Merge branch 'main' into jz/quantize_kwargs

Author: jackzhxng

.buckconfig

@@ -8,14 +8,15 @@
   root = .
   prelude = third-party/prelude
   shim = shim
+  shim_et = shim_et
 
 [repository_aliases]
   config = prelude
   ovr_config = prelude
-  toolchains = shim
-  fbcode = shim
-  fbcode_macros = shim
-  fbsource = shim
+  toolchains = shim_et
+  fbcode = shim_et
+  fbcode_macros = shim_et
+  fbsource = shim_et
   buck = shim
 
 [cxx]

.ci/scripts/build_llama_android.sh

@@ -14,7 +14,6 @@ if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
   PYTHON_EXECUTABLE=python3
 fi
 which "${PYTHON_EXECUTABLE}"
-CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
 
 install_executorch_and_backend_lib() {
   echo "Installing executorch and xnnpack backend"
@@ -28,7 +27,6 @@ install_executorch_and_backend_lib() {
     -DANDROID_ABI="${ANDROID_ABI}" \
     -DCMAKE_INSTALL_PREFIX=cmake-android-out \
     -DCMAKE_BUILD_TYPE=Release \
-    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
     -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
     -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
@@ -54,7 +52,6 @@ build_llama_runner() {
     -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
-    -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
     -Bcmake-android-out/examples/models/llama examples/models/llama
 
     cmake --build cmake-android-out/examples/models/llama -j4 --config Release

.ci/scripts/gather_test_models.py

@@ -104,8 +104,12 @@ def model_should_run_on_target_os(model: str, target_os: str) -> bool:
     For example, a big model can be disabled in macos due to the limited macos resources.
     """
     if target_os == "macos":
+        # Disabled in macos due to limited resources, and should stay that way even if
+        # we otherwise re-enable.
         return model not in ["llava"]
-    return True
+    # Disabled globally because we have test-llava-runner-linux that does a more
+    # comprehensive E2E test of llava.
+    return model not in ["llava"]
 
 
 def export_models_for_ci() -> dict[str, dict]:

.ci/scripts/setup-linux.sh

@@ -22,7 +22,7 @@ fi
 # have already been installed, so we use PyTorch build from source here instead
 # of nightly. This allows CI to test against latest commits from PyTorch
 install_executorch "use-pt-pinned-commit"
-build_executorch_runner "${BUILD_TOOL}"
+build_executorch_runner "${BUILD_TOOL}" "${2:-Release}"
 
 if [[ "${GITHUB_BASE_REF:-}" == *main* || "${GITHUB_BASE_REF:-}" == *gh* ]]; then
   do_not_use_nightly_on_ci

.ci/scripts/setup-macos.sh

@@ -136,7 +136,7 @@ install_pytorch_and_domains
 # We build PyTorch from source here instead of using nightly. This allows CI to test against
 # the pinned commit from PyTorch
 install_executorch "use-pt-pinned-commit"
-build_executorch_runner "${BUILD_TOOL}"
+build_executorch_runner "${BUILD_TOOL}" "${2:-Release}"
 
 if [[ "${GITHUB_BASE_REF:-}" == *main* || "${GITHUB_BASE_REF:-}" == *gh* ]]; then
   do_not_use_nightly_on_ci

.ci/scripts/test_llama.sh

@@ -154,7 +154,6 @@ cmake_install_executorch_libraries() {
     rm -rf cmake-out
     retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')" \
         -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \

.ci/scripts/test_llava.sh

@@ -31,7 +31,6 @@ NPROC=8
 if hash nproc &> /dev/null; then NPROC=$(nproc); fi
 
 python_lib=$($PYTHON_EXECUTABLE -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())')
-CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
 EXECUTORCH_COMMON_CMAKE_ARGS="                      \
         -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}         \
         -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}      \
@@ -48,7 +47,6 @@ EXECUTORCH_COMMON_CMAKE_ARGS="                      \
 cmake_install_executorch_libraries() {
     cmake                               \
         ${EXECUTORCH_COMMON_CMAKE_ARGS} \
-        "-DCMAKE_PREFIX_PATH=${CMAKE_PREFIX_PATH}" \
         -B${BUILD_DIR} .
 
     cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${CMAKE_BUILD_TYPE}
@@ -59,7 +57,6 @@ cmake_install_executorch_libraries_for_android() {
         -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
         -DANDROID_ABI=arm64-v8a                                                 \
         ${EXECUTORCH_COMMON_CMAKE_ARGS}                                         \
-        "-DCMAKE_PREFIX_PATH=${CMAKE_PREFIX_PATH}" \
         -B${BUILD_DIR} .
 
     cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${CMAKE_BUILD_TYPE}
@@ -80,7 +77,7 @@ cmake_build_llava_runner() {
 
     cmake                                 \
         ${LLAVA_COMMON_CMAKE_ARGS}        \
-        -DCMAKE_PREFIX_PATH="$python_lib;${CMAKE_PREFIX_PATH}" \
+        -DCMAKE_PREFIX_PATH="$python_lib" \
         -B${BUILD_DIR}/${dir}             \
         ${dir}
 
@@ -96,7 +93,7 @@ cmake_build_llava_runner_for_android() {
         -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
         -DANDROID_ABI=arm64-v8a                                                 \
         ${LLAVA_COMMON_CMAKE_ARGS}                                              \
-        -DCMAKE_PREFIX_PATH="$python_lib;${CMAKE_PREFIX_PATH}"                  \
+        -DCMAKE_PREFIX_PATH="$python_lib"                  \
         -DLLAVA_RUNNER_NO_TORCH_DUMMY_IMAGE=ON                                  \
         -B${BUILD_DIR}/${dir}                                                   \
         ${dir}

.ci/scripts/test_model.sh

@@ -50,12 +50,10 @@ prepare_artifacts_upload() {
 
 build_cmake_executor_runner() {
   echo "Building executor_runner"
-  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   rm -rf ${CMAKE_OUTPUT_DIR}
   cmake -DCMAKE_BUILD_TYPE=Debug \
       -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
       -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
-      -DCMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH" \
       -B${CMAKE_OUTPUT_DIR} .
 
   cmake --build ${CMAKE_OUTPUT_DIR} -j4 --config Debug
@@ -93,21 +91,29 @@ test_model() {
     # Install requirements for llama vision.
     bash examples/models/llama3_2_vision/install_requirements.sh
   fi
-  # python3 -m examples.portable.scripts.export --model_name="llama2" should works too
+  if [[ "${MODEL_NAME}" == "qwen2_5" ]]; then
+      # Install requirements for export_llama
+      bash examples/models/llama/install_requirements.sh
+      # Test export_llama script: python3 -m examples.models.llama.export_llama.
+      # Use Llama random checkpoint with Qwen 2.5 1.5b model configuration.
+      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/qwen2_5/1_5b_config.json
+      rm "./${MODEL_NAME}.pte"
+      return  # Skip running with portable executor runnner since portable doesn't support Qwen's biased linears.
+  fi
+
+  # Export a basic .pte and run the model.
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}" "${STRICT}"
   run_portable_executor_runner
 }
 
 build_cmake_xnn_executor_runner() {
   echo "Building xnn_executor_runner"
-  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
 
   (rm -rf ${CMAKE_OUTPUT_DIR} \
     && mkdir ${CMAKE_OUTPUT_DIR} \
     && cd ${CMAKE_OUTPUT_DIR} \
     && retry cmake -DCMAKE_BUILD_TYPE=Release \
       -DEXECUTORCH_BUILD_XNNPACK=ON \
-      -DCMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH" \
       -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" ..)
 
   cmake --build ${CMAKE_OUTPUT_DIR} -j4

.ci/scripts/test_phi_3_mini.sh

@@ -22,10 +22,8 @@ NPROC=8
 if hash nproc &> /dev/null; then NPROC=$(nproc); fi
 
 cmake_install_executorch_libraries() {
-  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   cmake -DPYTHON_EXECUTABLE=python \
       -DCMAKE_INSTALL_PREFIX=${BUILD_DIR} \
-      -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
       -DEXECUTORCH_ENABLE_LOGGING=1 \
       -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
       -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
@@ -41,10 +39,8 @@ cmake_install_executorch_libraries() {
 }
 
 cmake_build_phi_3_mini() {
-  CMAKE_PREFIX_PATH="$(python3 -c 'import torch as _; print(_.__path__[0])')"
   cmake -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
       -DCMAKE_INSTALL_PREFIX=${BUILD_DIR} \
-      -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}" \
       -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
       -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
       -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \

.ci/scripts/test_qnn_static_llama.sh

@@ -34,11 +34,11 @@ $PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o to
 
 set +e
 # Compile only as weight sharing is not applicable on x86
-$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleScript.test_stories_single_llama --model SM8650 --build_folder build-android/ --executorch_root . --artifact_dir . --compile_only
+$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-android/ --executorch_root . --artifact_dir . --llama_artifacts . --compile_only
 exit_code1=$?
 
 # Checks accuracy with weight sharing disabled since x86 does not support weight sharing.
-$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleScript.test_stories_single_llama --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir . --enable_x86_64
+$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir . --llama_artifacts . --enable_x86_64
 exit_code2=$?
 
 # Check the exit codes and print messages