start cuda circle config (#2256)

* rebase

* fix metal kernel linking issue on cuda

* start cuda circle config

mlx/CMakeLists.txt

@@ -55,6 +55,9 @@ endif()
 
 if(MLX_BUILD_CUDA)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda)
+else()
+  target_sources(mlx
+                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda/no_cuda.cpp)
 endif()
 
 if(MLX_BUILD_METAL OR MLX_BUILD_CUDA)

mlx/backend/cuda/CMakeLists.txt

@@ -12,6 +12,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_dynamic.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_input.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu

mlx/backend/cuda/cuda.cpp

@@ -0,0 +1,11 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cuda.h"
+
+namespace mlx::core::cu {
+
+bool is_available() {
+  return true;
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/cuda.h

@@ -0,0 +1,10 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+namespace mlx::core::cu {
+
+/* Check if the CUDA backend is available. */
+bool is_available();
+
+} // namespace mlx::core::cu

mlx/backend/cuda/no_cuda.cpp

@@ -0,0 +1,11 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cuda.h"
+
+namespace mlx::core::cu {
+
+bool is_available() {
+  return false;
+}
+
+} // namespace mlx::core::cu

mlx/backend/metal/no_metal.cpp

@@ -3,8 +3,11 @@
 #include <stdexcept>
 
 #include "mlx/backend/metal/metal.h"
+#include "mlx/fast.h"
 
-namespace mlx::core::metal {
+namespace mlx::core {
+
+namespace metal {
 
 bool is_available() {
   return false;
@@ -19,4 +22,21 @@ device_info() {
       "[metal::device_info] Cannot get device info without metal backend");
 };
 
-} // namespace mlx::core::metal
+} // namespace metal
+
+namespace fast {
+
+MetalKernelFunction metal_kernel(
+    const std::string&,
+    const std::vector<std::string>&,
+    const std::vector<std::string>&,
+    const std::string&,
+    const std::string&,
+    bool ensure_row_contiguous,
+    bool atomic_outputs) {
+  throw std::runtime_error("[metal_kernel] No GPU back-end.");
+}
+
+} // namespace fast
+
+} // namespace mlx::core

mlx/backend/no_gpu/primitives.cpp

@@ -2,7 +2,6 @@
 
 #include "mlx/primitives.h"
 #include "mlx/distributed/primitives.h"
-#include "mlx/fast.h"
 #include "mlx/fast_primitives.h"
 
 #define NO_GPU_MULTI(func)                                             \
@@ -156,18 +155,6 @@ NO_GPU_USE_FALLBACK(RoPE)
 NO_GPU(ScaledDotProductAttention)
 NO_GPU_MULTI(AffineQuantize)
 NO_GPU_MULTI(CustomKernel)
-
-MetalKernelFunction metal_kernel(
-    const std::string&,
-    const std::vector<std::string>&,
-    const std::vector<std::string>&,
-    const std::string&,
-    const std::string&,
-    bool ensure_row_contiguous,
-    bool atomic_outputs) {
-  throw std::runtime_error("[metal_kernel] No GPU back-end.");
-}
-
 } // namespace fast
 
 namespace distributed {

mlx/mlx.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include "mlx/array.h"
+#include "mlx/backend/cuda/cuda.h"
 #include "mlx/backend/metal/metal.h"
 #include "mlx/compile.h"
 #include "mlx/device.h"