Update

[ghstack-poisoned]

Author: swolchok

.ci/scripts/test_qnn_static_llama.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# 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.
+
+set -exu
+
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
+export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
+export PYTHONPATH=".."
+cp schema/program.fbs exir/_serialize/program.fbs
+cp schema/scalar_type.fbs exir/_serialize/scalar_type.fbs
+cp -f build-x86/backends/qualcomm/PyQnnManagerAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
+cp -f build-x86/backends/qualcomm/PyQnnWrapperAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+
+which "${PYTHON_EXECUTABLE}"
+
+# Although static llama CI does not require graphviz, it is required by test_qnn_delegate.py
+pip install graphviz
+
+# Download stories llama110m artifacts
+download_stories_model_artifacts
+echo "Creating tokenizer.bin"
+$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+
+set +e
+# Compile only as weight sharing is not applicable on x86
+$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
+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
+exit_code2=$?
+
+# Check the exit codes and print messages
+if [ $exit_code1 -ne 0 ]; then
+    echo "Static Llama compile only with weight sharing test failed. $exit_code1."
+fi
+
+if [ $exit_code2 -ne 0 ]; then
+    echo "Static Llama accuracy test failed. $exit_code2."
+fi
+
+# Return failure if either program failed
+if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ]; then
+    exit 1
+else
+    exit 0
+fi
+set -e

.github/workflows/pull.yml

@@ -212,7 +212,7 @@ jobs:
       docker-image: executorch-ubuntu-22.04-clang12
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 180
+      timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
@@ -437,6 +437,39 @@ jobs:
         # 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}"
 
+  test-static-llama-qnn-linux:
+    name: test-static-llama-qnn-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-qnn-sdk
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 180
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        BUILD_TOOL="cmake"
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
+
+        # Setup executorch
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
+
+        # Setup install_requirements for llama
+        PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
+        
+        # Test static llama weight sharing and accuracy
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_qnn_static_llama.sh
+
   test-qnn-models-linux:
     name: test-qnn-models-linux
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -526,7 +559,7 @@ jobs:
       docker-image: executorch-ubuntu-22.04-clang12
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 180
+      timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")

.lintrunner.toml

@@ -78,6 +78,8 @@ exclude_patterns = [
     # File contains @generated
     'extension/llm/custom_ops/spinquant/fast_hadamard_transform_special.h',
     'extension/llm/custom_ops/spinquant/test/fast_hadamard_transform_special_unstrided_cpu.h',
+    # Want to be able to keep c10 in sync with PyTorch core.
+    'runtime/core/portable_type/c10/**',
 ]
 command = [
     'python',
@@ -261,6 +263,8 @@ exclude_patterns = [
     'extension/**',
     'kernels/optimized/**',
     'runtime/core/exec_aten/**',
+    # Want to be able to keep c10 in sync with PyTorch core.
+    'runtime/core/portable_type/c10/**',
     'runtime/executor/tensor_parser_aten.cpp',
     'scripts/**',
     'test/**',

CMakeLists.txt

@@ -182,6 +182,10 @@ option(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER "Build the Data Loader extension"
        OFF
 )
 
+option(EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR "Build the Flat Tensor extension"
+       OFF
+)
+
 option(EXECUTORCH_BUILD_EXTENSION_MODULE "Build the Module extension" OFF)
 
 option(EXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL "Build the Runner Util extension"
@@ -240,6 +244,9 @@ cmake_dependent_option(
   "NOT EXECUTORCH_BUILD_ARM_BAREMETAL" OFF
 )
 
+if(EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR)
+  set(EXECUTORCH_BUILF_EXTENSION_DATA_LOADER ON)
+endif()
 
 if(EXECUTORCH_BUILD_EXTENSION_TRAINING)
   set(EXECUTORCH_BUILD_EXTENSION_TENSOR ON)
@@ -366,7 +373,7 @@ if(NOT "${_repo_dir_name}" STREQUAL "executorch")
       "fix for this restriction."
   )
 endif()
-set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/..)
+set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/.. ${CMAKE_CURRENT_SOURCE_DIR}/runtime/core/portable_type)
 
 #
 # The `_<target>_srcs` lists are defined by including ${EXECUTORCH_SRCS_FILE}.
@@ -549,6 +556,7 @@ endif()
 target_include_directories(
   executorch_core PUBLIC ${_common_include_directories}
 )
+target_compile_definitions(executorch_core PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch_core PUBLIC ${_common_compile_options})
 if(MAX_KERNEL_NUM)
   target_compile_definitions(
@@ -569,6 +577,7 @@ if(EXECUTORCH_BUILD_PYBIND AND APPLE)
   target_include_directories(
     executorch_core_shared PUBLIC ${_common_include_directories}
   )
+  target_compile_definitions(executorch_core_shared PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
   target_compile_options(
     executorch_core_shared PUBLIC ${_common_compile_options}
   )
@@ -589,6 +598,7 @@ endif()
 add_library(executorch ${_executorch__srcs})
 target_link_libraries(executorch PRIVATE executorch_core)
 target_include_directories(executorch PUBLIC ${_common_include_directories})
+target_compile_definitions(executorch PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch PUBLIC ${_common_compile_options})
 target_link_options_shared_lib(executorch)
 
@@ -622,6 +632,12 @@ endif()
 
 # Install `executorch` library as well as `executorch-config.cmake` under
 # ${CMAKE_INSTALL_PREFIX}/
+install(DIRECTORY runtime/core/  DESTINATION include/executorch/runtime/core FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY runtime/kernel/  DESTINATION include/executorch/runtime/kernel FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY runtime/platform/  DESTINATION include/executorch/runtime/platform FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY extension/kernel_util/  DESTINATION include/executorch/extension/kernel_util FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY extension/tensor/  DESTINATION include/executorch/extension/tensor FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY extension/threadpool/  DESTINATION include/executorch/extension/threadpool FILES_MATCHING PATTERN "*.h")
 install(
   TARGETS executorch executorch_core
   DESTINATION lib
@@ -694,6 +710,11 @@ if(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/data_loader)
 endif()
 
+if(EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/flat_tensor)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/flat_tensor/serialize)
+endif()
+
 if(EXECUTORCH_BUILD_EXTENSION_MODULE)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/module)
 endif()
@@ -780,6 +801,8 @@ if(EXECUTORCH_BUILD_PYBIND)
   target_include_directories(
     util PUBLIC ${_common_include_directories} ${TORCH_INCLUDE_DIRS}
   )
+  target_compile_definitions(util PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
+
   target_compile_options(util PUBLIC ${_pybind_compile_options})
   target_link_libraries(util PRIVATE torch c10 executorch extension_tensor)
 

CONTRIBUTING.md

@@ -215,6 +215,14 @@ must work with threading**
 
 ## Testing
 
+### Running Tests Locally
+
+CI is run automatically on all pull requests. However, if you want to run tests locally, here are some example commands (not exhaustive):
+
+- The `sh test/build_size_test.sh` script will compile the C++runtime along with portable kernels.
+- The `test/run_oss_cpp_tests.sh` script will build and run C++ tests locally
+- Running `pytest` from the root directory will run Python tests locally.
+
 ### Writing Tests
 To help keep code quality high, ExecuTorch uses a combination of unit tests and
 end-to-end (e2e) tests. If you add a new feature or fix a bug, please add tests
@@ -229,8 +237,6 @@ If it's not clear how to add a test for your PR, take a look at the blame for
 the code you're modifying and find an author who has more context. Ask them
 for their help in the PR comments.
 
-The `test/run_oss_cpp_tests.sh` script will build and run C++ tests locally.
-
 ### Continuous Integration
 See https://hud.pytorch.org/hud/pytorch/executorch/main for the current state of
 the CI (continuous integration) jobs. If `main` is broken, consider rebasing

README.md

@@ -7,7 +7,7 @@
 <div align="center">
   <a href="https://github.com/pytorch/executorch/graphs/contributors"><img src="https://img.shields.io/github/contributors/pytorch/executorch?style=for-the-badge&color=blue" alt="Contributors"></a>
   <a href="https://github.com/pytorch/executorch/stargazers"><img src="https://img.shields.io/github/stars/pytorch/executorch?style=for-the-badge&color=blue" alt="Stargazers"></a>
-  <a href="https://discord.gg/MeacgB7A"><img src="https://img.shields.io/badge/Discord-Join%20Us-purple?logo=discord&logoColor=white&style=for-the-badge" alt="Join our Discord community"></a>
+  <a href="https://discord.gg/Dh43CKSAdc"><img src="https://img.shields.io/badge/Discord-Join%20Us-purple?logo=discord&logoColor=white&style=for-the-badge" alt="Join our Discord community"></a>
   <a href="https://pytorch.org/executorch/stable/index.html"><img src="https://img.shields.io/badge/Documentation-000?logo=googledocs&logoColor=FFE165&style=for-the-badge" alt="Check out the documentation"></a>
   <hr>
 </div>
@@ -55,11 +55,11 @@ To get started you can:
 ## Feedback and Engagement
 
 We welcome any feedback, suggestions, and bug reports from the community to help
-us improve our technology. Check out the [Discussion Board](https://github.com/pytorch/executorch/discussions) or chat real time with us on [Discord](https://discord.gg/MeacgB7A)
+us improve our technology. Check out the [Discussion Board](https://github.com/pytorch/executorch/discussions) or chat real time with us on [Discord](https://discord.gg/Dh43CKSAdc)
 
 ## Contributing
 
-We welcome contributions. To get started review the [guidelines](CONTRIBUTING.md) and chat with us on [Discord](https://discord.gg/MeacgB7A)
+We welcome contributions. To get started review the [guidelines](CONTRIBUTING.md) and chat with us on [Discord](https://discord.gg/Dh43CKSAdc)
 
 
 ## Directory Structure

backends/apple/coreml/CMakeLists.txt

@@ -134,6 +134,8 @@ target_include_directories(
   coremldelegate PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/runtime/util
 )
 target_include_directories(coremldelegate PRIVATE ${EXECUTORCH_ROOT}/..)
+target_include_directories(coremldelegate PRIVATE ${EXECUTORCH_ROOT}/runtime/core/portable_type)
+target_compile_definitions(coremldelegate PRIVATE C10_USING_CUSTOM_GENERATED_MACROS)
 target_link_libraries(coremldelegate PRIVATE executorch_core)
 
 if(EXECUTORCH_BUILD_DEVTOOLS)

backends/apple/coreml/runtime/workspace/executorchcoreml.xcodeproj/project.pbxproj

@@ -830,6 +830,7 @@
 				GCC_OPTIMIZATION_LEVEL = 0;
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
+                                        "C10_USING_CUSTOM_GENERATED_MACROS",
 					"$(inherited)",
 				);
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
@@ -911,6 +912,7 @@
 				DEVELOPMENT_TEAM = "";
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
+                                        "C10_USING_CUSTOM_GENERATED_MACROS",
 					"ET_EVENT_TRACER_ENABLED=1",
 					"$(inherited)",
 				);
@@ -920,6 +922,7 @@
 					"$(SRCROOT)/../kvstore",
 					"$(SRCROOT)/../inmemoryfs",
 					"$(SRCROOT)/../include",
+					"$(SRCROOT)/../include/executorch/runtime/core/portable_type",
 					"$(SRCROOT)/../sdk",
 					"$(SRCROOT)/../util",
 					"$(SRCROOT)/../../third-party/nlohmann_json/single_include",
@@ -951,6 +954,7 @@
 					"$(SRCROOT)/../kvstore",
 					"$(SRCROOT)/../inmemoryfs",
 					"$(SRCROOT)/../include",
+					"$(SRCROOT)/../include/executorch/runtime/core/portable_type",
 					"$(SRCROOT)/../sdk",
 					"$(SRCROOT)/../util",
 					"$(SRCROOT)/../../third-party/nlohmann_json/single_include",

backends/arm/CMakeLists.txt

@@ -14,7 +14,8 @@ endif()
 
 include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 
-set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+set(_common_include_directories ${EXECUTORCH_ROOT}/.. ${EXECUTORCH_ROOT}/runtime/core/portable_type)
+add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 
 # Third-party folder and Ethos-U driver inclued
 set(THIRD_PARTY_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/third-party")

backends/arm/arm_partitioner.py

@@ -7,14 +7,14 @@
 
 import logging
 import os
-from typing import Callable, final, List, Optional, Tuple
+from typing import Callable, final, List, Optional, Sequence, Tuple
 
 import torch
 from executorch.backends.arm.arm_backend import (  # type: ignore[attr-defined]
     ArmBackend,
 )  # usort: skip
 from executorch.backends.arm.operator_support.tosa_supported_operators import (
-    TOSASupportedOperators,
+    tosa_support_factory,
 )
 from executorch.backends.arm.tosa_specification import TosaSpecification
 from executorch.exir.backend.compile_spec_schema import CompileSpec
@@ -27,6 +27,8 @@
 from executorch.exir.dialects._ops import ops as exir_ops
 from torch.export.exported_program import ExportedProgram
 from torch.fx.passes.infra.partitioner import CapabilityBasedPartitioner
+from torch.fx.passes.operator_support import OperatorSupportBase
+
 
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.WARNING)
@@ -54,8 +56,13 @@ def is_dequant_node(node: torch.fx.node.Node) -> bool:
 
 @final
 class ArmPartitioner(Partitioner):
-    def __init__(self, compile_spec: List[CompileSpec]) -> None:
+    def __init__(
+        self,
+        compile_spec: List[CompileSpec],
+        additional_checks: Optional[Sequence[OperatorSupportBase]] = None,
+    ) -> None:
         self.delegation_spec = DelegationSpec(ArmBackend.__name__, compile_spec)
+        self.additional_checks = additional_checks
 
     def partition(self, exported_program: ExportedProgram) -> PartitionResult:
         # Run the CapabilityBasedPartitioner to return the largest possible
@@ -72,7 +79,7 @@ def partition(self, exported_program: ExportedProgram) -> PartitionResult:
 
         capability_partitioner = CapabilityBasedPartitioner(
             exported_program.graph_module,
-            TOSASupportedOperators(tosa_spec),
+            tosa_support_factory(tosa_spec, self.additional_checks),
             allows_single_node_partition=True,
         )
         partition_list = capability_partitioner.propose_partitions()

backends/arm/operator_support/convolution_support.py

@@ -24,7 +24,7 @@ class ConvolutionSupported(SupportedTOSAOperatorCheck):
         TosaSpecification.create_from_string("TOSA-0.80+MI"),
     ]
 
-    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+    def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
 
         # Not implemented
         transposed = cast(bool, node.args[6])

backends/arm/operator_support/pool_2d_support.py

@@ -43,7 +43,7 @@ class AvgPool2dSupported(SupportedTOSAOperatorCheck):
         TosaSpecification.create_from_string("TOSA-0.80+MI"),
     ]
 
-    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+    def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
         if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
             return True
 
@@ -73,7 +73,7 @@ class MaxPool2dSupported(SupportedTOSAOperatorCheck):
         TosaSpecification.create_from_string("TOSA-0.80+MI"),
     ]
 
-    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+    def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
         if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
             return True