add preprocess to ci

.ci/scripts/test_model.sh

@@ -53,6 +53,7 @@ build_cmake_executor_runner() {
   rm -rf ${CMAKE_OUTPUT_DIR}
   cmake -DCMAKE_BUILD_TYPE=Debug \
       -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+      -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
       -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
       -B${CMAKE_OUTPUT_DIR} .
 
@@ -82,6 +83,9 @@ test_model() {
     rm "./${MODEL_NAME}.pte"
   fi
   STRICT="--strict"
+  if [[ "${MODEL_NAME}" == "clip" ]]; then
+    STRICT="--no-strict"
+  fi
   if [[ "${MODEL_NAME}" == "llava" ]]; then
     # Install requirements for llava
     bash examples/models/llava/install_requirements.sh

CMakeLists.txt

@@ -820,6 +820,11 @@ if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
     list(APPEND _executor_runner_libs quantized_ops_lib)
   endif()
 
+  if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
+    list(APPEND _executor_runner_libs custom_ops)
+    target_link_options_shared_lib(custom_ops)
+  endif()
+
   add_executable(executor_runner ${_executor_runner__srcs})
   if(CMAKE_BUILD_TYPE STREQUAL "Release")
     if(APPLE)

examples/models/__init__.py

@@ -30,6 +30,7 @@
     "resnet18": ("resnet", "ResNet18Model"),
     "resnet50": ("resnet", "ResNet50Model"),
     "llava": ("llava", "LlavaModel"),
+    "clip": ("llama3_2_vision.preprocess", "CLIPImageTransformModel"),
 }
 
 __all__ = [

examples/models/llama3_2_vision/install_requirements.sh

@@ -10,4 +10,4 @@ pip install --pre torchtune --extra-index-url https://download.pytorch.org/whl/n
 
 # Install torchao.
 TORCHAO_VERSION=$(cat "$(dirname "$0")"/../../../.ci/docker/ci_commit_pins/torchao.txt)
-pip install --no-use-pep517 "git+https://github.com/pytorch/ao.git@${TORCHAO_VERSION}"
+pip install "git+https://github.com/pytorch/ao.git@${TORCHAO_VERSION}"

examples/models/llama3_2_vision/preprocess/__init__.py

@@ -0,0 +1,11 @@
+# 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.
+
+from .model import CLIPImageTransformModel
+
+__all__ = [
+    CLIPImageTransformModel,
+]

extension/llm/custom_ops/targets.bzl

@@ -38,6 +38,7 @@ def define_common_targets():
             deps = [
                 "//executorch/kernels/portable/cpu/util:reduce_util",
                 "//executorch/extension/llm/custom_ops/spinquant:fast_hadamard_transform",
+                ":op_tile_crop",
             ],
             compiler_flags = ["-Wno-missing-prototypes", "-Wno-global-constructors"] + get_compiler_optimization_flags(),
             visibility = [
@@ -55,10 +56,8 @@ def define_common_targets():
             srcs = [
                 "op_fast_hadamard_transform_aten.cpp",
                 "op_sdpa_aot.cpp",
-                "op_tile_crop.cpp",
                 "op_tile_crop_aot.cpp",
             ],
-            headers = ["op_tile_crop.h"],
             compiler_flags = ["-Wno-global-constructors"],
             visibility = [
                 "//executorch/...",
@@ -69,6 +68,7 @@ def define_common_targets():
             ],
             deps = [
                 ":custom_ops" + mkl_dep,
+                ":op_tile_crop",
                 "//executorch/extension/aten_util:aten_bridge",
             ],
         )