[WIP] refactor: unify pretrain entry for GPT model

CMakeLists.txt

@@ -72,6 +72,9 @@ endif()
 # Framework core sources (*.cc), excluding cpu kernels (they are built separately)
 file(GLOB_RECURSE SRC ${PROJECT_SOURCE_DIR}/infini_train/src/*.cc)
 list(FILTER SRC EXCLUDE REGEX ".*kernels/cpu/.*")
+# CLI flags are opt-in: linking the core must not register names such as
+# --device or --batch_size that applications may define themselves.
+list(FILTER SRC EXCLUDE REGEX ".*/training/arguments\\.cc$")
 
 # Exclude backend-specific runtime/ccl translation units when the corresponding
 # backend is disabled. This keeps each build self-contained and avoids pulling
@@ -228,52 +231,20 @@ link_infini_train_exe(infini_train_executable)
 add_library(InfiniTrain::infini_train_executable ALIAS infini_train_executable)
 unset(_infini_train_executable_archives)
 
+# ------------------------------------------------------------------------------
+# Training CLI (opt-in for applications)
+# ------------------------------------------------------------------------------
+
+add_library(infini_train_training_cli STATIC infini_train/src/training/arguments.cc)
+add_library(InfiniTrain::training_cli ALIAS infini_train_training_cli)
+target_link_libraries(infini_train_training_cli PUBLIC InfiniTrain::infini_train)
+
 # ------------------------------------------------------------------------------
 # Examples
 # ------------------------------------------------------------------------------
 
 if(PROJECT_IS_TOP_LEVEL)
-add_executable(mnist
-  example/mnist/main.cc
-  example/mnist/dataset.cc
-  example/mnist/net.cc
-)
-link_infini_train_exe(mnist)
-
-add_executable(gpt2
-  example/gpt2/main.cc
-  example/common/tiny_shakespeare_dataset.cc
-  example/common/utils.cc
-  example/common/tokenizer.cc
-  example/gpt2/checkpoint_loader.cc
-)
-link_infini_train_exe(gpt2)
-
-add_executable(mixtral
-  example/mixtral/main.cc
-  example/common/tiny_shakespeare_dataset.cc
-  example/common/utils.cc
-  example/mixtral/checkpoint_loader.cc
-)
-link_infini_train_exe(mixtral)
-
-add_executable(llama3
-  example/llama3/main.cc
-  example/common/tiny_shakespeare_dataset.cc
-  example/common/utils.cc
-  example/common/tokenizer.cc
-  example/llama3/checkpoint_loader.cc
-)
-link_infini_train_exe(llama3)
-
-add_executable(qwen3
-  example/qwen3/main.cc
-  example/common/tiny_shakespeare_dataset.cc
-  example/common/utils.cc
-  example/qwen3/checkpoint_loader.cc
-  example/common/tokenizer.cc
-)
-link_infini_train_exe(qwen3)
+  add_subdirectory(example)
 endif()
 
 # Tools

README.md

@@ -140,6 +140,11 @@ HF_TOKEN=hf_xxx \
 
 #### Model Examples
 
+GPT-2, LLaMA3, Qwen3, and Mixtral share the `pretrain_gpt` training program.
+Model presets are selected with `--flagfile=example/<model>/train.flags`;
+subsequent CLI options override preset values. The named executables load
+`configs/<model>/train.flags` beside the binary automatically.
+
 The generated files can be passed directly to the corresponding executables:
 
 ##### MNIST

cmake/gpt_examples.cmake

@@ -0,0 +1,51 @@
+include_guard(GLOBAL)
+
+function(infini_train_add_gpt_executable target_name)
+  add_executable(${target_name})
+  target_link_libraries(${target_name} PRIVATE infini_train_gpt_sources InfiniTrain::training_cli)
+  # Keep the framework archives as direct link entries in their whole-archive
+  # group; transitive de-duplication can otherwise move the core outside it.
+  link_infini_train_exe(${target_name})
+endfunction()
+
+# Called by example/<model>/CMakeLists.txt. Each model contributes source files
+# to the shared GPT program and declares a preset-loading executable.
+function(infini_train_add_gpt_model model_name)
+  cmake_parse_arguments(MODEL "" "FLAGFILE" "SOURCES" ${ARGN})
+  if(MODEL_UNPARSED_ARGUMENTS OR MODEL_KEYWORDS_MISSING_VALUES)
+    message(FATAL_ERROR
+      "infini_train_add_gpt_model(${model_name}): invalid or incomplete arguments")
+  endif()
+  if(NOT MODEL_SOURCES OR NOT MODEL_FLAGFILE)
+    message(FATAL_ERROR
+      "infini_train_add_gpt_model(${model_name}): SOURCES and FLAGFILE are required")
+  endif()
+  if(NOT TARGET infini_train_gpt_sources)
+    message(FATAL_ERROR "GPT models must be declared from the example build")
+  endif()
+
+  # Absolute source paths retain the declaring directory's meaning when the
+  # interface sources are consumed by executables in other model directories.
+  foreach(source IN LISTS MODEL_SOURCES)
+    get_filename_component(source_path "${source}" ABSOLUTE BASE_DIR "${CMAKE_CURRENT_SOURCE_DIR}")
+    target_sources(infini_train_gpt_sources INTERFACE "${source_path}")
+  endforeach()
+  get_filename_component(flagfile "${MODEL_FLAGFILE}" ABSOLUTE BASE_DIR "${CMAKE_CURRENT_SOURCE_DIR}")
+  if(NOT EXISTS "${flagfile}" OR IS_DIRECTORY "${flagfile}")
+    message(FATAL_ERROR "Training preset not found: ${flagfile}")
+  endif()
+
+  infini_train_add_gpt_executable(${model_name})
+  target_compile_definitions(${model_name} PRIVATE
+    INFINITRAIN_GPT_DEFAULT_FLAGFILE="configs/${model_name}/train.flags")
+
+  # Target-file directories also cover multi-config generators and custom output
+  # locations. Preset edits are copied on the next build without recompilation.
+  add_custom_target(${model_name}_training_config
+    COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:${model_name}>/configs/${model_name}"
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different
+      "${flagfile}" "$<TARGET_FILE_DIR:${model_name}>/configs/${model_name}/train.flags"
+    VERBATIM
+  )
+  add_dependencies(${model_name} ${model_name}_training_config)
+endfunction()