Merge branch 'main' into refactor-bitwise-logical-tests

Author: YufengShi-dudu

.ci/scripts/build_android_instrumentation.sh

@@ -13,9 +13,11 @@ fi
 which "${PYTHON_EXECUTABLE}"
 
 build_android_test() {
-  pushd extension/android_test
-  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew testDebugUnitTest
-  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew build assembleAndroidTest
+  mkdir -p extension/android/executorch_android/src/androidTest/resources
+  cp extension/module/test/resources/add.pte extension/android/executorch_android/src/androidTest/resources
+  pushd extension/android
+  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:testDebugUnitTest
+  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:assembleAndroidTest
   popd
 }
 
@@ -24,8 +26,7 @@ collect_artifacts_to_be_uploaded() {
   # Collect Java library test
   JAVA_LIBRARY_TEST_DIR="${ARTIFACTS_DIR_NAME}/library_test_dir"
   mkdir -p "${JAVA_LIBRARY_TEST_DIR}"
-  cp extension/android_test/build/outputs/apk/debug/*.apk "${JAVA_LIBRARY_TEST_DIR}"
-  cp extension/android_test/build/outputs/apk/androidTest/debug/*.apk "${JAVA_LIBRARY_TEST_DIR}"
+  cp extension/android/executorch_android/build/outputs/apk/androidTest/debug/*.apk "${JAVA_LIBRARY_TEST_DIR}"
 }
 
 main() {

.github/workflows/_android.yml

@@ -28,14 +28,16 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool buck2
         export ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded
 
+        mkdir -p ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
+        bash examples/models/llama/install_requirements.sh
+        bash ".ci/scripts/test_llama.sh" -model stories110M -build_tool cmake -dtype fp16 -mode portable -upload ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
+
         # Build LLM Demo for Android
         export BUILD_AAR_DIR=aar-out
+        mkdir -p $BUILD_AAR_DIR
         bash build/build_android_library.sh ${ARTIFACTS_DIR_NAME}
         bash .ci/scripts/build_android_instrumentation.sh ${ARTIFACTS_DIR_NAME}
 
-        mkdir -p ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
-        bash ".ci/scripts/test_llama.sh" -model stories110M -build_tool cmake -dtype fp16 -mode portable -upload ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
-
         mkdir -p examples/demo-apps/android/LlamaDemo/app/libs
         cp aar-out/executorch.aar examples/demo-apps/android/LlamaDemo/app/libs
         pushd examples/demo-apps/android/LlamaDemo
@@ -94,7 +96,6 @@ jobs:
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug-androidTest.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/fp32-xnnpack-custom/model.zip
-          curl -o android-test-debug.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug.apk
           curl -o android-test-debug-androidTest.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug-androidTest.apk
           unzip model.zip
           mv *.pte model.pte

.github/workflows/doc-build.yml

@@ -26,7 +26,7 @@ jobs:
     with:
       job-name: Build doc
       runner: linux.2xlarge
-      docker-image: executorch-ubuntu-22.04-clang12
+      docker-image: executorch-ubuntu-22.04-clang12-android
       submodules: 'true'
       repository: pytorch/executorch
       upload-artifact: docs
@@ -70,8 +70,8 @@ jobs:
 
         # Build javadoc:
         cd extension/android
-        ./gradlew javadoc
-        cp -rf build/docs/javadoc "${RUNNER_DOCS_DIR}"
+        ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:javaDocReleaseGeneration
+        cp -rf executorch_android/build/intermediates/java_doc_dir/release/javaDocReleaseGeneration "${RUNNER_DOCS_DIR}/javadoc"
         cd ../..
 
         # If it's main branch, add noindex tag to all .html files to exclude from Google Search indexing.

.github/workflows/lint.yml

@@ -76,8 +76,8 @@ jobs:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 90
       script: |
-        FILES_NEEDS_FORMAT=$(/opt/google-java-format -n extension/android/src/main/java/org/pytorch/executorch/*.java \
-          examples/demo-apps/android/ExecuTorchDemo/app/src/main/java/com/example/executorchdemo/*.java \
+        FILES_NEEDS_FORMAT=$(/opt/google-java-format -n \
+          extension/android/executorch_android/src/main/java/org/pytorch/executorch/*.java \
           examples/demo-apps/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/*.java \
           extension/benchmark/android/benchmark/app/src/main/java/org/pytorch/minibench/*.java)
         if [ -n "$FILES_NEEDS_FORMAT" ]; then

.github/workflows/pull.yml

@@ -60,7 +60,7 @@ jobs:
           - runner: linux.arm64.2xlarge
             docker-image: executorch-ubuntu-22.04-clang12
           # TODO: Need to figure out why buck2 doesnt work on Graviton instances.
-          - runner: linux.arm64.2xlarge 
+          - runner: linux.arm64.2xlarge
             build-tool: buck2
       fail-fast: false
     with:
@@ -420,7 +420,6 @@ jobs:
     permissions:
       id-token: write
       contents: read
-    needs: test-llama-runner-linux
 
   unittest:
     uses: ./.github/workflows/_unittest.yml

backends/qualcomm/README.md

@@ -124,6 +124,7 @@ PRs are always welcome to help improve the codebase in a comprehensive manner. B
     - [shewu-quic](https://github.com/shewu-quic)
     - [chunit-quic](https://github.com/chunit-quic)
     - [winskuo-quic](https://github.com/winskuo-quic)
+    - [DannyYuyang-quic](https://github.com/DannyYuyang-quic)
     - [haowhsu-quic](https://github.com/haowhsu-quic)
 
 Thanks again for your contribution!

backends/qualcomm/_passes/__init__.py

@@ -2,7 +2,6 @@
 from .annotate_quant_attrs import AnnotateQuantAttrs
 from .constant_i64_to_i32 import ConstantI64toI32
 from .convert_bmm_to_matmul import ConvertBmmToMatmul
-from .convert_interpolate_with_upsample2d import ConvertInterpolateWithUpsample2D
 from .convert_to_linear import ConvertToLinear
 from .decompose_any import DecomposeAny
 from .decompose_einsum import DecomposeEinsum
@@ -30,7 +29,6 @@
     AnnotateQuantAttrs,
     ConstantI64toI32,
     ConvertBmmToMatmul,
-    ConvertInterpolateWithUpsample2D,
     RecomposePReLU,
     ConvertToLinear,
     DecomposeAny,

backends/qualcomm/_passes/annotate_quant_attrs.py

@@ -10,6 +10,7 @@
 from executorch.backends.qualcomm.builders.utils import get_parameter, set_parameter
 from executorch.backends.qualcomm.utils.constants import (
     QCOM_AXIS,
+    QCOM_BLOCK_SIZE,
     QCOM_DTYPE,
     QCOM_ENCODING,
     QCOM_QUANT_ATTRS,
@@ -122,13 +123,25 @@ def _dequant_fold_params(self, n, quant_attrs, param):
             scales = self._expand(quant_attrs[QCOM_SCALES], dim, axis)
             offsets = self._expand(quant_attrs[QCOM_ZERO_POINTS], dim, axis)
             param = param.sub(offsets).mul(scales).to(torch.float32).contiguous()
-            set_parameter(param, n.args[0], self.edge_program)
+        elif quant_attrs[QCOM_ENCODING] in [
+            exir_ops.edge.pt2e_quant.dequantize_affine.default
+        ]:
+            param = torch.ops.pt2e_quant.dequantize_affine(
+                param,
+                block_size=quant_attrs[QCOM_BLOCK_SIZE],
+                scale=quant_attrs[QCOM_SCALE],
+                zero_point=quant_attrs[QCOM_ZERO_POINT],
+                input_dtype=quant_attrs[QCOM_DTYPE],
+                quant_min=quant_attrs[QCOM_QUANT_MIN],
+                quant_max=quant_attrs[QCOM_QUANT_MAX],
+                output_dtype=torch.float32,
+            )
         else:
             scale = quant_attrs[QCOM_SCALE]
             offset = quant_attrs[QCOM_ZERO_POINT]
             param = param.sub(offset).mul(scale).to(torch.float32).contiguous()
-            set_parameter(param, n.args[0], self.edge_program)
 
+        set_parameter(param, n.args[0], self.edge_program)
         n.args[0].meta["val"] = param
 
     def _annotate_quant_attrs(

backends/qualcomm/_passes/convert_interpolate_with_upsample2d.py

@@ -39,6 +39,7 @@ class LayoutTransform(ExportPass):
         exir_ops.edge.aten.pixel_shuffle.default,
         exir_ops.edge.aten.pixel_unshuffle.default,
         exir_ops.edge.aten.upsample_bilinear2d.default,
+        exir_ops.edge.aten.upsample_bilinear2d.vec,
         exir_ops.edge.aten.upsample_nearest2d.default,
         exir_ops.edge.aten.upsample_nearest2d.vec,
     }

backends/qualcomm/_passes/layout_transform.py

@@ -6,7 +6,7 @@
 
 import torch
 from executorch.backends.qualcomm.builders.utils import get_parameter
-from executorch.backends.qualcomm.utils.constants import QCOM_ENCODING
+from executorch.backends.qualcomm.utils.constants import QCOM_DTYPE, QCOM_ENCODING
 from executorch.exir.dialects._ops import ops as exir_ops
 from torch._subclasses import FakeTensor
 
@@ -42,6 +42,10 @@ def get_quant_attrs(
                 value = get_parameter(attr_n, edge_program)
         quant_attrs[quant_attr_keys[i - 1]] = value
 
+    # remap key for compatibility - block quantization only
+    if dtype := quant_attrs.get("input_dtype", None):
+        quant_attrs[QCOM_DTYPE] = dtype
+
     quant_attrs[QCOM_ENCODING] = quant_node.target
     return quant_attrs
 
@@ -62,7 +66,6 @@ def get_passes_dependency_for_capture_program():
         AnnotateQuantAttrs,
         ConstantI64toI32,
         ConvertBmmToMatmul,
-        ConvertInterpolateWithUpsample2D,
         ConvertToLinear,
         DecomposeAny,
         DecomposeLinalgVectorNorm,
@@ -85,11 +88,9 @@ def get_passes_dependency_for_capture_program():
             ConvertToLinear,
             RecomposePReLU,
             ConvertBmmToMatmul,
-            ConvertInterpolateWithUpsample2D,
         ],
-        ConstantI64toI32: [ConvertInterpolateWithUpsample2D],
+        ConstantI64toI32: [RemoveRedundancy],
         ConvertBmmToMatmul: [ConvertToLinear],
-        ConvertInterpolateWithUpsample2D: [RemoveRedundancy],
         ConvertToLinear: [RecomposePixelUnshuffle],
         DecomposeAny: [RemoveRedundancy],
         DecomposeLinalgVectorNorm: [RemoveRedundancy],

backends/qualcomm/_passes/utils.py

@@ -54,6 +54,13 @@ enum QuantizeType : byte {
     AXIS_SCALE_OFFSET,
     BW_SCALE_OFFSET,
     BW_AXIS_SCALE_OFFSET,
+    BLOCKWISE_EXPANSION,
+    UNDEFINED,
+}
+
+enum BlockScaleStorageType: byte {
+    BITWIDTH_SCALE_STORAGE_8 = 0,
+    BITWIDTH_SCALE_STORAGE_16,
     UNDEFINED,
 }
 
@@ -72,6 +79,10 @@ table QuantizeParam {
     offsets: [int];
     // used by general quantization
     data: [ScaleOffset];
+    // used by block quantization
+    num_blocks_per_axis: uint;
+    block_scale_storage_type: BlockScaleStorageType;
+    block_scale: [ubyte];
 }
 
 table Tensor {