Merge branch 'main' into dev_lift_const

Author: chunit-quic

.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/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_model.sh

@@ -91,7 +91,17 @@ 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
 }

.ci/scripts/unittest-linux.sh

@@ -14,25 +14,39 @@ else
   exit 1
 fi
 
+BUILD_MODE=$2
+if [[ "${BUILD_MODE:-}" =~ ^(Debug|Release)$ ]]; then
+    echo "Running tests in build mode ${BUILD_MODE} ..."
+else
+    echo "Unsupported build mode ${BUILD_MODE}, options are Debug or Release."
+    exit 1
+fi
+
 # The generic Linux job chooses to use base env, not the one setup by the image
 eval "$(conda shell.bash hook)"
 CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
 conda activate "${CONDA_ENV}"
 
-# Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
-source .ci/scripts/setup-vulkan-linux-deps.sh
+if [[ "$BUILD_TOOL" == "cmake" ]]; then
+    # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
+    source .ci/scripts/setup-vulkan-linux-deps.sh
 
-PYTHON_EXECUTABLE=python \
-EXECUTORCH_BUILD_PYBIND=ON \
-CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
-.ci/scripts/setup-linux.sh "$BUILD_TOOL"
+    PYTHON_EXECUTABLE=python \
+    EXECUTORCH_BUILD_PYBIND=ON \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    .ci/scripts/setup-linux.sh "$BUILD_TOOL" "$BUILD_MODE"
 
-# Install llama3_2_vision dependencies.
-PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
+    # Install llama3_2_vision dependencies.
+    PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
 
-if [[ "$BUILD_TOOL" == "cmake" ]]; then
     .ci/scripts/unittest-linux-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
+    # Removing this breaks sccache in the Buck build, apparently
+    # because TMPDIR gets messed up? Please feel free to fix this and
+    # speed up this CI job!
+    PYTHON_EXECUTABLE=python \
+    .ci/scripts/setup-linux.sh "$BUILD_TOOL" "$BUILD_MODE"
+
     .ci/scripts/unittest-buck2.sh
 else
     echo "Unknown build tool $BUILD_TOOL"

.ci/scripts/unittest-macos.sh

@@ -14,6 +14,14 @@ else
   exit 1
 fi
 
+BUILD_MODE=$2
+if [[ $BUILD_MODE =~ ^(Debug|Release)$ ]]; then
+    echo "Running tests in build mode ${BUILD_MODE} ..."
+else
+    echo "Unsupported build mode ${BUILD_MODE}, options are Debug or Release."
+    exit 1
+fi
+
 bash .ci/scripts/setup-conda.sh
 eval "$(conda shell.bash hook)"
 
@@ -22,17 +30,19 @@ export TMP_DIR=$(mktemp -d)
 export PATH="${TMP_DIR}:$PATH"
 trap 'rm -rfv ${TMP_DIR}' EXIT
 
-# Setup MacOS dependencies as there is no Docker support on MacOS atm
-PYTHON_EXECUTABLE=python \
-EXECUTORCH_BUILD_PYBIND=ON \
-CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
-${CONDA_RUN} --no-capture-output \
-.ci/scripts/setup-macos.sh cmake
+if [[ "$BUILD_TOOL" == "cmake" ]]; then
+    # Setup MacOS dependencies as there is no Docker support on MacOS atm
+    PYTHON_EXECUTABLE=python \
+    EXECUTORCH_BUILD_PYBIND=ON \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    ${CONDA_RUN} --no-capture-output \
+    .ci/scripts/setup-macos.sh "${BUILD_TOOL}" "${BUILD_MODE}"
 
-# Install llama3_2_vision dependencies.
-PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
+    # Install llama3_2_vision dependencies.
+    PYTHON_EXECUTABLE=python \
+    ${CONDA_RUN} --no-capture-output \
+    ./examples/models/llama3_2_vision/install_requirements.sh
 
-if [[ "$BUILD_TOOL" == "cmake" ]]; then
     .ci/scripts/unittest-macos-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
     .ci/scripts/unittest-buck2.sh

.ci/scripts/utils.sh

@@ -109,7 +109,7 @@ build_executorch_runner_cmake() {
   pushd "${CMAKE_OUTPUT_DIR}" || return
   # This command uses buck2 to gather source files and buck2 could crash flakily
   # on MacOS
-  retry cmake -DPYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" -DCMAKE_BUILD_TYPE=Release ..
+  retry cmake -DPYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" -DCMAKE_BUILD_TYPE="${1:-Release}" ..
   popd || return
 
   if [ "$(uname)" == "Darwin" ]; then
@@ -124,7 +124,7 @@ build_executorch_runner() {
   if [[ $1 == "buck2" ]]; then
     build_executorch_runner_buck2
   elif [[ $1 == "cmake" ]]; then
-    build_executorch_runner_cmake
+    build_executorch_runner_cmake "$2"
   else
     echo "Invalid build tool $1. Only buck2 and cmake are supported atm"
     exit 1

.github/scripts/extract_benchmark_results.py

@@ -229,21 +229,22 @@ def extract_ios_metric(
 
     elif method == "forward":
         if metric_name == "Clock Monotonic Time, s":
-            benchmark_result["metric"] = (
-                "generate_time(ms)"
-                if "llama" in test_name
-                else "avg_inference_latency(ms)"
-            )
+            benchmark_result["metric"] = "avg_inference_latency(ms)"
             benchmark_result["actualValue"] = metric_value * 1000
 
         elif metric_name == "Memory Peak Physical, kB":
             # NB: Showing the value in mB is friendlier IMO
             benchmark_result["metric"] = "peak_inference_mem_usage(mb)"
             benchmark_result["actualValue"] = metric_value / 1024
 
-    elif method == "generate" and metric_name == "Tokens Per Second, t/s":
-        benchmark_result["metric"] = "token_per_sec"
-        benchmark_result["actualValue"] = metric_value
+    elif method == "generate":
+        if metric_name == "Clock Monotonic Time, s":
+            benchmark_result["metric"] = "generate_time(ms)"
+            benchmark_result["actualValue"] = metric_value * 1000
+
+        elif metric_name == "Tokens Per Second, t/s":
+            benchmark_result["metric"] = "token_per_sec"
+            benchmark_result["actualValue"] = metric_value
 
     return benchmark_result
 

.github/workflows/_android.yml

@@ -29,7 +29,7 @@ jobs:
         export ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded
 
         # Build LLM Demo for Android
-        bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+        bash build/build_android_library.sh ${ARTIFACTS_DIR_NAME}
 
   # Running Android emulator directly on the runner and not using Docker
   run-emulator: