2024-09-18 nightly release (1e4c316)

Author: pytorchbot

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-00e3eea170ce5db8ea9c62ce5e48f13886cd6d20
+aec9b2ab77389967ef39bb9c10662fd0fe3e185a

.github/workflows/trunk.yml

@@ -223,7 +223,6 @@ jobs:
     strategy:
       matrix:
         dtype: [fp32]
-        build-tool: [buck2, cmake]
         mode: [portable, xnnpack+kv+custom, mps, coreml]
       fail-fast: false
     with:
@@ -235,25 +234,12 @@ jobs:
       script: |
 
         DTYPE=${{ matrix.dtype }}
-        BUILD_TOOL=${{ matrix.build-tool }}
         MODE=${{ matrix.mode }}
 
-        if [[ "${BUILD_TOOL}" == "buck2" ]]; then
-          # TODO: Will add more modes that don't support buck2
-          if [[ "${MODE}" == "mps" ]]; then
-            echo "mps doesn't support buck2."
-            exit 0
-          fi
-          if [[ "${MODE}" == "coreml" ]]; then
-            echo "coreml doesn't support buck2."
-            exit 0
-          fi
-        fi
-
         bash .ci/scripts/setup-conda.sh
 
         # Setup executorch
-        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh "${BUILD_TOOL}"
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh cmake
 
         if [[ "${MODE}" == "mps" ]]; then
           # Install mps delegate
@@ -268,7 +254,7 @@ jobs:
         # Install requirements for export_llama
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash examples/models/llama2/install_requirements.sh
         # Test llama2
-        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_llama.sh stories110M "${BUILD_TOOL}" "${DTYPE}" "${MODE}"
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_llama.sh stories110M cmake "${DTYPE}" "${MODE}"
 
   # # TODO(jackzhxng): Runner consistently runs out of memory before test finishes. Try to find a more powerful runner.
   # test-llava-runner-macos:
@@ -406,6 +392,10 @@ jobs:
         echo "::endgroup::"
 
         echo "::group::Set up HuggingFace Dependencies"
+        if [ -z "$SECRET_EXECUTORCH_HF_TOKEN" ]; then
+          echo "::error::SECRET_EXECUTORCH_HF_TOKEN is empty. For security reason secrets won't be accessible on forked PRs. Please make sure you submit a non-forked PR."
+          exit 1
+        fi
         pip install -U "huggingface_hub[cli]"
         huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
         pip install accelerate sentencepiece

backends/apple/coreml/scripts/install_requirements.sh

@@ -24,7 +24,7 @@ rm -rf "$COREML_DIR_PATH/third-party"
 mkdir "$COREML_DIR_PATH/third-party"
 
 echo "${green}ExecuTorch: Cloning coremltools."
-git clone --depth 1 --branch 8.0b2 "https://github.com/apple/coremltools.git" $COREMLTOOLS_DIR_PATH
+git clone --depth 1 --branch 8.0 "https://github.com/apple/coremltools.git" $COREMLTOOLS_DIR_PATH
 cd $COREMLTOOLS_DIR_PATH
 
 STATUS=$?

backends/arm/test/models/test_mobilenet_v2_arm.py

@@ -84,7 +84,7 @@ def test_mv2_tosa_BI(self):
         )
 
     def test_mv2_u55_BI(self):
-        (
+        tester = (
             ArmTester(
                 self.mv2,
                 example_inputs=self.model_inputs,
@@ -96,4 +96,9 @@ def test_mv2_u55_BI(self):
             .check(list(self.operators_after_quantization))
             .partition()
             .to_executorch()
+            .serialize()
         )
+        if common.is_option_enabled("corstone300"):
+            tester.run_method_and_compare_outputs(
+                atol=1.0, qtol=1, inputs=self.model_inputs
+            )

backends/cadence/aot/utils.py

@@ -104,11 +104,11 @@ def get_ops_count(graph_module: torch.fx.GraphModule) -> Dict[str, int]:
             ):
                 continue
             # If the op is already present, increment the count
-            if get_edge_overload_packet(node.target).__name__ in freq:
-                freq[get_edge_overload_packet(node.target).__name__] += 1
+            if node.target._name in freq:
+                freq[node.target._name] += 1
             # else, add a new entry
             else:
-                freq[get_edge_overload_packet(node.target).__name__] = 1
+                freq[node.target._name] = 1
     return freq
 
 

backends/cadence/hifi/kernels/TARGETS

@@ -0,0 +1,5 @@
+load("targets.bzl", "define_common_targets")
+
+oncall("odai_jarvis")
+
+define_common_targets()

backends/cadence/hifi/kernels/kernels.cpp

@@ -6,9 +6,9 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <executorch/backends/cadence/reference/kernels/kernels.h>
-#include "xa_nnlib_common.h"
-#include "xa_nnlib_common_macros.h"
+#include <executorch/backends/cadence/hifi/kernels/kernels.h>
+#include <xa_nnlib_common.h>
+#include <xa_nnlib_common_macros.h>
 
 namespace impl {
 namespace HiFi {

backends/cadence/hifi/kernels/kernels.h

@@ -8,12 +8,9 @@
 
 #pragma once
 
-#include "inttypes.h"
-#include "stddef.h"
-#include "xa_type_def.h"
-
-/* For NNLIB APIs */
-#include "xa_nnlib_kernels_api.h"
+#include <inttypes.h>
+#include <stddef.h>
+#include <xa_type_def.h>
 
 namespace impl {
 namespace HiFi {

backends/cadence/hifi/kernels/targets.bzl

@@ -0,0 +1,18 @@
+load("@fbsource//tools/build_defs:platform_defs.bzl", "CXX")
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+def define_common_targets():
+    runtime.cxx_library(
+        name = "kernels",
+        srcs = ["kernels.cpp"],
+        exported_headers = [
+            "kernels.h",
+        ],
+        visibility = [
+            "//executorch/backends/cadence/...",
+        ],
+        exported_deps = [
+            "fbsource//third-party/nnlib-hifi4/xa_nnlib:libxa_nnlib_common",
+        ],
+        platforms = CXX,
+    )

backends/cadence/hifi/operators/TARGETS

@@ -0,0 +1,5 @@
+load("targets.bzl", "define_common_targets")
+
+oncall("odai_jarvis")
+
+define_common_targets()