Update base for Update on "Introduce data schema to store raw tensors"

Differential Revision: [D65156641](https://our.internmc.facebook.com/intern/diff/D65156641)

[ghstack-poisoned]

Author: lucylq

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-bd5482c7c3e1197e10c46ff739027f917d9c1fcc
+c8a648d4dffb9f0133ff4a2ea0e660b42105d3ad

.ci/docker/common/install_clang.sh

@@ -13,7 +13,7 @@ install_ubuntu() {
   apt-get install -y --no-install-recommends clang-"$CLANG_VERSION"
   apt-get install -y --no-install-recommends llvm-"$CLANG_VERSION"
   # Also require LLD linker from llvm and libomp to build PyTorch from source
-  apt-get install -y lld "libomp-${CLANG_VERSION}-dev"
+  apt-get install -y lld "libomp-${CLANG_VERSION}-dev" "libc++-${CLANG_VERSION}-dev"
 
   # Use update-alternatives to make this version the default
   update-alternatives --install /usr/bin/clang clang /usr/bin/clang-"$CLANG_VERSION" 50

.ci/docker/requirements-ci.txt

@@ -1,5 +1,5 @@
 mpmath==1.3.0
-numpy==1.22.0; python_version == '3.10'
+numpy==1.21.3; python_version == '3.10'
 numpy==1.23.2; python_version == '3.11'
 numpy; python_version >= '3.12'
 PyYAML==6.0.1

.ci/scripts/build-qnn-sdk.sh

@@ -42,7 +42,7 @@ set_up_aot() {
 
   # Workaround for fbs files in exir/_serialize
   cp schema/program.fbs exir/_serialize/program.fbs
-  cp schema/scalar_type.fbs exir/_serialize/scalar_type.fbs
+  cp schema/common.fbs exir/_serialize/common.fbs
 }
 
 build_qnn_backend

.ci/scripts/setup-linux.sh

@@ -19,6 +19,7 @@ else
 fi
 
 # As Linux job is running inside a Docker container, all of its dependencies
-# have already been installed
-install_executorch
+# 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}"

.ci/scripts/setup-qnn-deps.sh

@@ -31,8 +31,9 @@ install_qnn() {
 }
 
 setup_libc++() {
+  clang_version=$1
   sudo apt-get update
-  pkgs_to_check=('libc++-dev')
+  pkgs_to_check=("libc++-${clang_version}-dev")
   j=0
   while [ $j -lt ${#pkgs_to_check[*]} ]; do
     install_status=$(verify_pkg_installed ${pkgs_to_check[$j]})
@@ -47,5 +48,6 @@ setup_libc++() {
   done
 }
 
-setup_libc++
+# This needs to match with the clang version from the Docker image
+setup_libc++ 12
 install_qnn

.ci/scripts/test_llama.sh

@@ -79,7 +79,7 @@ if [[ "${MODE}" =~ .*qnn.* ]]; then
   export LD_LIBRARY_PATH="${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 schema/common.fbs exir/_serialize/common.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
 

.ci/scripts/test_llama_runner_eager.sh

@@ -42,11 +42,12 @@ run_and_verify() {
 	-d fp32 \
 	--max_seq_length 32 \
 	--temperature 0 \
+    --show_tokens \
 	--prompt "Once upon a time," > result.txt
 
     # Verify result.txt
     RESULT=$(cat result.txt)
-    EXPECTED_RESULT="there was a little girl"
+    EXPECTED_RESULT="727, 471, 263, 2217, 7826, 4257, 365, 2354, 29889, 2296, 18012, 304, 1708, 5377, 297, 278, 6575, 845, 457, 29889, 3118, 2462, 29892, 1183, 4446, 263"
     if [[ "${RESULT}" == *"${EXPECTED_RESULT}"* ]]; then
         echo "Actual result: ${RESULT}"
         echo "Success"

.ci/scripts/utils.sh

@@ -20,8 +20,11 @@ install_executorch() {
   which pip
   # Install executorch, this assumes that Executorch is checked out in the
   # current directory.
-  # TODO(T199538337): clean up install scripts to use install_requirements.sh
-  ./install_requirements.sh --pybind xnnpack
+  if [[ "${1:-}" == "use-pt-pinned-commit" ]]; then
+    ./install_requirements.sh --pybind xnnpack --use-pt-pinned-commit
+  else
+    ./install_requirements.sh --pybind xnnpack
+  fi
   # Just print out the list of packages for debugging
   pip list
 }

.github/pull_request_template.md

@@ -0,0 +1,9 @@
+### Summary
+[PLEASE REMOVE] See [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests) for ExecuTorch PR guidelines.
+
+[PLEASE REMOVE] If this PR closes an issue, please add a `Fixes #<issue-id>` line.
+
+[PLEASE REMOVE] If this PR introduces a fix or feature that should be the upcoming release notes, please add a "Release notes: <area>" label. For a list of available release notes labels, check out [CONTRIBUTING.md's Pull Requests](https://github.com/pytorch/executorch/blob/main/CONTRIBUTING.md#pull-requests).
+
+### Test plan
+[PLEASE REMOVE] How did you test this PR? Please write down any manual commands you used and note down tests that you have written if applicable.

.github/workflows/_unittest.yml

@@ -37,6 +37,9 @@ jobs:
         CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
         .ci/scripts/setup-linux.sh cmake
 
+        # Install llama3_2_vision dependencies.
+        PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
+
         # Run pytest with coverage
         pytest -n auto --cov=./ --cov-report=xml
         # Run gtest
@@ -67,6 +70,10 @@ jobs:
         ${CONDA_RUN} --no-capture-output \
         .ci/scripts/setup-macos.sh cmake
 
+        # Install llama3_2_vision dependencies.
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
+        ./examples/models/llama3_2_vision/install_requirements.sh
+
         # Run pytest with coverage
         ${CONDA_RUN} pytest -n auto --cov=./ --cov-report=xml
         # Run gtest

.github/workflows/ghstack_land.yml

@@ -5,6 +5,7 @@ on:
     branches:
       - 'gh/cccclai/[0-9]+/base'
       - 'gh/dbort/[0-9]+/base'
+      - 'gh/dvorjackz/[0-9]+/base'
       - 'gh/guangy10/[0-9]+/base'
       - 'gh/helunwencser/[0-9]+/base'
       - 'gh/jorgep31415/[0-9]+/base'

.github/workflows/trunk.yml

@@ -137,6 +137,7 @@ jobs:
       docker-image: executorch-ubuntu-22.04-arm-sdk
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
@@ -162,6 +163,7 @@ jobs:
       docker-image: executorch-ubuntu-22.04-arm-sdk
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")

.gitmodules

@@ -64,3 +64,6 @@
 [submodule "third-party/pybind11"]
 	path = third-party/pybind11
 	url = https://github.com/pybind/pybind11.git
+[submodule "third-party/ao"]
+	path = third-party/ao
+	url = https://github.com/pytorch/ao.git

CMakeLists.txt

@@ -721,10 +721,15 @@ if(EXECUTORCH_BUILD_PYBIND)
       -fPIC
       -frtti
       -fexceptions
-      # libtorch is built with the old ABI, so we need to do the same for any
-      # .cpp files that include torch, c10, or ATen targets.
-      -D_GLIBCXX_USE_CXX11_ABI=0
   )
+  if(EXECUTORCH_DO_NOT_USE_CXX11_ABI)
+      # libtorch is built with the old ABI, so we need to do the same for any
+      # .cpp files that include torch, c10, or ATen targets. Note that PyTorch
+      # nightly binary is built with _GLIBCXX_USE_CXX11_ABI set to 0 while its
+      # CI build sets this to 1 (default)
+    list(APPEND _pybind_compile_options -D_GLIBCXX_USE_CXX11_ABI=0)
+  endif()
+
   # util lib
   add_library(
     util ${CMAKE_CURRENT_SOURCE_DIR}/extension/evalue_util/print_evalue.cpp

CONTRIBUTING.md

@@ -242,14 +242,27 @@ for basics.
    - Give the PR a clear and thorough description. Don't just describe what the PR
      does: the diff will do that. Explain *why* you are making this change, in a
      way that will make sense to someone years from now.
-   - Add the line `Test Plan:` (with that spelling, capitalization, and trailing
-     colon character), followed by lines containing repeatable instructions for
+   - Explain how you have tested your changes by including repeatable instructions for
      testing the PR.
      - If you added tests, this can be as simple as the command you used to run the
        tests.
      - If you tested the PR manually, include the steps and the outputs. Help a
        future editor understand how to test the code that you're modifying
        today.
+   - If your PR contains or is representative of a feature/bug fix that should be
+     called out in the release notes, please add a label for "Release notes: \<area\>",
+	 where \<area\> describes which part of ExecuTorch the change pertains to, e.g.
+	 "Release notes: runtime". Here are all of the categories:
+     - `Release notes: runtime`: changes related to the core runtime which loads the program methods, initializes delegates, and runs the lowered graph.
+     - `Release notes: exir`: changes to any internal representations, such as any edge-related dialects. Also any changes to passes that may modify the exir, such as memory planning.
+     - `Release notes: quantization`: changes to quantization.
+     - `Release notes: ops & kernels`: changes to the opset and any new / changed kernel implementations.
+     - `Release notes: api`: changes to public facing apis (any interfaces, pybinded runtime methods, etc.).
+     - `Release notes: backends`: changes to any of the backend delegates.
+     - `Release notes: build`: changes related to the build system, including major dependency upgrades, notable build flags, optimizations, etc.
+     - `Release notes: devtools`: changes to any of ExecuTorch's developer tools, for example the debugger & profiler.
+     - `Release notes: examples`: changes to any code under `examples/`.
+     - `Release notes: misc`: anything notable that doesn't belong in the above categories.
    - See https://github.com/pytorch/executorch/pull/3612 for an example PR that
      follows this advice.
 1. Before asking for a review, ensure that all [CI (continuous integration)

README.md

@@ -43,6 +43,11 @@ We recommend using the latest release tag from the
 See [CONTRIBUTING.md](CONTRIBUTING.md) for details about issues, PRs, code
 style, CI jobs, and other development topics.
 
+To connect with us and other community members, we invite you to join PyTorch Slack community by filling out this [form](https://docs.google.com/forms/d/e/1FAIpQLSeADnUNW36fjKjYzyHDOzEB_abKQE9b6gqqW9NXse6O0MWh0A/viewform). Once you've joined, you can:
+* Head to the `#executorch-general` channel for general questions, discussion, and community support.
+* Join the `#executorch-contributors` channel if you're interested in contributing directly to project development.
+
+
 ## Directory Structure
 
 ```

backends/apple/coreml/README.md

@@ -65,7 +65,7 @@ To quantize a Program in a Core ML favored way, the client may utilize **CoreMLQ
 import torch
 import executorch.exir
 
-from torch._export import capture_pre_autograd_graph
+from torch.export import export_for_training
 from torch.ao.quantization.quantize_pt2e import (
     convert_pt2e,
     prepare_pt2e,
@@ -93,7 +93,7 @@ class Model(torch.nn.Module):
 source_model = Model()
 example_inputs = (torch.randn((1, 3, 256, 256)), )
 
-pre_autograd_aten_dialect = capture_pre_autograd_graph(model, example_inputs)
+pre_autograd_aten_dialect = export_for_training(model, example_inputs).module()
 
 quantization_config = LinearQuantizerConfig.from_dict(
     {

backends/arm/TARGETS

@@ -1,3 +1,4 @@
+# @noautodeps
 load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
 
 python_library(
@@ -69,6 +70,18 @@ python_library(
     ],
 )
 
+python_library(
+    name = "tosa_specification",
+    srcs = [
+        "tosa_specification.py",
+    ],
+    typing = True,
+    deps = [
+        "fbsource//third-party/pypi/packaging:packaging",
+        "//executorch/exir/backend:compile_spec_schema",
+    ],
+)
+
 python_library(
     name = "tosa_utils",
     srcs = [

backends/arm/_passes/TARGETS

@@ -7,6 +7,7 @@ python_library(
     deps = [
         "//executorch/backends/arm:tosa_quant_utils",
         "//executorch/backends/arm:tosa_utils",
+        "//executorch/backends/xnnpack/_passes:xnnpack_passes",
         "//executorch/exir:lib",
     ],
 )

backends/arm/_passes/annotate_channels_last_dim_order_pass.py

@@ -14,7 +14,7 @@
     get_first_fake_tensor,
     insert_q_dq_pair,
 )
-from executorch.backends.arm.tosa_quant_utils import dq_op, q_op
+from executorch.backends.arm.tosa_quant_utils import dq_op, q_op, register_passable_op
 from executorch.backends.arm.tosa_utils import is_consumer_node_depthwise_conv2d
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass, PassResult
@@ -42,6 +42,9 @@ def _transpose_impl(*args, **kwargs):
     return args[0]
 
 
+register_passable_op(torch.ops.passthrough_to_tosa._transpose)
+
+
 class AnnotateChannelsLastDimOrder(ExportPass):
     """
     Annotates each node with a tosa_dim_order. tosa_dim_order can be seen as a channels-last dim-order

backends/arm/_passes/arm_pass_manager.py

@@ -23,6 +23,7 @@
 from executorch.backends.arm._passes.decompose_layernorm_pass import (
     DecomposeLayerNormPass,
 )
+from executorch.backends.arm._passes.decompose_linear_pass import DecomposeLinearPass
 from executorch.backends.arm._passes.decompose_meandim_pass import DecomposeMeanDimPass
 from executorch.backends.arm._passes.decompose_softmaxes_pass import (
     DecomposeSoftmaxesPass,
@@ -43,6 +44,7 @@
 from executorch.backends.arm._passes.unsqueeze_scalar_placeholders_pass import (
     UnsqueezeScalarPlaceholdersPass,
 )
+from executorch.backends.xnnpack._passes.remove_getitem_op import RemoveGetItemPass
 from executorch.exir import ExportedProgram
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.pass_manager import PassManager
@@ -58,6 +60,7 @@ def transform_to_backend_pipeline(
     ):
         """Apply passes before transforming program to backend"""
         self.add_pass(CastInt64ToInt32Pass(exported_program))
+        self.add_pass(RemoveGetItemPass())
         self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
         self.add_pass(SizeAdjustConv2DPass())
         self.add_pass(RemoveClonePass())
@@ -72,6 +75,7 @@ def transform_to_backend_pipeline(
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(Conv1dUnsqueezePass(exported_program))
         self.add_pass(DecomposeSoftmaxesPass())
+        self.add_pass(DecomposeLinearPass())
         for spec in compile_spec:
             if spec.key == "permute_memory_format":
                 memory_format = spec.value.decode()