fix bug when kernels upload (#242)

Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com>
Co-authored-by: Liu, Kaixuan <kaixuan.liu@intel.com>

Author: danieldk

build2cmake/src/templates/windows.cmake

@@ -140,6 +140,8 @@ endfunction()
 function(add_local_install_target TARGET_NAME PACKAGE_NAME BUILD_VARIANT_NAME)
     # Define your local, folder based, installation directory
     set(LOCAL_INSTALL_DIR "${CMAKE_SOURCE_DIR}/build/${BUILD_VARIANT_NAME}/${PACKAGE_NAME}")
+    # Variant directory is where metadata.json should go (for kernels upload discovery)
+    set(VARIANT_DIR "${CMAKE_SOURCE_DIR}/build/${BUILD_VARIANT_NAME}")
 
     # Glob Python files at configure time
     file(GLOB PYTHON_FILES "${CMAKE_SOURCE_DIR}/torch-ext/${PACKAGE_NAME}/*.py")
@@ -164,12 +166,14 @@ function(add_local_install_target TARGET_NAME PACKAGE_NAME BUILD_VARIANT_NAME)
             # Copy metadata.json if it exists
             COMMAND ${CMAKE_COMMAND} -E copy_if_different
             ${CMAKE_SOURCE_DIR}/metadata.json
-            ${LOCAL_INSTALL_DIR}/
+            ${VARIANT_DIR}/
 
             COMMENT "Copying shared library and Python files to ${LOCAL_INSTALL_DIR}"
             COMMAND_EXPAND_LISTS
     )
 
+    # Create both directories: variant dir for metadata.json, package dir for binaries
+    file(MAKE_DIRECTORY ${VARIANT_DIR})
     file(MAKE_DIRECTORY ${LOCAL_INSTALL_DIR})
     message(STATUS "Added install rules for ${TARGET_NAME} -> build/${BUILD_VARIANT_NAME}/${PACKAGE_NAME}")
 endfunction()

build2cmake/src/templates/xpu/preamble.cmake

@@ -80,3 +80,11 @@ add_compile_definitions(USE_XPU)
 set(sycl_link_flags "-fsycl;--offload-compress;-fsycl-targets=spir64_gen,spir64;-Xs;-device pvc,xe-lpg,ats-m150 -options ' -cl-intel-enable-auto-large-GRF-mode -cl-poison-unsupported-fp64-kernels -cl-intel-greater-than-4GB-buffer-required';")
 set(sycl_flags "-fsycl;-fhonor-nans;-fhonor-infinities;-fno-associative-math;-fno-approx-func;-fno-sycl-instrument-device-code;--offload-compress;-fsycl-targets=spir64_gen,spir64;")
 message(STATUS "Configuring for Intel XPU backend using SYCL")
+
+{% if platform == 'windows' %}
+# Include Windows-specific functions for local_install and kernels_install targets
+include(${CMAKE_CURRENT_LIST_DIR}/cmake/windows.cmake)
+
+# Generate build variant name for XPU (e.g., torch291-xpu-x86_64-windows)
+generate_build_name(BUILD_VARIANT_NAME "${TORCH_VERSION}" "xpu" "${DPCPP_VERSION}")
+{% endif %}

build2cmake/src/templates/xpu/torch-extension.cmake

@@ -10,3 +10,14 @@ define_gpu_extension_target(
 # Add XPU/SYCL specific linker flags
 target_link_options({{ ops_name }} PRIVATE ${sycl_link_flags})
 target_link_libraries({{ ops_name }} PRIVATE dnnl)
+
+{% if platform == 'windows' %}
+# These methods below should be included from preamble.cmake on windows platform.
+
+# Add kernels_install target for huggingface/kernels library layout
+add_kernels_install_target({{ ops_name }} "{{ name }}" "${BUILD_VARIANT_NAME}")
+
+# Add local_install target for local development with get_local_kernel()
+add_local_install_target({{ ops_name }} "{{ name }}" "${BUILD_VARIANT_NAME}")
+
+{% endif %}

build2cmake/src/torch/xpu.rs

@@ -15,6 +15,7 @@ use crate::FileSet;
 
 static CMAKE_UTILS: &str = include_str!("../templates/utils.cmake");
 static REGISTRATION_H: &str = include_str!("../templates/registration.h");
+static WINDOWS_UTILS: &str = include_str!("../templates/windows.cmake");
 
 pub fn write_torch_ext_xpu(
     env: &Environment,
@@ -137,6 +138,14 @@ fn write_cmake(
         .entry(utils_path.clone())
         .extend_from_slice(CMAKE_UTILS.as_bytes());
 
+    // Add windows.cmake for Windows-specific install targets
+    let mut windows_utils_path = PathBuf::new();
+    windows_utils_path.push("cmake");
+    windows_utils_path.push("windows.cmake");
+    file_set
+        .entry(windows_utils_path.clone())
+        .extend_from_slice(WINDOWS_UTILS.as_bytes());
+
     let cmake_writer = file_set.entry("CMakeLists.txt");
 
     render_preamble(
@@ -296,6 +305,7 @@ pub fn render_preamble(
                 name => name,
                 torch_minver => torch_minver.map(|v| v.to_string()),
                 torch_maxver => torch_maxver.map(|v| v.to_string()),
+                platform => std::env::consts::OS,
             },
             &mut *write,
         )