[Build] Fixes for CMake install (#18570)

CMakeLists.txt

@@ -23,6 +23,9 @@ include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 # Suppress potential warnings about unused manually-specified variables
 set(ignoreMe "${VLLM_PYTHON_PATH}")
 
+# Prevent installation of dependencies (cutlass) by default.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
@@ -785,5 +788,7 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
+
+    # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()

cmake/external_projects/vllm_flash_attn.cmake

@@ -46,22 +46,38 @@ else()
 endif()
 
 
+# Ensure the vllm/vllm_flash_attn directory exists before installation
+install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" ALL_COMPONENTS)
+
+# Make sure vllm-flash-attn install rules are nested under vllm/
+# This is here to support installing all components under the same prefix with cmake --install.
+# setup.py installs every component separately but uses the same prefix for all.
+# ALL_COMPONENTS is used to avoid duplication for FA2 and FA3,
+# and these statements don't hurt when installing neither component.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" ALL_COMPONENTS)
+install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_COMPONENTS)
+
 # Fetch the vllm-flash-attn library
 FetchContent_MakeAvailable(vllm-flash-attn)
 message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
 
+# Restore the install prefix
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 # Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
 # case only one is built, in the case both are built redundant work is done)
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
+  DESTINATION vllm/vllm_flash_attn
   COMPONENT _vllm_fa2_C
   FILES_MATCHING PATTERN "*.py"
 )
 
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
+  DESTINATION vllm/vllm_flash_attn
   COMPONENT _vllm_fa3_C
   FILES_MATCHING PATTERN "*.py"
 )

cmake/utils.cmake

@@ -76,7 +76,7 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
   add_custom_target(
     hipify${NAME}
-    COMMAND ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
+    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
     DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${SRCS}
     BYPRODUCTS ${HIP_SRCS}
     COMMENT "Running hipify on ${NAME} extension source files.")

setup.py

@@ -251,11 +251,8 @@ class cmake_build_ext(build_ext):
 
             # CMake appends the extension prefix to the install path,
             # and outdir already contains that prefix, so we need to remove it.
-            # We assume only the final component of extension prefix is added by
-            # CMake, this is currently true for current extensions but may not
-            # always be the case.
             prefix = outdir
-            if '.' in ext.name:
+            for _ in range(ext.name.count('.')):
                 prefix = prefix.parent
 
             # prefix here should actually be the same for all components