Merge branch 'main' into arm-passes-init

Author: digantdesai

.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/unittest-linux.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 ${BUILD_MODE}, options are Debug or Release."
+    exit 1
+fi
+
 # The generic Linux job chooses to use base env, not the one setup by the image
 eval "$(conda shell.bash hook)"
 CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
@@ -25,7 +33,7 @@ source .ci/scripts/setup-vulkan-linux-deps.sh
 PYTHON_EXECUTABLE=python \
 EXECUTORCH_BUILD_PYBIND=ON \
 CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
-.ci/scripts/setup-linux.sh "$BUILD_TOOL"
+.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

.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)"
 
@@ -27,7 +35,7 @@ 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" \
 ${CONDA_RUN} --no-capture-output \
-.ci/scripts/setup-macos.sh cmake
+.ci/scripts/setup-macos.sh "${BUILD_TOOL}" "${BUILD_MODE}"
 
 # Install llama3_2_vision dependencies.
 PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh

.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/_unittest.yml

@@ -7,6 +7,10 @@ on:
         required: true
         type: string
         description: Name of the docker image to use.
+      build-mode:
+        required: true
+        type: string
+        description: Build mode to use, Debug or Release.
       build-tool:
         required: true
         type: string
@@ -30,7 +34,7 @@ jobs:
       timeout: 90
       script: |
         set -eux
-        .ci/scripts/unittest-linux.sh "${{ inputs.build-tool }}"
+        .ci/scripts/unittest-linux.sh "${{ inputs.build-tool }}" "${{ inputs.build-mode }}"
 
   macos:
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
@@ -41,4 +45,4 @@ jobs:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
         set -eux
-        .ci/scripts/unittest-macos.sh "${{ inputs.build-tool }}"
+        .ci/scripts/unittest-macos.sh "${{ inputs.build-tool }}" "${{ inputs.build-mode }}"

.github/workflows/pull.yml

@@ -367,6 +367,7 @@ jobs:
       id-token: write
       contents: read
     with:
+      build-mode: Debug
       build-tool: cmake
       docker-image: executorch-ubuntu-22.04-clang12
 
@@ -376,6 +377,7 @@ jobs:
       id-token: write
       contents: read
     with:
+      build-mode: Debug
       build-tool: buck2
       docker-image: executorch-ubuntu-22.04-clang12
 

.github/workflows/trunk.yml

@@ -489,3 +489,13 @@ jobs:
         PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
         # Test llama2
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh -model stories110M -build_tool "${BUILD_TOOL}" -mode "${MODE}" -dtype "${DTYPE}" -pt2e_quantize "${PT2E_QUANTIZE}"
+
+  unittest-release:
+    uses: ./.github/workflows/_unittest.yml
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      build-mode: Release
+      build-tool: cmake
+      docker-image: executorch-ubuntu-22.04-clang12

backends/arm/_passes/fuse_quantized_activation_pass.py

@@ -13,7 +13,8 @@
 
 
 class FuseQuantizedActivationPass(ExportPass):
-    def _is_fuseable_quantized_activation(self, node: Node):
+    @staticmethod
+    def _is_fuseable_quantized_activation(node: Node):
         """Fuse activations that have a 0 lower bound and quantized with a qmin zero-point"""
         is_fuseable = node.target == exir_ops.edge.aten.relu.default
         if node.target == exir_ops.edge.aten.hardtanh.default:
@@ -29,7 +30,8 @@ def _is_fuseable_quantized_activation(self, node: Node):
         else:
             return False
 
-    def _is_fuseable_input(self, node: Node):
+    @staticmethod
+    def _is_fuseable_input(node: Node):
         return (
             node.target
             in (
@@ -45,11 +47,11 @@ def call(self, graph_module: torch.fx.GraphModule):
             if node.op != "call_function":
                 continue
 
-            if not self._is_fuseable_quantized_activation(node):
+            if not FuseQuantizedActivationPass._is_fuseable_quantized_activation(node):
                 continue
 
             input_node = node.args[0]
-            if not self._is_fuseable_input(input_node):
+            if not FuseQuantizedActivationPass._is_fuseable_input(input_node):
                 continue
 
             node.replace_all_uses_with(input_node)