[None][refactor] Split thop.attention into phased fmha

cpp/tensorrt_llm/CMakeLists.txt

@@ -22,6 +22,55 @@ set(API_INCLUDE_DIR ${PROJECT_SOURCE_DIR}/include)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/cutlass_extensions/include
                     ${API_INCLUDE_DIR})
 
+get_filename_component(TRTLLM_REPO_ROOT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.."
+                       ABSOLUTE)
+set(TRTLLM_FMHA_PARAMS_GENERATOR
+    "${TRTLLM_REPO_ROOT_DIR}/scripts/generate_fmha_params.py")
+set(TRTLLM_FMHA_PARAMS_SCHEMAS
+    "${TRTLLM_REPO_ROOT_DIR}/tensorrt_llm/_torch/attention/backends/fmha/phased.py"
+    "${TRTLLM_REPO_ROOT_DIR}/tensorrt_llm/_torch/attention/backends/interface.py"
+    "${TRTLLM_REPO_ROOT_DIR}/tensorrt_llm/_torch/attention/backends/trtllm.py"
+    "${TRTLLM_REPO_ROOT_DIR}/tensorrt_llm/_torch/attention/backends/sparse/params.py"
+)
+# The C++ consumer decides which members of the mirrored runtime classes
+# (TrtllmAttention, TrtllmAttentionMetadata) are emitted: a member crosses when
+# these sources read it as attn.<name> / meta.<name>.
+set(TRTLLM_FMHA_PARAMS_USAGE_SOURCES
+    "${CMAKE_CURRENT_SOURCE_DIR}/thop/attentionOp.h"
+    "${CMAKE_CURRENT_SOURCE_DIR}/thop/attentionFeature.h"
+    "${CMAKE_CURRENT_SOURCE_DIR}/thop/attentionVariant.h"
+    "${CMAKE_CURRENT_SOURCE_DIR}/thop/attentionOp.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/nanobind/thop/bindings.cpp")
+set(TRTLLM_FMHA_PARAMS_SCHEMA_ARGS "")
+foreach(_schema ${TRTLLM_FMHA_PARAMS_SCHEMAS})
+  list(APPEND TRTLLM_FMHA_PARAMS_SCHEMA_ARGS --module-path "${_schema}")
+endforeach()
+foreach(_usage ${TRTLLM_FMHA_PARAMS_USAGE_SOURCES})
+  list(APPEND TRTLLM_FMHA_PARAMS_SCHEMA_ARGS --usage-source "${_usage}")
+endforeach()
+set(TRTLLM_FMHA_PARAMS_GENERATED_INCLUDE_DIR
+    "${CMAKE_CURRENT_BINARY_DIR}/generated")
+# Generate during configuration so clangd can resolve the headers before a
+# build.
+set_property(
+  DIRECTORY
+  APPEND
+  PROPERTY CMAKE_CONFIGURE_DEPENDS ${TRTLLM_FMHA_PARAMS_GENERATOR}
+           ${TRTLLM_FMHA_PARAMS_SCHEMAS} ${TRTLLM_FMHA_PARAMS_USAGE_SOURCES})
+execute_process(
+  COMMAND
+    "${Python3_EXECUTABLE}" "${TRTLLM_FMHA_PARAMS_GENERATOR}"
+    ${TRTLLM_FMHA_PARAMS_SCHEMA_ARGS} --class-name FmhaParams --out-dir
+    "${TRTLLM_FMHA_PARAMS_GENERATED_INCLUDE_DIR}/tensorrt_llm/thop"
+  RESULT_VARIABLE TRTLLM_FMHA_PARAMS_RESULT
+  OUTPUT_VARIABLE TRTLLM_FMHA_PARAMS_OUTPUT
+  ERROR_VARIABLE TRTLLM_FMHA_PARAMS_OUTPUT)
+if(NOT TRTLLM_FMHA_PARAMS_RESULT EQUAL 0)
+  message(
+    FATAL_ERROR
+      "Generating native FmhaParams failed:\n${TRTLLM_FMHA_PARAMS_OUTPUT}")
+endif()
+
 set(TARGET_ARCH "unknown")
 
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")

cpp/tensorrt_llm/common/CMakeLists.txt

@@ -76,7 +76,6 @@ set_target_properties(tllm_common_environment
 
 add_library(
   common_src OBJECT
-  attentionOp.cpp
   cublasMMWrapper.cpp
   cudaFp8Utils.cu
   cudaProfilerUtils.cpp