Merge branch 'main' into export-D70507395

Author: zonglinpeng

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-27e35de6c288bffad1b4d18b393579c1d1a95547
+08434df1f2f88c9770e59246caa2ff9c6f613270

.ci/scripts/test_model.sh

@@ -100,6 +100,14 @@ test_model() {
       rm "./${MODEL_NAME}.pte"
       return  # Skip running with portable executor runnner since portable doesn't support Qwen's biased linears.
   fi
+  if [[ "${MODEL_NAME}" == "phi4_mini" ]]; then
+      # Install requirements for export_llama
+      bash examples/models/llama/install_requirements.sh
+      # Test export_llama script: python3 -m examples.models.llama.export_llama.
+      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/phi-4-mini/config.json
+      run_portable_executor_runner
+      rm "./${MODEL_NAME}.pte"
+  fi
 
   # Export a basic .pte and run the model.
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}" "${STRICT}"

.github/workflows/trunk.yml

@@ -229,7 +229,7 @@ jobs:
         # see if we can import the module successfully
         ${CONDA_RUN} python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
 
-    test-static-llama-ane:
+  test-static-llama-ane:
     name: test-static-llama-ane
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:

CMakeLists.txt

@@ -749,9 +749,9 @@ endif()
 
 if(EXECUTORCH_BUILD_PTHREADPOOL
    AND EXECUTORCH_BUILD_CPUINFO
-   AND CMAKE_CXX_STANDARD GREATER_EQUAL 14
 )
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/threadpool)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/parallel)
 endif()
 
 if(EXECUTORCH_BUILD_PYBIND)

backends/arm/scripts/build_executorch_runner.sh

@@ -14,8 +14,9 @@ toolchain_cmake=${et_root_dir}/examples/arm/ethos-u-setup/arm-none-eabi-gcc.cmak
 pte_file=""
 target="ethos-u55-128"
 build_type="Release"
-system_config=""
 bundleio=false
+system_config=""
+memory_mode=""
 build_with_etdump=false
 extra_build_flags=""
 output_folder_set=false
@@ -32,9 +33,12 @@ help() {
     echo "  --pte=<PTE_FILE>                pte file (genrated by the aot_arm_compier from the model to include in the elf"
     echo "  --target=<TARGET>               Target to build and run for Default: ${target}"
     echo "  --build_type=<TYPE>             Build with Release, Debug or RelWithDebInfo, default is ${build_type}"
-    echo "  --system_config=<CONFIG>        System configuration to select from the Vela configuration file (see vela.ini). Default: Ethos_U55_High_End_Embedded for EthosU55 targets, Ethos_U85_SYS_DRAM_Mid for EthosU85 targets."
-    echo "                                     NOTE: If given, this option must match the given target. This option also sets timing adapter values customized for specific hardware, see ./executor_runner/CMakeLists.txt."
     echo "  --bundleio                      Support both pte and Bundle IO bpte using Devtools BundelIO with Input/RefOutput included"
+    echo "  --system_config=<CONFIG>        System configuration to select from the Vela configuration file (see vela.ini). Default: Ethos_U55_High_End_Embedded for EthosU55 targets, Ethos_U85_SYS_DRAM_Mid for EthosU85 targets."
+    echo "                                     NOTE: If given, this option must match the given target. This option along with the memory_mode sets timing adapter values customized for specific hardware, see ./executor_runner/CMakeLists.txt."
+    echo "  --memory_mode=<CONFIG>          Vela memory mode, used for setting the Timing Adapter parameters of the Corstone platforms."
+    echo "                                  Valid values are Shared_Sram(for Ethos-U55, Ethos-U65, Ethos-85), Sram_Only(for Ethos-U55, Ethos-U65, Ethos-U85) or Dedicated_Sram(for Ethos-U65, Ethos-U85)."
+    echo "                                  Default: Shared_Sram for the Ethos-U55 and Sram_Only for the Ethos-U85"
     echo "  --etdump                        Adds Devtools etdump support to track timing, etdump area will be base64 encoded in the log"
     echo "  --extra_build_flags=<FLAGS>     Extra flags to pass to cmake like -DET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE=60000 Default: none "
     echo "  --output=<FOLDER>               Output folder Default: <MODEL>/<MODEL>_<TARGET INFO>.pte"
@@ -49,8 +53,9 @@ for arg in "$@"; do
       --pte=*) pte_file="${arg#*=}";;
       --target=*) target="${arg#*=}";;
       --build_type=*) build_type="${arg#*=}";;
-      --system_config=*) system_config="${arg#*=}";;
       --bundleio) bundleio=true ;;
+      --system_config=*) system_config="${arg#*=}";;
+      --memory_mode=*) memory_mode="${arg#*=}";;
       --etdump) build_with_etdump=true ;;
       --extra_build_flags=*) extra_build_flags="${arg#*=}";;
       --output=*) output_folder="${arg#*=}" ; output_folder_set=true ;;
@@ -83,6 +88,15 @@ then
     fi
 fi
 
+if [[ ${memory_mode} == "" ]]
+then
+    memory_mode="Shared_Sram"
+    if [[ ${target} =~ "ethos-u85" ]]
+    then
+        memory_mode="Sram_Only"
+    fi
+fi
+
 output_folder=$(realpath ${output_folder})
 
 if [[ ${target} == *"ethos-u55"*  ]]; then
@@ -91,7 +105,7 @@ else
     target_cpu=cortex-m85
 fi
 echo "--------------------------------------------------------------------------------"
-echo "Build Arm Baremetal executor_runner for ${target} with ${pte_file} using ${system_config} ${extra_build_flags} to '${output_folder}/cmake-out'"
+echo "Build Arm Baremetal executor_runner for ${target} with ${pte_file} using ${system_config} ${memory_mode} ${extra_build_flags} to '${output_folder}/cmake-out'"
 echo "--------------------------------------------------------------------------------"
 
 cd ${et_root_dir}/examples/arm/executor_runner
@@ -120,6 +134,7 @@ cmake \
     ${build_with_etdump_flags}                  \
     -DPYTHON_EXECUTABLE=$(which python3)        \
     -DSYSTEM_CONFIG=${system_config}            \
+    -DMEMORY_MODE=${memory_mode}                \
     ${extra_build_flags}                        \
     -B ${output_folder}/cmake-out
 

backends/cadence/hifi/operators/op_quantized_conv_out.cpp

@@ -1,7 +1,7 @@
 // (c) Meta Platforms, Inc. and affiliates. Confidential and proprietary.
 
 #include <executorch/backends/cadence/hifi/kernels/kernels.h>
-#include <executorch/backends/cadence/reference/operators/operators.h>
+#include <executorch/backends/cadence/hifi/operators/operators.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
 #define ALIGN_PTR(x, bytes) ((((unsigned)(x)) + (bytes - 1)) & (~(bytes - 1)))
@@ -1108,4 +1108,5 @@ void quantized_conv_per_tensor_out(
 } // namespace native
 } // namespace HiFi
 } // namespace impl
-} // namespace cadence
+} // namespace cadence
+} // namespace cadence

backends/qualcomm/CMakeLists.txt

@@ -126,7 +126,6 @@ add_library(qnn_executorch_logging STATIC)
 add_library(qnn_factory STATIC)
 add_library(qnn_function_interface INTERFACE)
 add_library(qnn_graph STATIC)
-add_library(qnn_header INTERFACE)
 add_library(qnn_implementation STATIC)
 add_library(qnn_logger STATIC)
 add_library(qnn_manager STATIC)
@@ -143,16 +142,12 @@ add_library(utils STATIC)
 # declare dependency
 #
 target_link_libraries(qcir_utils PRIVATE qcir)
-target_link_libraries(wrappers PRIVATE qnn_header qnn_executorch_logging)
-target_link_libraries(qnn_function_interface INTERFACE qnn_header)
+target_link_libraries(wrappers PRIVATE qnn_executorch_logging)
 target_link_libraries(
-  qnn_implementation PRIVATE qnn_function_interface qnn_header
-                             qnn_executorch_logging ${CMAKE_DL_LIBS}
+  qnn_implementation PRIVATE qnn_function_interface qnn_executorch_logging ${CMAKE_DL_LIBS}
 )
-target_link_libraries(qnn_sys_function_interface INTERFACE qnn_header)
 target_link_libraries(
-  qnn_sys_implementation PRIVATE qnn_sys_function_interface qnn_header
-                                 qnn_executorch_logging ${CMAKE_DL_LIBS}
+  qnn_sys_implementation PRIVATE qnn_sys_function_interface qnn_executorch_logging ${CMAKE_DL_LIBS}
 )
 target_link_libraries(qnn_executorch_logging PRIVATE qnn_schema)
 target_link_libraries(qnn_profiler PRIVATE qnn_executorch_logging)
@@ -178,9 +173,7 @@ target_link_libraries(
 )
 
 target_link_libraries(
-  qnn_factory
-  PUBLIC qnn_header
-  PRIVATE qnn_schema qnn_backend qnn_device qnn_context qnn_graph
+  qnn_factory PRIVATE qnn_schema qnn_backend qnn_device qnn_context qnn_graph
           qnn_mem_manager qnn_custom_protocol
 )
 target_link_libraries(

backends/qualcomm/runtime/backends/CMakeLists.txt

@@ -122,54 +122,3 @@ target_sources(
   PUBLIC ${CMAKE_CURRENT_LIST_DIR}/QnnCustomProtocol.h
   PRIVATE ${CMAKE_CURRENT_LIST_DIR}/QnnCustomProtocol.cpp
 )
-
-set(qnn_header_basenames
-    QnnBackend.h
-    QnnCommon.h
-    QnnContext.h
-    QnnDevice.h
-    GPU/QnnGpuBackend.h
-    GPU/QnnGpuCommon.h
-    GPU/QnnGpuContext.h
-    GPU/QnnGpuGraph.h
-    QnnGraph.h
-    HTP/QnnHtpCommon.h
-    HTP/QnnHtpDevice.h
-    HTP/QnnHtpGraph.h
-    HTP/QnnHtpMem.h
-    HTP/QnnHtpPerfInfrastructure.h
-    HTP/QnnHtpProfile.h
-    HTP/QnnHtpProperty.h
-    HTP/QnnHtpSystemContext.h
-    QnnInterface.h
-    QnnLog.h
-    QnnMem.h
-    QnnOpDef.h
-    QnnOpPackage.h
-    QnnProfile.h
-    QnnProperty.h
-    Saver/QnnSaver.h
-    Saver/QnnSaverCommon.h
-    QnnSdkBuildId.h
-    QnnSignal.h
-    QnnTensor.h
-    QnnTypes.h
-    System/QnnSystemCommon.h
-    System/QnnSystemContext.h
-    System/QnnSystemInterface.h
-)
-
-set(QNN_HEADER_DIR_DST ${CMAKE_CURRENT_BINARY_DIR}/QNN/include)
-
-# add the custom commands to copy each headers
-foreach(_qnn_header ${qnn_header_basenames})
-  # copy at generation time to make below target_sources(qnn_header) happy.
-  configure_file(
-    ${QNN_SDK_ROOT}/include/QNN/${_qnn_header}
-    ${QNN_HEADER_DIR_DST}/${_qnn_header} COPYONLY
-  )
-  list(APPEND qnn_header_files ${QNN_HEADER_DIR_DST}/${_qnn_header})
-endforeach()
-
-# qnn_header
-target_sources(qnn_header INTERFACE ${qnn_header_files})

backends/vulkan/targets.bzl

@@ -5,7 +5,13 @@ load("@fbsource//tools/build_defs:platform_defs.bzl", "ANDROID", "CXX", "FBCODE"
 
 
 def get_vulkan_compiler_flags():
-    return ["-Wno-missing-prototypes", "-Wno-global-constructors"]
+    return select({
+        "DEFAULT": [
+            "-Wno-global-constructors",
+            "-Wno-missing-prototypes",
+        ],
+        "ovr_config//os:windows": [],
+    })
 
 def get_labels(no_volk):
     if no_volk:

backends/xnnpack/partition/config/gemm_configs.py

@@ -96,9 +96,9 @@ def _detect_precision(self, node: torch.fx.Node) -> ConfigPrecisionType:
     def _overwrite_precision(self, node: torch.fx.Node):
         precision = self._detect_precision(node)
         if precision not in self.enabled_precision_types:
-            # detected precision is not enabled, lets try to partition it as fp32
+            # detected precision is not enabled, try to partition it as fp32
             if self.enabled_precision_types == [ConfigPrecisionType.FP32]:
-                # if only fp32 is enabled, then we can still partition fp32 gemms
+                # when only fp32 is enabled, then we can still partition fp32 gemms
                 # even with in a quantized graph
                 if precision in [
                     ConfigPrecisionType.STATIC_QUANT,
@@ -107,6 +107,7 @@ def _overwrite_precision(self, node: torch.fx.Node):
                     precision = ConfigPrecisionType.FP32
                     logging.info(f"Overwriting precision, partitioning {node} as FP32")
                     return True, precision
+
         return False, precision
 
     def get_deps(
@@ -226,8 +227,11 @@ def _get_bias_deps(
         self, node: torch.fx.Node, ep: ExportedProgram, precision: ConfigPrecisionType
     ) -> Tuple[bool, List[torch.fx.Node]]:
         gemm_deps = []
-        if precision == ConfigPrecisionType.FP32 and self.force_fp32_dynamic_linear:
-            # if force force_fp32_dynamic_linear is enabled, then we
+        if (
+            precision == ConfigPrecisionType.FP32
+            and self.force_non_static_weights_for_f32_linear
+        ):
+            # if force_non_static_weights_for_f32_linear is enabled, then we
             # do not partition the weight node
             return (True, gemm_deps)
 
@@ -305,8 +309,11 @@ def get_original_aten(self) -> Optional[torch._ops.OpOverload]:
     def _get_weight_deps(
         self, node: torch.fx.Node, ep: ExportedProgram, precision: ConfigPrecisionType
     ) -> Tuple[bool, List[torch.fx.Node]]:
-        if precision == ConfigPrecisionType.FP32 and self.force_fp32_dynamic_linear:
-            # if force fp32_dynamic_linear is enabled, then we
+        if (
+            precision == ConfigPrecisionType.FP32
+            and self.force_non_static_weights_for_f32_linear
+        ):
+            # if force_non_static_weights_for_f32_linear is enabled, then we
             # do not partition the weight node
             return (True, [])
 
@@ -412,9 +419,11 @@ def __init__(self, **kwargs):
     def _get_weight_deps(
         self, node: torch.fx.Node, ep: ExportedProgram, precision: ConfigPrecisionType
     ) -> Tuple[bool, List[torch.fx.Node]]:
-        # TODO(maxren, T210537195):
-        if precision == ConfigPrecisionType.FP32 and self.force_fp32_dynamic_linear:
-            # if force fp32_dynamic_linear is on and we detected this as fp32, then we
+        if (
+            precision == ConfigPrecisionType.FP32
+            and self.force_non_static_weights_for_f32_linear
+        ):
+            # if force_non_static_weights_for_f32_linear is on and we detected this as fp32, then we
             # do not partition the weight node
             return (True, [])
 
@@ -501,11 +510,11 @@ def find_partition_args(input_node):
         node.args = old_args
         node.users = old_users
 
-        # When using force_fp32_dynamic_linear, we want to get_deps to overwrite the source partition nodes.
+        # When using force_non_static_weights_for_f32_linear, we want to get_deps to overwrite the source partition nodes.
         # Else we want to be greedy.
         ret_deps = (
             list(set(deps) & set(src_partition.nodes))
-            if self.force_fp32_dynamic_linear
+            if self.force_non_static_weights_for_f32_linear
             else list(set(deps) | set(src_partition.nodes))
         )
 
@@ -531,8 +540,11 @@ def __init__(self, **kwargs):
     def _get_weight_deps(
         self, node: torch.fx.Node, ep: ExportedProgram, precision: ConfigPrecisionType
     ) -> Tuple[bool, List[torch.fx.Node]]:
-        if precision == ConfigPrecisionType.FP32 and self.force_fp32_dynamic_linear:
-            # if force fp32_dynamic_linear is on and we detected this as fp32, then we
+        if (
+            precision == ConfigPrecisionType.FP32
+            and self.force_non_static_weights_for_f32_linear
+        ):
+            # if force_non_static_weights_for_f32_linear is on and we detected this as fp32, then we
             # do not partition the weight node
             return (True, [])
 

backends/xnnpack/partition/config/xnnpack_config.py

@@ -41,7 +41,9 @@ def __init__(self, **kwargs):
         super().__init__()
         self.enabled_precision_types = self.supported_precision_types()
         # Flag used in GEMMConfig()
-        self.force_fp32_dynamic_linear = kwargs.get("force_fp32_dynamic_linear", False)
+        self.force_non_static_weights_for_f32_linear = kwargs.get(
+            "force_non_static_weights_for_f32_linear", False
+        )
 
     def get_partition(
         self, node: torch.fx.Node, ep: ExportedProgram

backends/xnnpack/test/ops/test_linear.py

@@ -948,7 +948,7 @@ def test_linear_qd8_as_fp32(self):
                 },
             )
 
-    def test_linear_fp32_with_force_as_mm(self):
+    def test_linear_with_force_non_static_weights_for_f32_linear(self):
         def check_signature(
             signature: ExportGraphSignature,
             force_flag: bool,
@@ -981,7 +981,7 @@ def check_signature(
                     inputs = module.get_inputs()
                     tester = Tester(module, inputs).export()
                     partitioner = XnnpackPartitioner(
-                        force_fp32_dynamic_linear=force_flag
+                        force_non_static_weights_for_f32_linear=force_flag
                     )
                     if legacy_mode:
                         tester.to_edge()

backends/xnnpack/test/ops/test_lstm.py

@@ -43,18 +43,20 @@ def test_fp32_lstm(self):
             .run_method_and_compare_outputs()
         )
 
-    def test_fp32_lstm_force_dynamic_linear(self):
+    def test_lstm_with_force_non_static_weights_for_f32_linear(self):
         (
             Tester(self.LSTMLinear(32, 32, 10), (torch.rand(1, 32, 32),))
             .export()
             .to_edge_transform_and_lower(
                 ToEdgeTransformAndLower(
-                    partitioners=[XnnpackPartitioner(force_fp32_dynamic_linear=True)]
+                    partitioners=[
+                        XnnpackPartitioner(force_non_static_weights_for_f32_linear=True)
+                    ]
                 )
             )
             .check_not(["executorch_exir_dialects_edge__ops_aten_addmm_default"])
             # Weights are supplied as input to linears
-            # Biases are not owned by delegates when force_fp32_dynamic_linear is set
+            # Biases are not owned by delegates when force_non_static_weights_for_f32_linear is set
             .check(["p_lstm_weight_hh_l0", "p_lstm_weight_ih_l0", "p_lstm_bias"])
             .to_executorch()
             .serialize()

build/build_android_library.sh

@@ -178,7 +178,9 @@ collect_artifacts_to_be_uploaded() {
 }
 
 main() {
-  BUILD_AAR_DIR="$(mktemp -d)"
+  if [[ -z "${BUILD_AAR_DIR:-}" ]]; then
+    BUILD_AAR_DIR="$(mktemp -d)"
+  fi
   export BUILD_AAR_DIR
   if [ -z "$ANDROID_ABIS" ]; then
     ANDROID_ABIS=("arm64-v8a" "x86_64")