[llm] Add CMake for llama runner

.github/workflows/pull.yml

@@ -213,3 +213,27 @@ jobs:
     uses: ./.github/workflows/_unittest.yml
     with:
       docker-image: executorch-ubuntu-22.04-clang12
+
+  test-llama-runner-cmake:
+    name: test-llama-runner
+    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    strategy:
+      matrix:
+        include:
+          - build-tool: cmake
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-clang12
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        BUILD_TOOL=${{ matrix.build-tool }}
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
+        # Test selective build
+        PYTHON_EXECUTABLE=python bash examples/models/llama2/test_llama_runner.sh "${BUILD_TOOL}"

build/cmake_deps.toml

@@ -13,21 +13,7 @@ excludes = [
   "^third-party",
 ]
 
-[targets.executor_runner]
-buck_targets = [
-  "//examples/portable/executor_runner:executor_runner",
-]
-filters = [
-  ".cpp$",
-]
-excludes = [
-  "^codegen",
-]
-deps = [
-  "executorch",
-  "portable_kernels",
-  "quantized_kernels",
-]
+# ---------------------------------- core start ----------------------------------
 
 [targets.executorch]
 buck_targets = [
@@ -40,94 +26,106 @@ filters = [
   ".cpp$",
 ]
 
-[targets.extension_data_loader]
+
+[targets.portable_kernels]
 buck_targets = [
-  "//extension/data_loader:buffer_data_loader",
-  "//extension/data_loader:file_data_loader",
-  "//extension/data_loader:mmap_data_loader",
-  "//extension/data_loader:shared_ptr_data_loader",
+  # //kernels/portable:operators would be more appropriate, but buck2 doesn't
+  # think it has any "inputs" since its srcs list is empty.
+  "//kernels/portable:generated_lib_all_ops",
 ]
 filters = [
   ".cpp$",
 ]
+excludes = [
+  # Exclude the codegen templates, which are picked up because the buck target
+  # is the generated_lib and not the unwrapped set of kernels.
+  "^codegen/templates",
+]
 deps = [
   "executorch",
 ]
 
-[targets.extension_module]
+[targets.quantized_kernels]
 buck_targets = [
-  "//extension/module:module",
+  "//kernels/quantized:generated_lib",
 ]
 filters = [
   ".cpp$",
 ]
+excludes = [
+  # Exclude the codegen templates, which are picked up because the buck target
+  # is the generated_lib and not the unwrapped set of kernels.
+  "^codegen/templates",
+]
 deps = [
   "executorch",
-  "extension_data_loader",
 ]
 
-[targets.extension_runner_util]
+[targets.program_schema]
 buck_targets = [
-  "//extension/runner_util:inputs",
+  "//schema:program",
 ]
 filters = [
-  ".cpp$",
-]
-deps = [
-  "executorch",
+  ".fbs$",
 ]
 
-[targets.portable_kernels]
+# ---------------------------------- core end ----------------------------------
+# ---------------------------------- extension start ----------------------------------
+
+[targets.extension_data_loader]
 buck_targets = [
-  # //kernels/portable:operators would be more appropriate, but buck2 doesn't
-  # think it has any "inputs" since its srcs list is empty.
-  "//kernels/portable:generated_lib_all_ops",
+  "//extension/data_loader:buffer_data_loader",
+  "//extension/data_loader:file_data_loader",
+  "//extension/data_loader:mmap_data_loader",
+  "//extension/data_loader:shared_ptr_data_loader",
 ]
 filters = [
   ".cpp$",
 ]
-excludes = [
-  # Exclude the codegen templates, which are picked up because the buck target
-  # is the generated_lib and not the unwrapped set of kernels.
-  "^codegen/templates",
-]
 deps = [
   "executorch",
 ]
 
-[targets.quantized_kernels]
+[targets.extension_module]
 buck_targets = [
-  "//kernels/quantized:generated_lib",
+  "//extension/module:module",
 ]
 filters = [
   ".cpp$",
 ]
-excludes = [
-  # Exclude the codegen templates, which are picked up because the buck target
-  # is the generated_lib and not the unwrapped set of kernels.
-  "^codegen/templates",
-]
 deps = [
   "executorch",
+  "extension_data_loader",
 ]
 
-[targets.program_schema]
+[targets.extension_runner_util]
 buck_targets = [
-  "//schema:program",
+  "//extension/runner_util:inputs",
 ]
 filters = [
-  ".fbs$",
+  ".cpp$",
+]
+deps = [
+  "executorch",
 ]
 
-[targets.print_evalue]
+# ---------------------------------- extension end ----------------------------------
+# ---------------------------------- binary start ----------------------------------
+
+[targets.executor_runner]
 buck_targets = [
-  "//extension/evalue_util:print_evalue",
+  "//examples/portable/executor_runner:executor_runner",
 ]
 filters = [
   ".cpp$",
 ]
+excludes = [
+  "^codegen",
+]
 deps = [
   "executorch",
+  "portable_kernels",
+  "quantized_kernels",
 ]
 
 [targets.size_test]
@@ -143,7 +141,8 @@ excludes = [
 deps = [
   "executorch",
 ]
-
+# ---------------------------------- binary end ----------------------------------
+# ---------------------------------- MPS start ----------------------------------
 [targets.mps_executor_runner]
 buck_targets = [
   "//examples/apple/mps/executor_runner:mps_executor_runner",
@@ -178,6 +177,9 @@ filters = [
   ".fbs$",
 ]
 
+# ---------------------------------- MPS end ----------------------------------
+# ---------------------------------- XNNPACK start ----------------------------------
+
 [targets.xnn_executor_runner]
 buck_targets = [
   "//examples/xnnpack:xnn_executor_runner",
@@ -220,3 +222,24 @@ buck_targets = [
 filters = [
   ".fbs$",
 ]
+# ---------------------------------- XNNPACK end ----------------------------------
+# ---------------------------------- LLama start ----------------------------------
+
+[targets.llama_runner]
+buck_targets = [
+  "//examples/models/llama2/runner:runner",
+]
+filters = [
+  ".cpp$",
+]
+excludes = [
+  "^codegen",
+]
+deps = [
+  "executorch",
+  "extension_module",
+  "portable_kernels",
+  "quantized_kernels",
+  "xnnpack_backend",
+]
+# ---------------------------------- LLama start ----------------------------------

build/executorch-config.cmake

@@ -43,7 +43,7 @@ endif()
 
 set(lib_list
     etdump bundled_program extension_data_loader ${FLATCC_LIB} mpsdelegate
-    qnn_executorch_backend
+    qnn_executorch_backend portable_ops_lib extension_module
 )
 foreach(lib ${lib_list})
     # Name of the variable which stores result of the find_library search
@@ -53,7 +53,11 @@ foreach(lib ${lib_list})
         message("${lib} library is not found.
             If needed rebuild with the proper options in CMakeLists.txt")
     else()
-        add_library(${lib} STATIC IMPORTED)
+        if("${lib}" STREQUAL "extension_module")
+            add_library(${lib} SHARED IMPORTED)
+        else()
+            add_library(${lib} STATIC IMPORTED)
+        endif()
         set_target_properties(
             ${lib} PROPERTIES IMPORTED_LOCATION "${${lib_var}}"
         )

examples/models/llama2/CMakeLists.txt

@@ -0,0 +1,64 @@
+# 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.
+
+#
+# Simple CMake build system for selective build demo.
+#
+# ### Editing this file ###
+#
+# This file should be formatted with
+# ~~~
+# cmake-format --first-comment-is-literal=True CMakeLists.txt
+# ~~~
+# It should also be cmake-lint clean.
+#
+cmake_minimum_required(VERSION 3.19)
+project(llama_runner)
+
+if(NOT PYTHON_EXECUTABLE)
+  set(PYTHON_EXECUTABLE python3)
+endif()
+
+set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
+set(TORCH_ROOT ${EXECUTORCH_ROOT}/third-party/pytorch)
+
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+  # Can't set to 11 due to executor_runner.cpp make_unique
+endif()
+
+set(_common_compile_options -Wno-deprecated-declarations -fPIC)
+
+# Let files say "include <executorch/path/to/header.h>".
+set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+
+find_package(gflags REQUIRED PATHS
+             ${CMAKE_CURRENT_BINARY_DIR}/../../../third-party)
+
+#
+# llama_main: test binary to run llama, with tokenizer and sampler integrated
+#
+add_executable(llama_main main.cpp)
+if(CMAKE_BUILD_TYPE EQUAL "RELEASE")
+  target_link_options(llama_main PRIVATE "LINKER:--gc-sections")
+endif()
+
+# find `executorch` libraries
+find_package(executorch CONFIG REQUIRED)
+
+# llama_runner library
+add_subdirectory(runner)
+
+# XNNPACK pthreadpool cpuinfo
+target_link_libraries(llama_main PUBLIC gflags llama_runner portable_ops_lib)
+target_link_options(
+  llama_main PUBLIC "SHELL:LINKER:--whole-archive \
+                    $<TARGET_FILE:portable_ops_lib> \
+                    LINKER:--no-whole-archive")
+target_compile_options(llama_main PUBLIC ${_common_compile_options})
+
+# Print all summary
+executorch_print_configuration_summary()

examples/models/llama2/runner/CMakeLists.txt

@@ -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.
+
+#
+# Simple CMake build system for selective build demo.
+#
+# ### Editing this file ###
+#
+# This file should be formatted with
+# ~~~
+# cmake-format --first-comment-is-literal=True CMakeLists.txt
+# ~~~
+# It should also be cmake-lint clean.
+#
+
+include(${EXECUTORCH_ROOT}/build/Utils.cmake)
+include(${EXECUTORCH_ROOT}/build/Codegen.cmake)
+#
+# The `_<target>_srcs` lists are defined by including ${EXECUTORCH_SRCS_FILE}.
+#
+set(EXECUTORCH_SRCS_FILE
+    "${CMAKE_CURRENT_BINARY_DIR}/../../../../executorch_srcs.cmake")
+
+extract_sources(${EXECUTORCH_SRCS_FILE})
+
+include(${EXECUTORCH_SRCS_FILE})
+
+# build llama_runner library
+list(TRANSFORM _llama_runner__srcs PREPEND "${EXECUTORCH_ROOT}/")
+
+target_include_directories(extension_module
+                           INTERFACE ${_common_include_directories})
+
+add_library(llama_runner SHARED ${_llama_runner__srcs})
+target_link_libraries(
+  llama_runner PUBLIC executorch portable_kernels extension_module
+                      extension_data_loader)