declared (and named `mlx`) before mlx-c is made available
+# so that our pinned/patched version wins over the one mlx-c declares. Because
+# we now pull mlx in first, we are also responsible for the option defaults that
+# mlx-c would normally set for us.
+set(MLX_BUILD_TESTS OFF)
+set(MLX_BUILD_EXAMPLES OFF)
+set(MLX_BUILD_BENCHMARKS OFF)
+set(MLX_BUILD_PYTHON_BINDINGS OFF)
+
+FetchContent_Declare(
+  mlx
+  GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
+  GIT_TAG "v0.32.2"
+  PATCH_COMMAND ${mlx_patch})
+FetchContent_MakeAvailable(mlx)
+
 # mlx-c
+set(MLX_C_BUILD_EXAMPLES OFF)
 FetchContent_Declare(
   mlx-c
   GIT_REPOSITORY "https://github.com/ml-explore/mlx-c.git"
-  GIT_TAG "v0.6.0")
+  GIT_TAG "c74db5307cc8ce122f48d97ef951b30578674e7f")
 FetchContent_MakeAvailable(mlx-c)
 
 # swift-numerics
-set(swift_numerics_patch git apply
-                         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/swift-numerics.patch)
+set(swift_numerics_patch
+    ${apply_patch}
+    -DPATCH=${CMAKE_CURRENT_SOURCE_DIR}/cmake/swift-numerics.patch
+    ${apply_patch_script})
 FetchContent_Declare(
   swift-numerics
   GIT_REPOSITORY "https://github.com/apple/swift-numerics.git"
   GIT_TAG "1.0.2"
-  PATCH_COMMAND ${swift_numerics_patch} || true)
+  PATCH_COMMAND ${swift_numerics_patch})
 FetchContent_MakeAvailable(swift-numerics)
 
 # MLX package
@@ -64,6 +106,10 @@ if(NOT MLX_BUILD_METAL)
        ${CMAKE_CURRENT_LIST_DIR}/Source/MLX/MLXArray+Metal.swift)
 endif()
 
+if(NOT MLX_BUILD_CUDA)
+  list(REMOVE_ITEM MLX-src ${CMAKE_CURRENT_LIST_DIR}/Source/MLX/GPU+CUDA.swift)
+endif()
+
 add_library(MLX STATIC ${MLX-src})
 target_include_directories(MLX
                            PUBLIC ${CMAKE_CURRENT_LIST_DIR}/Source/Cmlx/include)
@@ -103,7 +149,7 @@ add_library(MLXLinalg STATIC ${MLXLinalg-src})
 target_link_libraries(MLXLinalg PRIVATE MLX)
 
 # Examples
-if(MLX_BUILD_EXAMPLES)
+if(MLX_SWIFT_BUILD_EXAMPLES)
   add_executable(example1
                  ${CMAKE_CURRENT_LIST_DIR}/Source/Examples/Example1.swift)
   target_link_libraries(example1 PRIVATE MLX)

MAINTENANCE.md

@@ -91,12 +91,7 @@ dependencies: [.product(name: "MLX", package: "mlx-swift"),
                .product(name: "MLXOptimizers", package: "mlx-swift")]
 ```
 
-10. Update `tools/generate_integration_tests.py` as needed
-
-```
-import MLXNN
-@testable import MLXOptimizers
-```
+10. Update `tools/integration_tests/cases.py` as needed, regenerate tests if needed
 
 11. Update tests as needed