[slimtensor] Add CUDA Storage with DeviceTraits and memory allocation

backends/aoti/slim/c10/cuda/Exception.h

@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#ifdef CUDA_AVAILABLE
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <executorch/backends/aoti/slim/c10/macros/Macros.h>
+#include <executorch/runtime/platform/assert.h>
+#include <executorch/runtime/platform/log.h>
+
+/// Checks a CUDA expression and aborts on error.
+/// @param EXPR The CUDA expression to check.
+#define ET_CUDA_CHECK(EXPR)                                                 \
+  do {                                                                      \
+    const cudaError_t __err = EXPR;                                         \
+    ET_CHECK_MSG(                                                           \
+        __err == cudaSuccess, "CUDA error: %s", cudaGetErrorString(__err)); \
+  } while (0)
+
+/// Checks a CUDA expression and logs a warning on error (non-fatal).
+/// @param EXPR The CUDA expression to check.
+#define ET_CUDA_LOG_WARN(EXPR)                                      \
+  do {                                                              \
+    const cudaError_t __err = EXPR;                                 \
+    if (SLIMTENSOR_UNLIKELY(__err != cudaSuccess)) {                \
+      [[maybe_unused]] auto error_unused = cudaGetLastError();      \
+      ET_LOG(Error, "CUDA warning: %s", cudaGetErrorString(__err)); \
+    }                                                               \
+  } while (0)
+
+#endif // CUDA_AVAILABLE

backends/aoti/slim/c10/cuda/TARGETS

@@ -0,0 +1,6 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+load(":targets.bzl", "define_common_targets")
+
+oncall("executorch")
+
+define_common_targets()

backends/aoti/slim/c10/cuda/targets.bzl

@@ -0,0 +1,16 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+def define_common_targets():
+    """Define targets for SlimTensor CUDA exception handling module."""
+
+    runtime.cxx_library(
+        name = "exception",
+        exported_headers = [
+            "Exception.h",
+        ],
+        visibility = ["@EXECUTORCH_CLIENTS"],
+        exported_deps = [
+            "//executorch/backends/aoti/slim/c10/macros:macros",
+            "//executorch/runtime/platform:platform",
+        ],
+    )

backends/aoti/slim/core/Storage.h

@@ -10,12 +10,18 @@
 
 #include <cstring>
 
+#ifdef CUDA_AVAILABLE
+#include <executorch/backends/aoti/slim/c10/cuda/Exception.h>
+#include <executorch/backends/aoti/slim/cuda/Guard.h>
+#endif
+
 #include <executorch/backends/aoti/slim/c10/core/Device.h>
 #include <executorch/backends/aoti/slim/c10/core/ScalarType.h>
 #include <executorch/backends/aoti/slim/util/ArrayRefUtil.h>
 #include <executorch/backends/aoti/slim/util/SharedPtr.h>
 #include <executorch/backends/aoti/slim/util/SizeUtil.h>
 #include <executorch/runtime/platform/assert.h>
+#include <executorch/runtime/platform/log.h>
 
 namespace executorch::backends::aoti::slim {
 
@@ -30,6 +36,10 @@ inline void noop(void*) {}
 /// Default CPU device constant.
 inline const c10::Device CPU_DEVICE = c10::Device(c10::DeviceType::CPU, 0);
 
+/// Default CUDA device constant.
+inline const c10::Device DEFAULT_CUDA_DEVICE =
+    c10::Device(c10::DeviceType::CUDA, 0);
+
 /// DeviceTraits template for device-specific operations.
 /// Device-specific implementations provide allocate(), free(), and memcpy().
 template <c10::DeviceType D>
@@ -74,6 +84,93 @@ struct DeviceTraits<c10::DeviceType::CPU> {
   }
 };
 
+#ifdef CUDA_AVAILABLE
+/// CUDA specialization of DeviceTraits.
+/// Provides CUDA memory allocation and copy operations using
+/// cudaMalloc/cudaFree.
+template <>
+struct DeviceTraits<c10::DeviceType::CUDA> {
+  /// Allocates CUDA device memory.
+  /// @param nbytes Number of bytes to allocate.
+  /// @param device The target CUDA device.
+  /// @return Pointer to allocated device memory.
+  static void* allocate(size_t nbytes, const c10::Device& device) {
+    cuda::CUDAGuard guard(device);
+    void* data = nullptr;
+    ET_CUDA_CHECK(cudaMalloc(&data, nbytes));
+    return data;
+  }
+
+  /// Frees CUDA device memory.
+  /// @param ptr Pointer to device memory to free.
+  static void free(void* ptr) {
+    ET_CUDA_LOG_WARN(cudaFree(ptr));
+  }
+
+  /// Copies memory between CPU and CUDA or CUDA and CUDA.
+  /// @param dst Destination pointer.
+  /// @param src Source pointer.
+  /// @param nbytes Number of bytes to copy.
+  /// @param dst_device Destination device.
+  /// @param src_device Source device.
+  static void memcpy(
+      void* dst,
+      const void* src,
+      size_t nbytes,
+      const c10::Device& dst_device,
+      const c10::Device& src_device) {
+    cudaMemcpyKind direction = cudaMemcpyDeviceToDevice;
+    c10::Device cuda_device = dst_device;
+
+    if (src_device.is_cpu()) {
+      direction = cudaMemcpyHostToDevice;
+    } else if (dst_device.is_cpu()) {
+      direction = cudaMemcpyDeviceToHost;
+      cuda_device = src_device;
+    } else {
+      ET_CHECK_MSG(
+          src_device.index() == dst_device.index(),
+          "CUDA memcpy across different device indices not supported: %d != %d",
+          static_cast<int>(src_device.index()),
+          static_cast<int>(dst_device.index()));
+    }
+
+    cuda::CUDAGuard guard(cuda_device);
+    ET_CUDA_CHECK(cudaMemcpy(dst, src, nbytes, direction));
+  }
+};
+#else
+/// CUDA stub when CUDA_AVAILABLE is not defined.
+/// All operations abort with an error message.
+template <>
+struct DeviceTraits<c10::DeviceType::CUDA> {
+  static void* allocate(size_t nbytes, const c10::Device& device) {
+    (void)nbytes;
+    (void)device;
+    ET_CHECK_MSG(false, "Build with CUDA_AVAILABLE=1 to enable CUDA support");
+  }
+
+  static void free(void* ptr) {
+    (void)ptr;
+    ET_LOG(Error, "Build with CUDA_AVAILABLE=1 to enable CUDA support");
+  }
+
+  static void memcpy(
+      void* dst,
+      const void* src,
+      size_t nbytes,
+      const c10::Device& dst_device,
+      const c10::Device& src_device) {
+    (void)dst;
+    (void)src;
+    (void)nbytes;
+    (void)dst_device;
+    (void)src_device;
+    ET_CHECK_MSG(false, "Build with CUDA_AVAILABLE=1 to enable CUDA support");
+  }
+};
+#endif // CUDA_AVAILABLE
+
 /**
  * MaybeOwningStorage - A storage class that manages tensor data memory.
  *
@@ -93,17 +190,19 @@ struct DeviceTraits<c10::DeviceType::CPU> {
 class MaybeOwningStorage {
  public:
   /// Constructs owning storage with allocated memory.
-  /// @param device The device for storage (must be CPU).
+  /// @param device The device for storage (CPU or CUDA).
   /// @param nbytes Number of bytes to allocate.
   MaybeOwningStorage(const c10::Device& device, size_t nbytes)
       : device_(device), capacity_(nbytes), is_owning_(true) {
-    ET_CHECK_MSG(
-        device.is_cpu(),
-        "Only CPU device is currently supported, got: %s",
-        device.str().c_str());
-
-    data_ = DeviceTraits<c10::DeviceType::CPU>::allocate(nbytes, device);
-    deleter_ = DeviceTraits<c10::DeviceType::CPU>::free;
+    if (device.is_cpu()) {
+      data_ = DeviceTraits<c10::DeviceType::CPU>::allocate(nbytes, device);
+      deleter_ = DeviceTraits<c10::DeviceType::CPU>::free;
+    } else if (device.is_cuda()) {
+      data_ = DeviceTraits<c10::DeviceType::CUDA>::allocate(nbytes, device);
+      deleter_ = DeviceTraits<c10::DeviceType::CUDA>::free;
+    } else {
+      ET_CHECK_MSG(false, "Unsupported device type: %s", device.str().c_str());
+    }
   }
 
   /// Default constructor is deleted - storage must have a device.

backends/aoti/slim/core/targets.bzl

@@ -17,10 +17,12 @@ def define_common_targets():
             "//executorch/backends/aoti/slim/util:shared_ptr",
             "//executorch/backends/aoti/slim/util:size_util",
             "//executorch/runtime/platform:platform",
+            "//executorch/backends/aoti/slim/c10/cuda:exception",
+            "//executorch/backends/aoti/slim/cuda:guard",
         ],
     )
 
-    # Header-only library for SlimTensor
+    # Header-only library for SlimTensor (CPU-only for now)
     runtime.cxx_library(
         name = "slimtensor",
         headers = [

backends/aoti/slim/core/test/targets.bzl

@@ -1,17 +1,36 @@
+load("@fbcode_macros//build_defs/lib:re_test_utils.bzl", "re_test_utils")
 load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
 
+def get_backend_mode():
+    """Get the supported backend mode of slimtensor."""
+    return ["cuda", "cpu"]
+
 def define_common_targets():
     """Define test targets for SlimTensor core module."""
 
-    runtime.cxx_test(
-        name = "test_storage",
-        srcs = [
-            "test_storage.cpp",
-        ],
-        deps = [
-            "//executorch/backends/aoti/slim/core:storage",
-        ],
-    )
+    # GPU storage test with CUDA support
+    for backend_mode in get_backend_mode():
+        backend_suffix = "_" + backend_mode if backend_mode == "cuda" else ""
+
+        backend_kwargs = {
+            "external_deps": [("cuda", None, "cuda-lazy")],
+            "preprocessor_flags": ["-DCUDA_AVAILABLE=1"],
+            "keep_gpu_sections": True,
+            "remote_execution": re_test_utils.remote_execution(
+                platform = "gpu-remote-execution",
+            ),
+        } if backend_mode == "cuda" else {}
+
+        runtime.cxx_test(
+            name = "test_storage" + backend_suffix,
+            srcs = [
+                "test_storage.cpp",
+            ],
+            deps = [
+                "//executorch/backends/aoti/slim/core:storage",
+            ],
+            **backend_kwargs
+        )
 
     runtime.cxx_test(
         name = "test_slimtensor_basic",