From e1280571d3f8110d07b924d1da15bddae5b18c83 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 11:27:28 -0700
Subject: [PATCH 01/22] patch for Cuda builds

- local apply of https://github.com/ml-explore/mlx/pull/4480
- fixes crash on exit for cuda builds when using cmake
---
 .github/scripts/build-linux-cuda-cmake.sh |  5 ++---
 CMakeLists.txt                            | 13 +++++++++++++
 cmake/mlx.patch                           | 17 +++++++++++++++++
 3 files changed, 32 insertions(+), 3 deletions(-)
 create mode 100644 cmake/mlx.patch

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index 2a571ea6d..c84e02f27 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -17,7 +17,6 @@ mkdir -p build
 pushd build
 cmake -DMLX_BUILD_METAL=OFF -DMLX_BUILD_CUDA=ON -DMLX_C_BUILD_EXAMPLES=OFF .. -G Ninja
 ninja
-# TODO dkoski -- disabled for now until clear_streams is available
-# ./example1 --device gpu
-# ./tutorial --device gpu
+./example1 --device gpu
+./tutorial --device gpu
 popd
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 41404a01d..8b3fbcb41 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -26,6 +26,19 @@ if(POLICY CMP0135)
   cmake_policy(SET CMP0135 NEW)
 endif()
 
+# mlx
+
+# local apply of https://github.com/ml-explore/mlx/pull/4480 for CUDA
+set(mlx_patch git apply
+                         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch)
+
+FetchContent_Declare(
+  mlx-c
+  GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
+  GIT_TAG "1f8e74e3f12f31365464a6867c6579f0e9b29d85")
+  PATCH_COMMAND ${mlx_patch} || true)
+FetchContent_MakeAvailable(mlx)
+
 # mlx-c
 FetchContent_Declare(
   mlx-c
diff --git a/cmake/mlx.patch b/cmake/mlx.patch
new file mode 100644
index 000000000..384d7dd3f
--- /dev/null
+++ b/cmake/mlx.patch
@@ -0,0 +1,17 @@
+diff --git a/mlx/backend/cuda/device.cpp b/mlx/backend/cuda/device.cpp
+index 472b3d99..e7d8f262 100644
+--- a/mlx/backend/cuda/device.cpp
++++ b/mlx/backend/cuda/device.cpp
+@@ -617,8 +617,10 @@ std::unordered_map<int, CommandEncoder>& get_command_encoders() {
+ }
+ 
+ std::unordered_map<int, CommandEncoder>& get_global_command_encoders() {
+-  static std::unordered_map<int, CommandEncoder> encoders;
+-  return encoders;
++  // encoders are leaked intentionally as they would synchronize on process
++  // shutdown
++  static auto* encoders = new std::unordered_map<int, CommandEncoder>();
++  return *encoders;
+ }
+ 
+ } // namespace mlx::core::cu

From 5adc4d386086fa0af2f98c417f7203010e523ec2 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 14:24:04 -0700
Subject: [PATCH 02/22] fix cmake file

---
 CMakeLists.txt | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8b3fbcb41..9829d23ce 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -29,13 +29,12 @@ endif()
 # mlx
 
 # local apply of https://github.com/ml-explore/mlx/pull/4480 for CUDA
-set(mlx_patch git apply
-                         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch)
+set(mlx_patch git apply ${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch)
 
 FetchContent_Declare(
   mlx-c
   GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
-  GIT_TAG "1f8e74e3f12f31365464a6867c6579f0e9b29d85")
+  GIT_TAG "1f8e74e3f12f31365464a6867c6579f0e9b29d85"
   PATCH_COMMAND ${mlx_patch} || true)
 FetchContent_MakeAvailable(mlx)
 

From ae88266662ade658e414e0ccce1283d67d1ceb20 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 14:52:05 -0700
Subject: [PATCH 03/22] fix cmake issue

---
 CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9829d23ce..d2822d362 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -32,7 +32,7 @@ endif()
 set(mlx_patch git apply ${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch)
 
 FetchContent_Declare(
-  mlx-c
+  mlx
   GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
   GIT_TAG "1f8e74e3f12f31365464a6867c6579f0e9b29d85"
   PATCH_COMMAND ${mlx_patch} || true)

From 586ebb42f0bce6feb99db382c76a4f29bd611e55 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 17:51:05 -0700
Subject: [PATCH 04/22] fix cmake build, part n

---
 CMakeLists.txt | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index d2822d362..61b4af6a3 100644
--- a/CMakeLists.txt
+++ b/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)
 
@@ -31,6 +33,15 @@ endif()
 # local apply of https://github.com/ml-explore/mlx/pull/4480 for CUDA
 set(mlx_patch git apply ${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch)
 
+# 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"
@@ -39,6 +50,7 @@ FetchContent_Declare(
 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"
@@ -119,7 +131,7 @@ 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)

From 8972311e206d1ccabf16fdb2fb83df876d21f00a Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 19:38:49 -0700
Subject: [PATCH 05/22] debug cuda run

---
 .github/scripts/build-linux-cuda-cmake.sh | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index c84e02f27..3edbd44b7 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -16,6 +16,11 @@ rm -rf build
 mkdir -p build
 pushd build
 cmake -DMLX_BUILD_METAL=OFF -DMLX_BUILD_CUDA=ON -DMLX_C_BUILD_EXAMPLES=OFF .. -G Ninja
+
+cd _deps/mlx-src
+git diff
+cd ../..
+
 ninja
 ./example1 --device gpu
 ./tutorial --device gpu

From d69ddd674fd426269362db3cc561f8fa7b069f82 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 20:02:13 -0700
Subject: [PATCH 06/22] Revert "debug cuda run"

This reverts commit 8972311e206d1ccabf16fdb2fb83df876d21f00a.
---
 .github/scripts/build-linux-cuda-cmake.sh | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index 3edbd44b7..c84e02f27 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -16,11 +16,6 @@ rm -rf build
 mkdir -p build
 pushd build
 cmake -DMLX_BUILD_METAL=OFF -DMLX_BUILD_CUDA=ON -DMLX_C_BUILD_EXAMPLES=OFF .. -G Ninja
-
-cd _deps/mlx-src
-git diff
-cd ../..
-
 ninja
 ./example1 --device gpu
 ./tutorial --device gpu

From 4521f0051f4bde97f9f73bdc03967bd1d858d062 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 20:25:28 -0700
Subject: [PATCH 07/22] Reapply "debug cuda run"

This reverts commit d69ddd674fd426269362db3cc561f8fa7b069f82.
---
 .github/scripts/build-linux-cuda-cmake.sh | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index c84e02f27..3edbd44b7 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -16,6 +16,11 @@ rm -rf build
 mkdir -p build
 pushd build
 cmake -DMLX_BUILD_METAL=OFF -DMLX_BUILD_CUDA=ON -DMLX_C_BUILD_EXAMPLES=OFF .. -G Ninja
+
+cd _deps/mlx-src
+git diff
+cd ../..
+
 ninja
 ./example1 --device gpu
 ./tutorial --device gpu

From 0e8039164ce820730f2348f1530c83161b2ad6fc Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:09:02 -0700
Subject: [PATCH 08/22] try and clean up the patches

---
 .github/scripts/build-linux-cuda-cmake.sh |  9 ++++++---
 CMakeLists.txt                            | 21 ++++++++++++++++-----
 2 files changed, 22 insertions(+), 8 deletions(-)

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index 3edbd44b7..2fc81d589 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -17,9 +17,12 @@ mkdir -p build
 pushd build
 cmake -DMLX_BUILD_METAL=OFF -DMLX_BUILD_CUDA=ON -DMLX_C_BUILD_EXAMPLES=OFF .. -G Ninja
 
-cd _deps/mlx-src
-git diff
-cd ../..
+# 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
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 61b4af6a3..8f4c0ce5d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -30,8 +30,17 @@ 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)
+
 # local apply of https://github.com/ml-explore/mlx/pull/4480 for CUDA
-set(mlx_patch git apply ${CMAKE_CURRENT_SOURCE_DIR}/cmake/mlx.patch)
+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
@@ -46,7 +55,7 @@ FetchContent_Declare(
   mlx
   GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
   GIT_TAG "1f8e74e3f12f31365464a6867c6579f0e9b29d85"
-  PATCH_COMMAND ${mlx_patch} || true)
+  PATCH_COMMAND ${mlx_patch})
 FetchContent_MakeAvailable(mlx)
 
 # mlx-c
@@ -58,13 +67,15 @@ FetchContent_Declare(
 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

From 0f003b6aa22815a20da71f5ca9e45e800acf5474 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:32:46 -0700
Subject: [PATCH 09/22] add missing files

---
 cmake/apply-patch.cmake | 68 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 68 insertions(+)
 create mode 100644 cmake/apply-patch.cmake

diff --git a/cmake/apply-patch.cmake b/cmake/apply-patch.cmake
new file mode 100644
index 000000000..06e667ce9
--- /dev/null
+++ b/cmake/apply-patch.cmake
@@ -0,0 +1,68 @@
+# Idempotent `git apply` for FetchContent PATCH_COMMAND.
+#
+# Usage (from a FetchContent_Declare PATCH_COMMAND):
+#
+# PATCH_COMMAND ${CMAKE_COMMAND} -DREPO=<SOURCE_DIR>
+# -DPATCH=${CMAKE_CURRENT_SOURCE_DIR}/cmake/foo.patch -P
+# ${CMAKE_CURRENT_SOURCE_DIR}/cmake/apply-patch.cmake
+#
+# Unlike `git apply ... || true` this does not swallow real failures: an
+# already-applied patch is a no-op (the patch step re-runs on every update), but
+# a patch that fails to apply is a hard error instead of a silently unpatched
+# dependency.
+
+cmake_minimum_required(VERSION 3.16)
+
+if(NOT DEFINED PATCH OR PATCH STREQUAL "")
+  message(FATAL_ERROR "apply-patch.cmake: -DPATCH=<file> is required")
+endif()
+
+if(NOT DEFINED REPO
+   OR REPO STREQUAL ""
+   OR NOT IS_DIRECTORY "${REPO}")
+  # the patch step runs with the dependency source dir as its working directory
+  set(REPO "${CMAKE_CURRENT_BINARY_DIR}")
+endif()
+
+if(NOT EXISTS "${PATCH}")
+  message(FATAL_ERROR "apply-patch.cmake: no such patch file: ${PATCH}")
+endif()
+
+# note: deliberately not `find_package(Git REQUIRED)` -- that pulls in
+# FindPackageHandleStandardArgs in script mode for no benefit here.
+if(NOT DEFINED GIT_EXECUTABLE OR GIT_EXECUTABLE STREQUAL "")
+  find_program(GIT_EXECUTABLE NAMES git git.exe)
+endif()
+
+if(NOT GIT_EXECUTABLE)
+  message(FATAL_ERROR "apply-patch.cmake: git not found (pass -DGIT_EXECUTABLE=)")
+endif()
+
+message(STATUS "apply-patch: repo=${REPO} patch=${PATCH} git=${GIT_EXECUTABLE}")
+
+# already applied? (the patch step re-runs whenever the update step re-runs)
+execute_process(
+  COMMAND ${GIT_EXECUTABLE} apply --reverse --check "${PATCH}"
+  WORKING_DIRECTORY "${REPO}"
+  RESULT_VARIABLE reverse_check
+  OUTPUT_QUIET ERROR_QUIET)
+
+if(reverse_check EQUAL 0)
+  message(STATUS "apply-patch: already applied: ${PATCH}")
+  return()
+endif()
+
+# note: git's stdout/stderr are intentionally *not* captured so that the real
+# reason a patch failed always reaches the build log, even if this script dies
+# before it can format a message of its own.
+execute_process(
+  COMMAND ${GIT_EXECUTABLE} apply --verbose "${PATCH}"
+  WORKING_DIRECTORY "${REPO}"
+  RESULT_VARIABLE apply_result)
+
+if(NOT apply_result EQUAL 0)
+  message(FATAL_ERROR "apply-patch: failed to apply ${PATCH} in ${REPO} "
+                      "(git apply exit ${apply_result}, see output above)")
+endif()
+
+message(STATUS "apply-patch: applied ${PATCH}")

From 9c26f057b94ceb324049bda078e892395c4b51bb Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:35:16 -0700
Subject: [PATCH 10/22] need git

---
 .github/workflows/pull_request.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/pull_request.yml b/.github/workflows/pull_request.yml
index 56452955c..aa7577636 100644
--- a/.github/workflows/pull_request.yml
+++ b/.github/workflows/pull_request.yml
@@ -89,7 +89,7 @@ jobs:
       - name: Setup cmake
         shell: sh
         run: |
-          brew install cmake ninja
+          brew install cmake ninja git
 
       - name: Verify documentation
         run: scripts/verify-docs.sh

From 72356911c6d65fa7c5954aae0c913747575eed30 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:40:33 -0700
Subject: [PATCH 11/22] need git

---
 .github/workflows/pull_request.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/pull_request.yml b/.github/workflows/pull_request.yml
index aa7577636..a805ea144 100644
--- a/.github/workflows/pull_request.yml
+++ b/.github/workflows/pull_request.yml
@@ -24,7 +24,7 @@ jobs:
       - name: Setup pre-commit
         shell: sh
         run: |
-          uv pip install pre-commit
+          uv pip install pre-commit git
 
       - name: Get swift-format tag
         id: swift-format

From 973a6fe336357cb8fcd2a4807e27d342ce2ab36b Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:42:34 -0700
Subject: [PATCH 12/22] need git

---
 .github/workflows/pull_request.yml | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/pull_request.yml b/.github/workflows/pull_request.yml
index a805ea144..004a22fff 100644
--- a/.github/workflows/pull_request.yml
+++ b/.github/workflows/pull_request.yml
@@ -21,10 +21,15 @@ jobs:
         with:
           activate-environment: true
 
+      - name: Setup cmake
+        shell: sh
+        run: |
+          brew install cmake ninja git
+
       - name: Setup pre-commit
         shell: sh
         run: |
-          uv pip install pre-commit git
+          uv pip install pre-commit
 
       - name: Get swift-format tag
         id: swift-format

From 071270f236cdfd920d24dbd364a9c2b513500db7 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:45:55 -0700
Subject: [PATCH 13/22] git should be there actually

---
 .github/workflows/pull_request.yml | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/.github/workflows/pull_request.yml b/.github/workflows/pull_request.yml
index 004a22fff..56452955c 100644
--- a/.github/workflows/pull_request.yml
+++ b/.github/workflows/pull_request.yml
@@ -21,11 +21,6 @@ jobs:
         with:
           activate-environment: true
 
-      - name: Setup cmake
-        shell: sh
-        run: |
-          brew install cmake ninja git
-
       - name: Setup pre-commit
         shell: sh
         run: |
@@ -94,7 +89,7 @@ jobs:
       - name: Setup cmake
         shell: sh
         run: |
-          brew install cmake ninja git
+          brew install cmake ninja
 
       - name: Verify documentation
         run: scripts/verify-docs.sh

From ce7e09e714e1fad48fe2ebd8897b1a2ec05de7f8 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 21:48:52 -0700
Subject: [PATCH 14/22] haha, it was cmake format, not git

---
 cmake/apply-patch.cmake | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/cmake/apply-patch.cmake b/cmake/apply-patch.cmake
index 06e667ce9..63e04f83f 100644
--- a/cmake/apply-patch.cmake
+++ b/cmake/apply-patch.cmake
@@ -35,7 +35,8 @@ if(NOT DEFINED GIT_EXECUTABLE OR GIT_EXECUTABLE STREQUAL "")
 endif()
 
 if(NOT GIT_EXECUTABLE)
-  message(FATAL_ERROR "apply-patch.cmake: git not found (pass -DGIT_EXECUTABLE=)")
+  message(
+    FATAL_ERROR "apply-patch.cmake: git not found (pass -DGIT_EXECUTABLE=)")
 endif()
 
 message(STATUS "apply-patch: repo=${REPO} patch=${PATCH} git=${GIT_EXECUTABLE}")

From 77c1ba06b1d959b9bc4317d458816a489c71e36b Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Fri, 11 Sep 2026 22:32:36 -0700
Subject: [PATCH 15/22] try leaking thread local encoders as well

---
 cmake/mlx.patch | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)

diff --git a/cmake/mlx.patch b/cmake/mlx.patch
index 384d7dd3f..b1fc2c697 100644
--- a/cmake/mlx.patch
+++ b/cmake/mlx.patch
@@ -1,15 +1,28 @@
 diff --git a/mlx/backend/cuda/device.cpp b/mlx/backend/cuda/device.cpp
-index 472b3d99..e7d8f262 100644
+index 472b3d99..b310b47a 100644
 --- a/mlx/backend/cuda/device.cpp
 +++ b/mlx/backend/cuda/device.cpp
-@@ -617,8 +617,10 @@ std::unordered_map<int, CommandEncoder>& get_command_encoders() {
+@@ -612,13 +612,21 @@ CommandEncoder& get_command_encoder(Stream s) {
+ }
+ 
+ std::unordered_map<int, CommandEncoder>& get_command_encoders() {
+-  static thread_local std::unordered_map<int, CommandEncoder> encoders;
+-  return encoders;
++  // encoders are leaked intentionally: destroying a CommandEncoder stops its
++  // cu::Worker, and the detached worker thread owns the last reference to
++  // itself, so ~Worker (and cudaEventDestroy) runs on that thread at an
++  // arbitrary point during process shutdown -- typically after the CUDA runtime
++  // has started unloading, which turns into a throw from a destructor and
++  // terminates the process
++  static thread_local auto* encoders =
++      new std::unordered_map<int, CommandEncoder>();
++  return *encoders;
  }
  
  std::unordered_map<int, CommandEncoder>& get_global_command_encoders() {
 -  static std::unordered_map<int, CommandEncoder> encoders;
 -  return encoders;
-+  // encoders are leaked intentionally as they would synchronize on process
-+  // shutdown
++  // leaked for the same reason as above
 +  static auto* encoders = new std::unordered_map<int, CommandEncoder>();
 +  return *encoders;
  }

From 775202f22e58bc12ac20489a321c3db0e24d1db9 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Sat, 12 Sep 2026 07:33:48 -0700
Subject: [PATCH 16/22] debugging the patch

---
 .github/scripts/build-linux-cuda-cmake.sh |  19 ++++
 cmake/mlx.patch                           | 106 +++++++++++++++++++++-
 2 files changed, 120 insertions(+), 5 deletions(-)

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index 2fc81d589..7b3b3e93c 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -25,6 +25,25 @@ git -C _deps/mlx-src apply --reverse --check "$PWD/../cmake/mlx.patch" || {
 }
 
 ninja
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
+./example1 --device gpu
+echo ===================
 ./example1 --device gpu
 ./tutorial --device gpu
 popd
diff --git a/cmake/mlx.patch b/cmake/mlx.patch
index b1fc2c697..51f02b038 100644
--- a/cmake/mlx.patch
+++ b/cmake/mlx.patch
@@ -1,8 +1,56 @@
 diff --git a/mlx/backend/cuda/device.cpp b/mlx/backend/cuda/device.cpp
-index 472b3d99..b310b47a 100644
+index 472b3d99..5995e94b 100644
 --- a/mlx/backend/cuda/device.cpp
 +++ b/mlx/backend/cuda/device.cpp
-@@ -612,13 +612,21 @@ CommandEncoder& get_command_encoder(Stream s) {
+@@ -7,11 +7,27 @@
+ 
+ #include <fmt/format.h>
+ #include <nvtx3/nvtx3.hpp>
++#include <cstdio>
+ #include <future>
++#include <thread>
+ #include <unordered_set>
+ 
+ namespace mlx::core::cu {
+ 
++// TEMP DEBUG (mlx-swift): trace CommandEncoder lifetime so we can tell whether
++// the thread-local or the global (thread-unsafe) encoder map is in use.
++void encoder_debug_log(const char* what, int stream_index, const void* encoder) {
++  fprintf(
++      stderr,
++      "[mlx-swift/encoders] %s stream=%d encoder=%p thread=%zu main=%d\n",
++      what,
++      stream_index,
++      encoder,
++      std::hash<std::thread::id>{}(std::this_thread::get_id()),
++      is_main_thread() ? 1 : 0);
++  fflush(stderr);
++}
++
+ namespace {
+ 
+ bool use_cuda_graphs() {
+@@ -210,9 +226,11 @@ CommandEncoder::CommandEncoder(Device& d)
+       graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {
+   std::tie(max_ops_per_graph_, max_mb_per_graph_) = get_graph_limits(d);
+   worker_->start();
++  encoder_debug_log("CommandEncoder()", -1, this);
+ }
+ 
+ CommandEncoder::~CommandEncoder() {
++  encoder_debug_log("~CommandEncoder", -1, this);
+   synchronize();
+   worker_->stop();
+ }
+@@ -601,6 +619,7 @@ CommandEncoder& get_command_encoder(Stream s) {
+   auto it = encoders.find(s.index);
+   if (it == encoders.end()) {
+     auto& global_encoders = get_global_command_encoders();
++    encoder_debug_log("lookup falls back to global map", s.index, nullptr);
+     it = global_encoders.find(s.index);
+     if (it == global_encoders.end()) {
+       throw std::runtime_error(
+@@ -612,13 +631,26 @@ CommandEncoder& get_command_encoder(Stream s) {
  }
  
  std::unordered_map<int, CommandEncoder>& get_command_encoders() {
@@ -14,8 +62,10 @@ index 472b3d99..b310b47a 100644
 +  // arbitrary point during process shutdown -- typically after the CUDA runtime
 +  // has started unloading, which turns into a throw from a destructor and
 +  // terminates the process
-+  static thread_local auto* encoders =
-+      new std::unordered_map<int, CommandEncoder>();
++  static thread_local auto* encoders = []() {
++    encoder_debug_log("new thread_local encoder map", -1, nullptr);
++    return new std::unordered_map<int, CommandEncoder>();
++  }();
 +  return *encoders;
  }
  
@@ -23,8 +73,54 @@ index 472b3d99..b310b47a 100644
 -  static std::unordered_map<int, CommandEncoder> encoders;
 -  return encoders;
 +  // leaked for the same reason as above
-+  static auto* encoders = new std::unordered_map<int, CommandEncoder>();
++  static auto* encoders = []() {
++    encoder_debug_log("new global encoder map", -1, nullptr);
++    return new std::unordered_map<int, CommandEncoder>();
++  }();
 +  return *encoders;
  }
  
  } // namespace mlx::core::cu
+diff --git a/mlx/backend/cuda/device.h b/mlx/backend/cuda/device.h
+index 198f0b5a..2507202c 100644
+--- a/mlx/backend/cuda/device.h
++++ b/mlx/backend/cuda/device.h
+@@ -214,6 +214,9 @@ MLX_API Device& device(int cuda_device);
+ MLX_API Device& device(mlx::core::Device d);
+ MLX_API CommandEncoder& get_command_encoder(Stream s);
+ 
++// TEMP DEBUG (mlx-swift): see device.cpp
++void encoder_debug_log(const char* what, int stream_index, const void* encoder);
++
+ std::unordered_map<int, CommandEncoder>& get_command_encoders();
+ std::unordered_map<int, CommandEncoder>& get_global_command_encoders();
+ 
+diff --git a/mlx/backend/cuda/eval.cpp b/mlx/backend/cuda/eval.cpp
+index f8a71d71..0503d7c8 100644
+--- a/mlx/backend/cuda/eval.cpp
++++ b/mlx/backend/cuda/eval.cpp
+@@ -26,6 +26,7 @@ void init() {
+ 
+ void new_stream(Stream s) {
+   assert(s.device == Device::gpu);
++  cu::encoder_debug_log("new_stream (thread_local map)", s.index, nullptr);
+   auto& encoders = cu::get_command_encoders();
+   auto& d = cu::device(s.device);
+   encoders.try_emplace(s.index, d);
+@@ -33,6 +34,8 @@ void new_stream(Stream s) {
+ 
+ void new_thread_unsafe_stream(Stream s) {
+   assert(s.device == Device::gpu);
++  cu::encoder_debug_log(
++      "new_thread_unsafe_stream (global map)", s.index, nullptr);
+   auto& encoders = cu::get_global_command_encoders();
+   auto& d = cu::device(s.device);
+   encoders.try_emplace(s.index, d);
+@@ -87,6 +90,7 @@ void synchronize(Stream s) {
+ }
+ 
+ void clear_streams() {
++  cu::encoder_debug_log("clear_streams", -1, nullptr);
+   cu::get_command_encoders().clear();
+   if (is_main_thread()) {
+     cu::get_global_command_encoders().clear();

From ab10a8b3f891c6ee5cc22825d40633138a9621c9 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Sat, 12 Sep 2026 20:39:06 -0700
Subject: [PATCH 17/22] a different approach

---
 cmake/mlx.patch | 151 ++++++++++++------------------------------------
 1 file changed, 37 insertions(+), 114 deletions(-)

diff --git a/cmake/mlx.patch b/cmake/mlx.patch
index 51f02b038..47fb5bdd0 100644
--- a/cmake/mlx.patch
+++ b/cmake/mlx.patch
@@ -1,126 +1,49 @@
+diff --git a/mlx/backend/cuda/cuda_utils.h b/mlx/backend/cuda/cuda_utils.h
+index f8a234ee..cc075f7b 100644
+--- a/mlx/backend/cuda/cuda_utils.h
++++ b/mlx/backend/cuda/cuda_utils.h
+@@ -30,7 +30,22 @@ class CudaHandle {
+     if (cudaPeekAtLastError() != cudaSuccess) {
+       return;
+     }
+-    reset();
++    // Never CHECK here: a destructor must not throw.
++    //
++    // Handles are routinely destroyed at times when the CUDA runtime is no
++    // longer usable. The worst case is cu::Worker: it is started detached and
++    // holds the last shared_ptr to itself, so ~Worker -- and with it
++    // ~CudaEvent -> CudaEventPool::release() -> cudaEventDestroy, since the
++    // pool only recycles events on its own creation thread -- runs on the
++    // worker thread whenever it next wakes up. When that happens during
++    // process shutdown it races the CUDA runtime's own atexit teardown, and
++    // throwing out of this destructor terminates the process instead of
++    // reporting anything useful. Losing a handle we were about to destroy
++    // anyway is strictly better.
++    if (handle_ != nullptr) {
++      Destroy(handle_);
++      handle_ = nullptr;
++    }
+   }
+ 
+   CudaHandle(const CudaHandle&) = delete;
 diff --git a/mlx/backend/cuda/device.cpp b/mlx/backend/cuda/device.cpp
-index 472b3d99..5995e94b 100644
+index 472b3d99..5de6fdf1 100644
 --- a/mlx/backend/cuda/device.cpp
 +++ b/mlx/backend/cuda/device.cpp
-@@ -7,11 +7,27 @@
- 
- #include <fmt/format.h>
- #include <nvtx3/nvtx3.hpp>
-+#include <cstdio>
- #include <future>
-+#include <thread>
- #include <unordered_set>
- 
- namespace mlx::core::cu {
- 
-+// TEMP DEBUG (mlx-swift): trace CommandEncoder lifetime so we can tell whether
-+// the thread-local or the global (thread-unsafe) encoder map is in use.
-+void encoder_debug_log(const char* what, int stream_index, const void* encoder) {
-+  fprintf(
-+      stderr,
-+      "[mlx-swift/encoders] %s stream=%d encoder=%p thread=%zu main=%d\n",
-+      what,
-+      stream_index,
-+      encoder,
-+      std::hash<std::thread::id>{}(std::this_thread::get_id()),
-+      is_main_thread() ? 1 : 0);
-+  fflush(stderr);
-+}
-+
- namespace {
- 
- bool use_cuda_graphs() {
-@@ -210,9 +226,11 @@ CommandEncoder::CommandEncoder(Device& d)
-       graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {
-   std::tie(max_ops_per_graph_, max_mb_per_graph_) = get_graph_limits(d);
-   worker_->start();
-+  encoder_debug_log("CommandEncoder()", -1, this);
- }
- 
- CommandEncoder::~CommandEncoder() {
-+  encoder_debug_log("~CommandEncoder", -1, this);
-   synchronize();
-   worker_->stop();
- }
-@@ -601,6 +619,7 @@ CommandEncoder& get_command_encoder(Stream s) {
-   auto it = encoders.find(s.index);
-   if (it == encoders.end()) {
-     auto& global_encoders = get_global_command_encoders();
-+    encoder_debug_log("lookup falls back to global map", s.index, nullptr);
-     it = global_encoders.find(s.index);
-     if (it == global_encoders.end()) {
-       throw std::runtime_error(
-@@ -612,13 +631,26 @@ CommandEncoder& get_command_encoder(Stream s) {
- }
- 
- std::unordered_map<int, CommandEncoder>& get_command_encoders() {
--  static thread_local std::unordered_map<int, CommandEncoder> encoders;
--  return encoders;
-+  // encoders are leaked intentionally: destroying a CommandEncoder stops its
-+  // cu::Worker, and the detached worker thread owns the last reference to
-+  // itself, so ~Worker (and cudaEventDestroy) runs on that thread at an
-+  // arbitrary point during process shutdown -- typically after the CUDA runtime
-+  // has started unloading, which turns into a throw from a destructor and
-+  // terminates the process
-+  static thread_local auto* encoders = []() {
-+    encoder_debug_log("new thread_local encoder map", -1, nullptr);
-+    return new std::unordered_map<int, CommandEncoder>();
-+  }();
-+  return *encoders;
+@@ -617,8 +617,14 @@ std::unordered_map<int, CommandEncoder>& get_command_encoders() {
  }
  
  std::unordered_map<int, CommandEncoder>& get_global_command_encoders() {
 -  static std::unordered_map<int, CommandEncoder> encoders;
 -  return encoders;
-+  // leaked for the same reason as above
-+  static auto* encoders = []() {
-+    encoder_debug_log("new global encoder map", -1, nullptr);
-+    return new std::unordered_map<int, CommandEncoder>();
-+  }();
++  // Leaked intentionally. Unlike the thread-local map above -- which is
++  // destroyed when its thread exits, while CUDA is still fully alive, and which
++  // is the only thing that reclaims per-thread streams/events/worker threads --
++  // this map is a plain static, so its only destruction point is static
++  // destruction during exit(), racing the CUDA runtime teardown. There is
++  // nothing to gain from destroying it there: the process is going away.
++  static auto* encoders = new std::unordered_map<int, CommandEncoder>();
 +  return *encoders;
  }
  
  } // namespace mlx::core::cu
-diff --git a/mlx/backend/cuda/device.h b/mlx/backend/cuda/device.h
-index 198f0b5a..2507202c 100644
---- a/mlx/backend/cuda/device.h
-+++ b/mlx/backend/cuda/device.h
-@@ -214,6 +214,9 @@ MLX_API Device& device(int cuda_device);
- MLX_API Device& device(mlx::core::Device d);
- MLX_API CommandEncoder& get_command_encoder(Stream s);
- 
-+// TEMP DEBUG (mlx-swift): see device.cpp
-+void encoder_debug_log(const char* what, int stream_index, const void* encoder);
-+
- std::unordered_map<int, CommandEncoder>& get_command_encoders();
- std::unordered_map<int, CommandEncoder>& get_global_command_encoders();
- 
-diff --git a/mlx/backend/cuda/eval.cpp b/mlx/backend/cuda/eval.cpp
-index f8a71d71..0503d7c8 100644
---- a/mlx/backend/cuda/eval.cpp
-+++ b/mlx/backend/cuda/eval.cpp
-@@ -26,6 +26,7 @@ void init() {
- 
- void new_stream(Stream s) {
-   assert(s.device == Device::gpu);
-+  cu::encoder_debug_log("new_stream (thread_local map)", s.index, nullptr);
-   auto& encoders = cu::get_command_encoders();
-   auto& d = cu::device(s.device);
-   encoders.try_emplace(s.index, d);
-@@ -33,6 +34,8 @@ void new_stream(Stream s) {
- 
- void new_thread_unsafe_stream(Stream s) {
-   assert(s.device == Device::gpu);
-+  cu::encoder_debug_log(
-+      "new_thread_unsafe_stream (global map)", s.index, nullptr);
-   auto& encoders = cu::get_global_command_encoders();
-   auto& d = cu::device(s.device);
-   encoders.try_emplace(s.index, d);
-@@ -87,6 +90,7 @@ void synchronize(Stream s) {
- }
- 
- void clear_streams() {
-+  cu::encoder_debug_log("clear_streams", -1, nullptr);
-   cu::get_command_encoders().clear();
-   if (is_main_thread()) {
-     cu::get_global_command_encoders().clear();

From 55160e21aca2e9bedbd551b71844563f7c907ac7 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Mon, 14 Sep 2026 09:17:58 -0700
Subject: [PATCH 18/22] clean up changes

---
 CMakeLists.txt  |  9 ++++++++-
 cmake/mlx.patch | 25 +++++--------------------
 2 files changed, 13 insertions(+), 21 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8f4c0ce5d..94ae24abb 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -37,7 +37,14 @@ endif()
 set(apply_patch ${CMAKE_COMMAND} -DREPO=<SOURCE_DIR>)
 set(apply_patch_script -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/apply-patch.cmake)
 
-# local apply of https://github.com/ml-explore/mlx/pull/4480 for CUDA
+# 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})
diff --git a/cmake/mlx.patch b/cmake/mlx.patch
index 47fb5bdd0..dc8f2f7b8 100644
--- a/cmake/mlx.patch
+++ b/cmake/mlx.patch
@@ -7,18 +7,7 @@ index f8a234ee..cc075f7b 100644
        return;
      }
 -    reset();
-+    // Never CHECK here: a destructor must not throw.
-+    //
-+    // Handles are routinely destroyed at times when the CUDA runtime is no
-+    // longer usable. The worst case is cu::Worker: it is started detached and
-+    // holds the last shared_ptr to itself, so ~Worker -- and with it
-+    // ~CudaEvent -> CudaEventPool::release() -> cudaEventDestroy, since the
-+    // pool only recycles events on its own creation thread -- runs on the
-+    // worker thread whenever it next wakes up. When that happens during
-+    // process shutdown it races the CUDA runtime's own atexit teardown, and
-+    // throwing out of this destructor terminates the process instead of
-+    // reporting anything useful. Losing a handle we were about to destroy
-+    // anyway is strictly better.
++    // Inline reset() without throwing -- handle shutdown cases.
 +    if (handle_ != nullptr) {
 +      Destroy(handle_);
 +      handle_ = nullptr;
@@ -27,21 +16,17 @@ index f8a234ee..cc075f7b 100644
  
    CudaHandle(const CudaHandle&) = delete;
 diff --git a/mlx/backend/cuda/device.cpp b/mlx/backend/cuda/device.cpp
-index 472b3d99..5de6fdf1 100644
+index 472b3d99..e7d8f262 100644
 --- a/mlx/backend/cuda/device.cpp
 +++ b/mlx/backend/cuda/device.cpp
-@@ -617,8 +617,14 @@ std::unordered_map<int, CommandEncoder>& get_command_encoders() {
+@@ -617,8 +617,10 @@ std::unordered_map<int, CommandEncoder>& get_command_encoders() {
  }
  
  std::unordered_map<int, CommandEncoder>& get_global_command_encoders() {
 -  static std::unordered_map<int, CommandEncoder> encoders;
 -  return encoders;
-+  // Leaked intentionally. Unlike the thread-local map above -- which is
-+  // destroyed when its thread exits, while CUDA is still fully alive, and which
-+  // is the only thing that reclaims per-thread streams/events/worker threads --
-+  // this map is a plain static, so its only destruction point is static
-+  // destruction during exit(), racing the CUDA runtime teardown. There is
-+  // nothing to gain from destroying it there: the process is going away.
++  // encoders are leaked intentionally as they would synchronize on process
++  // shutdown
 +  static auto* encoders = new std::unordered_map<int, CommandEncoder>();
 +  return *encoders;
  }

From 1d7250eb5afd6db0d47ef6a88e7c3c870b68fe76 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Mon, 14 Sep 2026 09:21:28 -0700
Subject: [PATCH 19/22] cmake-format

---
 CMakeLists.txt | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 94ae24abb..b7390ecb7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -39,12 +39,12 @@ 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
+# * 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.
+# * 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})

From 1f466865c424e00adca7b54e34ccf5571375d515 Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Mon, 14 Sep 2026 09:27:35 -0700
Subject: [PATCH 20/22] fix patch

---
 cmake/mlx.patch | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cmake/mlx.patch b/cmake/mlx.patch
index dc8f2f7b8..de72da7bb 100644
--- a/cmake/mlx.patch
+++ b/cmake/mlx.patch
@@ -2,7 +2,7 @@ diff --git a/mlx/backend/cuda/cuda_utils.h b/mlx/backend/cuda/cuda_utils.h
 index f8a234ee..cc075f7b 100644
 --- a/mlx/backend/cuda/cuda_utils.h
 +++ b/mlx/backend/cuda/cuda_utils.h
-@@ -30,7 +30,22 @@ class CudaHandle {
+@@ -30,6 +30,11 @@ class CudaHandle {
      if (cudaPeekAtLastError() != cudaSuccess) {
        return;
      }

From 44e75efd15152655d3aad92c03cf2c74c7fe228b Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Mon, 14 Sep 2026 09:40:04 -0700
Subject: [PATCH 21/22] fix patch try 2

---
 cmake/mlx.patch | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/cmake/mlx.patch b/cmake/mlx.patch
index de72da7bb..de99b7ebc 100644
--- a/cmake/mlx.patch
+++ b/cmake/mlx.patch
@@ -1,8 +1,8 @@
 diff --git a/mlx/backend/cuda/cuda_utils.h b/mlx/backend/cuda/cuda_utils.h
-index f8a234ee..cc075f7b 100644
+index f8a234ee6..3b7a79f27 100644
 --- a/mlx/backend/cuda/cuda_utils.h
 +++ b/mlx/backend/cuda/cuda_utils.h
-@@ -30,6 +30,11 @@ class CudaHandle {
+@@ -30,7 +30,11 @@ class CudaHandle {
      if (cudaPeekAtLastError() != cudaSuccess) {
        return;
      }
@@ -16,7 +16,7 @@ index f8a234ee..cc075f7b 100644
  
    CudaHandle(const CudaHandle&) = delete;
 diff --git a/mlx/backend/cuda/device.cpp b/mlx/backend/cuda/device.cpp
-index 472b3d99..e7d8f262 100644
+index 472b3d99f..e7d8f2620 100644
 --- a/mlx/backend/cuda/device.cpp
 +++ b/mlx/backend/cuda/device.cpp
 @@ -617,8 +617,10 @@ std::unordered_map<int, CommandEncoder>& get_command_encoders() {

From c13b5c5fe643f3f3ad4cb263da89812aa791f3ea Mon Sep 17 00:00:00 2001
From: David Koski <dkoski@apple.com>
Date: Tue, 15 Sep 2026 15:18:15 -0700
Subject: [PATCH 22/22] remove debugging code

---
 .github/scripts/build-linux-cuda-cmake.sh | 19 -------------------
 CMakeLists.txt                            |  2 +-
 2 files changed, 1 insertion(+), 20 deletions(-)

diff --git a/.github/scripts/build-linux-cuda-cmake.sh b/.github/scripts/build-linux-cuda-cmake.sh
index 7b3b3e93c..2fc81d589 100755
--- a/.github/scripts/build-linux-cuda-cmake.sh
+++ b/.github/scripts/build-linux-cuda-cmake.sh
@@ -25,25 +25,6 @@ git -C _deps/mlx-src apply --reverse --check "$PWD/../cmake/mlx.patch" || {
 }
 
 ninja
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
-./example1 --device gpu
-echo ===================
 ./example1 --device gpu
 ./tutorial --device gpu
 popd
diff --git a/CMakeLists.txt b/CMakeLists.txt
index b7390ecb7..22f4b16ff 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -61,7 +61,7 @@ set(MLX_BUILD_PYTHON_BINDINGS OFF)
 FetchContent_Declare(
   mlx
   GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
-  GIT_TAG "1f8e74e3f12f31365464a6867c6579f0e9b29d85"
+  GIT_TAG "v0.32.2"
   PATCH_COMMAND ${mlx_patch})
 FetchContent_MakeAvailable(mlx)