feat: add FlashComm1 and MMRS fusion support for Qwen3.5 NPU.

CMakeLists.txt

@@ -334,9 +334,53 @@ find_package(benchmark CONFIG REQUIRED)
 find_package(nlohmann_json CONFIG REQUIRED)
 find_package(OpenCV CONFIG REQUIRED)
 find_package(FFMPEG REQUIRED)
-find_package(Python COMPONENTS Development REQUIRED)
+find_package(Python COMPONENTS Interpreter Development REQUIRED)
 find_package(pybind11 CONFIG REQUIRED)
 
+if(USE_NPU)
+  function(xllm_set_env_from_python VAR PYTHON_CODE)
+    if("$ENV{${VAR}}" STREQUAL "")
+      execute_process(
+        COMMAND ${Python_EXECUTABLE} -c "${PYTHON_CODE}"
+        OUTPUT_VARIABLE XLLM_INFERRED_ENV_VALUE
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_VARIABLE XLLM_INFERRED_ENV_ERROR
+        RESULT_VARIABLE XLLM_INFERRED_ENV_RESULT
+      )
+      if(NOT XLLM_INFERRED_ENV_RESULT EQUAL 0 OR "${XLLM_INFERRED_ENV_VALUE}" STREQUAL "")
+        message(FATAL_ERROR "Failed to infer ${VAR}: ${XLLM_INFERRED_ENV_ERROR}")
+      endif()
+      set(ENV{${VAR}} "${XLLM_INFERRED_ENV_VALUE}")
+      message(STATUS "Inferred ${VAR}: $ENV{${VAR}}")
+    endif()
+  endfunction()
+
+  if("$ENV{PYTHON_INCLUDE_PATH}" STREQUAL "")
+    set(ENV{PYTHON_INCLUDE_PATH} "${Python_INCLUDE_DIRS}")
+  endif()
+  if("$ENV{PYTHON_LIB_PATH}" STREQUAL "")
+    execute_process(
+      COMMAND ${Python_EXECUTABLE} -c "import sysconfig; print(sysconfig.get_config_var('LIBDIR') or '')"
+      OUTPUT_VARIABLE XLLM_PYTHON_LIB_PATH
+      OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(ENV{PYTHON_LIB_PATH} "${XLLM_PYTHON_LIB_PATH}")
+  endif()
+  xllm_set_env_from_python(PYTORCH_INSTALL_PATH "import importlib.util; spec = importlib.util.find_spec('torch'); print(spec.submodule_search_locations[0])")
+  xllm_set_env_from_python(PYTORCH_NPU_INSTALL_PATH "import importlib.util; spec = importlib.util.find_spec('torch_npu'); print(spec.submodule_search_locations[0])")
+  if(NOT EXISTS "$ENV{PYTORCH_NPU_INSTALL_PATH}/include/torch_npu/torch_npu.h"
+      AND EXISTS "/usr/local/libtorch_npu/include/torch_npu/torch_npu.h")
+    set(ENV{PYTORCH_NPU_INSTALL_PATH} "/usr/local/libtorch_npu")
+    message(STATUS "Using PYTORCH_NPU_INSTALL_PATH with compatibility header: $ENV{PYTORCH_NPU_INSTALL_PATH}")
+  endif()
+  if("$ENV{LIBTORCH_ROOT}" STREQUAL "")
+    set(ENV{LIBTORCH_ROOT} "$ENV{PYTORCH_INSTALL_PATH}")
+  endif()
+  if("$ENV{NPU_HOME_PATH}" STREQUAL "" AND NOT "$ENV{ASCEND_HOME_PATH}" STREQUAL "")
+    set(ENV{NPU_HOME_PATH} "$ENV{ASCEND_HOME_PATH}")
+  endif()
+endif()
+
 if (USE_CXX11_ABI)
   # only use jemalloc if using the new C++-11 ABI
   find_package(Jemalloc)

cmake/FindRust.cmake

@@ -10,12 +10,10 @@ else()
 	set(USER_HOME "$ENV{HOME}")
 endif()
 
-if(NOT DEFINED CARGO_HOME)
-	if("$ENV{CARGO_HOME}" STREQUAL "")
-		set(CARGO_HOME "${USER_HOME}/.cargo")
-	else()
-		set(CARGO_HOME "$ENV{CARGO_HOME}")
-	endif()
+if(NOT "$ENV{CARGO_HOME}" STREQUAL "")
+	set(CARGO_HOME "$ENV{CARGO_HOME}" CACHE PATH "Rust Cargo Home" FORCE)
+elseif(NOT DEFINED CARGO_HOME)
+	set(CARGO_HOME "${USER_HOME}/.cargo")
 endif()
 
 # Find cargo executable

cmake/cargo_library.cmake

@@ -65,7 +65,9 @@ function(cargo_library)
 
   file(GLOB_RECURSE LIB_SOURCES "*.rs")
 
-  set(CARGO_ENV_COMMAND ${CMAKE_COMMAND} -E env "CARGO_TARGET_DIR=${CMAKE_CURRENT_BINARY_DIR}")
+  set(CARGO_ENV_COMMAND ${CMAKE_COMMAND} -E env
+      "CARGO_HOME=${CARGO_HOME}"
+      "CARGO_TARGET_DIR=${CMAKE_CURRENT_BINARY_DIR}")
 
   # build the library target with cargo
   set(STATIC_LIB_NAME

cmake/cargo_shared_library.cmake

@@ -65,7 +65,9 @@ function(cargo_shared_library)
 
   file(GLOB_RECURSE LIB_SOURCES "*.rs")
 
-  set(CARGO_ENV_COMMAND ${CMAKE_COMMAND} -E env "CARGO_TARGET_DIR=${CMAKE_CURRENT_BINARY_DIR}")
+  set(CARGO_ENV_COMMAND ${CMAKE_COMMAND} -E env
+      "CARGO_HOME=${CARGO_HOME}"
+      "CARGO_TARGET_DIR=${CMAKE_CURRENT_BINARY_DIR}")
 
   # build the library target with cargo
   set(SHARED_LIB_NAME

xllm/core/common/CMakeLists.txt

@@ -17,16 +17,19 @@ cc_library(
     types.h
     device_monitor.h
     version_singleton.h
+    flash_comm1_context.h
   SRCS
     etcd_client.cpp
     metrics.cpp
     $<$<BOOL:${USE_NPU}>:mspti_helper.cpp>
     options.cpp
     rate_limiter.cpp
     device_monitor.cpp
+    flash_comm1_context.cpp
   DEPS
     :config
     util
+    parallel_state
     absl::random_random
     absl::strings
     torch