Update on "move rope related logic together"

Right now, rope related code scatters around a few different places in `llama_transformer`. It makes it hard to make changes to rope related things.

This PR moves all rope related logic into its own module.

Differential Revision: [D65173598](https://our.internmc.facebook.com/intern/diff/D65173598/)

[ghstack-poisoned]

Author: helunwencser

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-c8a648d4dffb9f0133ff4a2ea0e660b42105d3ad
+19eff28ff3f19b50da46f5a9ff5f4d4d213806fe

.ci/scripts/gather_test_models.py

@@ -24,6 +24,8 @@
         "ic4": "linux.12xlarge",
         "resnet50": "linux.12xlarge",
         "llava": "linux.12xlarge",
+        "llama3_2_vision_encoder": "linux.12xlarge",
+        "llama3_2_text_decoder": "linux.12xlarge",
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",

.ci/scripts/test_model.sh

@@ -77,7 +77,7 @@ test_model() {
     # Install requirements for export_llama
     bash examples/models/llama/install_requirements.sh
     # Test export_llama script: python3 -m examples.models.llama.export_llama
-    "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama -c examples/models/llama/params/demo_rand_params.pth -p examples/models/llama/params/demo_config.json
+    "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/llama/params/demo_config.json
     run_portable_executor_runner
     rm "./${MODEL_NAME}.pte"
   fi
@@ -87,6 +87,10 @@ test_model() {
     bash examples/models/llava/install_requirements.sh
     STRICT="--no-strict"
   fi
+  if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" || "$MODEL_NAME" == "llama3_2_text_decoder" ]]; then
+    # Install requirements for llama vision.
+    bash examples/models/llama3_2_vision/install_requirements.sh
+  fi
   # python3 -m examples.portable.scripts.export --model_name="llama2" should works too
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}" "${STRICT}"
   run_portable_executor_runner

.github/workflows/apple.yml

@@ -20,12 +20,26 @@ on:
       - extension/benchmark/apple/**
       - extension/module/**
   workflow_dispatch:
+  schedule:
+    - cron: '0 10 * * *'  # Runs daily at 2 AM PST
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
   cancel-in-progress: true
 
 jobs:
+  set-version:
+    runs-on: ubuntu-22.04
+    outputs:
+      version: ${{ steps.set_version.outputs.version }}
+    steps:
+      - name: Set VERSION variable
+        id: set_version
+        shell: bash
+        run: |
+          VERSION="0.4.0.$(TZ='PST8PDT' date +%Y%m%d)"
+          echo "version=$VERSION" >> "$GITHUB_OUTPUT"
+
   build-demo-ios:
     name: build-demo-ios
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
@@ -39,6 +53,8 @@ jobs:
       secrets-env: BUILD_CERTIFICATE_BASE64 EXECUTORCH_DEMO_BUILD_PROVISION_PROFILE_BASE64 KEYCHAIN_PASSWORD
       upload-artifact: ios-apps
       script: |
+        set -eux
+
         BUILD_TOOL=cmake
 
         .ci/scripts/setup-conda.sh
@@ -57,7 +73,7 @@ jobs:
 
         # Build and test iOS Demo App
         PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        build/test_ios_ci.sh ${ARTIFACTS_DIR_NAME}
+        build/test_ios_ci.sh "${ARTIFACTS_DIR_NAME}"
 
   # Upload the test demo app to S3
   upload-demo-ios:
@@ -75,6 +91,7 @@ jobs:
         shell: bash
         working-directory: ${{ runner.temp }}/artifacts/
         run: |
+          set -eux
           ls -lah ./
 
       - name: Upload the artifacts to S3
@@ -112,6 +129,7 @@ jobs:
 
   build-frameworks-ios:
     name: build-frameworks-ios
+    needs: set-version
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:
       runner: macos-latest-xlarge
@@ -121,8 +139,10 @@ jobs:
       upload-artifact: executorch-frameworks-ios
       timeout: 90
       script: |
+        set -eux
+
         BUILD_TOOL=cmake
-        VERSION="latest"
+        VERSION="${{ needs.set-version.outputs.version }}"
         FRAMEWORKS=(
           "executorch"
           "backend_coreml"
@@ -171,13 +191,17 @@ jobs:
 
   upload-frameworks-ios:
     runs-on: ubuntu-22.04
-    needs: build-frameworks-ios
+    needs: [build-frameworks-ios, set-version]
     timeout-minutes: 30
+    environment: ${{ github.ref == 'refs/heads/main' && 'cherry-pick-bot' || '' }}
     permissions:
       id-token: write
-      contents: read
+      contents: write
     steps:
       - uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+          token: ${{ secrets.GH_PYTORCHBOT_CHERRY_PICK_TOKEN || secrets.GITHUB_TOKEN }}
       - uses: actions/setup-python@v4
         with:
           python-version: '3.11'
@@ -194,15 +218,15 @@ jobs:
           name: executorch-frameworks-ios
           path: ${{ runner.temp }}/frameworks-ios/
       - name: Only push to S3 when running the workflow manually from main branch
-        if: ${{ github.event_name == 'workflow_dispatch' && github.ref == 'refs/heads/main' }}
+        if: ${{ (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && github.ref == 'refs/heads/main' }}
         shell: bash
         run: |
-          set -eux
           echo "UPLOAD_ON_MAIN=1" >> "${GITHUB_ENV}"
       - name: Upload the artifact to ossci-ios S3 bucket
         shell: bash
         run: |
           set -eux
+          VERSION="${{ needs.set-version.outputs.version }}"
 
           pip install awscli==1.32.18
 
@@ -213,9 +237,44 @@ jobs:
 
           for FILENAME in "${RUNNER_TEMP}"/frameworks-ios/*.zip; do
             [ -e "${FILENAME}" ] || continue
-            shasum -a 256 "${FILENAME}"
+            FRAMEWORK_NAME=$(basename "${FILENAME}" | sed "s/-${VERSION}.zip//")
+            CHECKSUM=$(shasum -a 256 "${FILENAME}" | cut -d ' ' -f1)
+            echo "${FRAMEWORK_NAME} ${CHECKSUM}" >> "${RUNNER_TEMP}/checksums.txt"
             ${AWS_CMD} "${FILENAME}" s3://ossci-ios/executorch/ --acl public-read
           done
+      - name: Update SwiftPM
+        shell: bash
+        run: |
+          set -eux
+          VERSION="${{ needs.set-version.outputs.version }}"
+          BRANCH="swiftpm-${VERSION}"
+
+          git checkout swiftpm
+
+          if git show-ref --verify --quiet refs/heads/${BRANCH}; then
+            git checkout "${BRANCH}"
+          else
+            git checkout -b "${BRANCH}"
+          fi
+
+          [[ -f Package.swift ]] || mv Package.swift.template Package.swift
+
+          sed -i "s/__VERSION__/${VERSION}/g" Package.swift
+
+          while read -r FRAMEWORK CHECKSUM; do
+            sed -i "s/__SHA256_${FRAMEWORK}__/${CHECKSUM}/g" Package.swift
+          done < "${RUNNER_TEMP}/checksums.txt"
+
+          if [[ "${UPLOAD_ON_MAIN:-0}" == "1" ]]; then
+            git config --global user.name "PyTorch Bot"
+            git config --global user.email "[email protected]"
+            git add Package.swift
+            git commit -am "${VERSION}"
+            git push -f origin "${BRANCH}"
+          else
+            echo "Draft Package.swift:"
+            cat Package.swift
+          fi
 
   build-benchmark-app:
     name: build-benchmark-app
@@ -281,5 +340,5 @@ jobs:
         echo "::group::Build ExecuTorch benchmark app"
         mkdir -p extension/benchmark/apple/Benchmark/Models
         ${CONDA_RUN} --no-capture-output \
-          build/build_apple_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+          build/build_apple_llm_demo.sh "${ARTIFACTS_DIR_NAME}"
         echo "::endgroup::"

.github/workflows/trunk.yml

@@ -58,7 +58,7 @@ jobs:
         bash .ci/scripts/setup-conda.sh
         # Setup MacOS dependencies as there is no Docker support on MacOS atm
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh "${BUILD_TOOL}"
-        # Build and test xecutorch
+        # Build and test executorch
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "${BACKEND}" "${DEMO_BACKEND_DELEGATION}"
 
   test-custom-ops-macos:
@@ -411,8 +411,6 @@ jobs:
         pip install -U "huggingface_hub[cli]"
         huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
         pip install accelerate sentencepiece
-        # TODO(guangyang): Switch to use released transformers library after all required patches are included
-        pip install "git+https://github.com/huggingface/transformers.git@6cc4dfe3f1e8d421c6d6351388e06e9b123cbfe1"
         pip list
         echo "::endgroup::"
 

.lintrunner.toml

@@ -196,3 +196,91 @@ command = [
     '@{{PATHSFILE}}',
 ]
 is_formatter = true
+
+[[linter]]
+code = 'NOSTDINC'
+include_patterns = [
+    "**/*.c",
+    "**/*.cpp",
+    "**/*.h",
+    "**/*.hpp",
+]
+exclude_patterns = [
+    '**/devtools/**',
+    '**/test/**',
+    '**/testing_util/**',
+    '**/third-party/**',
+    'backends/**',
+    'devtools/**',
+    'examples/**',
+    'extension/**',
+    'kernels/optimized/**',
+    'scripts/**',
+    'third-party/**',
+    'util/**',
+]
+command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'grep_linter',
+    '--pattern=([^\\S\r\n]*#include\s*<(deque|exception|forward_list|functional|list|map|multimap|multiset|priority_queue|queue|set|stack|string|unordered_map|unordered_multimap|unordered_multiset|unordered_set|vector)>)',
+    '--linter-name=NOSTDINC',
+    '--error-name=Standard C++ container include in core',
+    """--error-description=\
+        Standard library containers should not be included in ExecuTorch core \
+        because they may call malloc, which is not allowed in core. \
+    """,
+    '--',
+    '@{{PATHSFILE}}',
+]
+
+[[linter]]
+code = 'NOTORCHINC'
+include_patterns = [
+    "**/*.c",
+    "**/*.cpp",
+    "**/*.h",
+    "**/*.hpp",
+]
+exclude_patterns = [
+    '**/devtools/**',
+    '**/fb/**',
+    '**/test/**',
+    '**/tests/**',
+    '**/testing_util/**',
+    '**/third-party/**',
+    'backends/**',
+    'codegen/templates/RegisterDispatchKeyCustomOps.cpp',
+    'codegen/templates/RegisterSchema.cpp',
+    'devtools/**',
+    'examples/**',
+    'exir/verification/bindings.cpp',
+    'extension/**',
+    'kernels/optimized/**',
+    'runtime/core/exec_aten/**',
+    'runtime/executor/tensor_parser_aten.cpp',
+    'scripts/**',
+    'test/**',
+    'third-party/**',
+    'util/**',
+]
+command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'grep_linter',
+    '--pattern=#include\s+[<"](aten/|ATen/|torch/)',
+    '--linter-name=NOTORCHINC',
+    '--error-name=ATen or torch include',
+    """--error-description=\
+        PyTorch includes in ExecuTorch core are prohibited to prevent \
+        accidentally breaking core's requirements; please make sure this \
+        header complies (e.g., no streams/malloc/syscalls) and then include \
+        a patch to update this linter.\
+    """,
+    '--',
+    '@{{PATHSFILE}}',
+]

CMakeLists.txt

@@ -324,6 +324,19 @@ message(STATUS "Using python executable '${PYTHON_EXECUTABLE}'")
 set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 
 # Let files say "include <executorch/path/to/header.h>".
+# TODO(#6475): This requires/assumes that the repo lives in a directory named
+# exactly `executorch`. Check the assumption first. Remove this check once we
+# stop relying on the assumption.
+cmake_path(GET CMAKE_CURRENT_SOURCE_DIR FILENAME _repo_dir_name)
+if(NOT "${_repo_dir_name}" STREQUAL "executorch")
+  message(
+    FATAL_ERROR
+      "The ExecuTorch repo must be cloned into a directory named exactly "
+      "`executorch`; found `${_repo_dir_name}`. See "
+      "https://github.com/pytorch/executorch/issues/6475 for progress on a "
+      "fix for this restriction."
+  )
+endif()
 set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/..)
 
 #

backends/arm/TARGETS

@@ -26,12 +26,27 @@ python_library(
         "fbsource//third-party/serialization_lib/python/serializer:serializer",
         "fbsource//third-party/serialization_lib/python/tosa:tosa",
         ":arm_vela",
+        ":process_node",
         "//executorch/backends/arm/operators:lib",
         "//executorch/backends/arm/operators:node_visitor",
         "//executorch/backends/arm/_passes:passes",
     ],
 )
 
+python_library(
+    name = "process_node",
+    srcs = ["process_node.py"],
+    typing = True,
+    deps = [
+        "fbsource//third-party/serialization_lib/python/tosa:tosa",
+        "//executorch/backends/arm/operators:node_visitor",
+        "//executorch/backends/arm:tosa_mapping",
+        "//executorch/backends/arm:tosa_quant_utils",
+        "//executorch/backends/arm:tosa_utils",
+        "//executorch/exir:lib",
+    ],
+)
+
 python_library(
     name = "arm_vela",
     srcs = [

backends/arm/_passes/arm_pass_utils.py

@@ -5,6 +5,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 from typing import Optional
 
 import torch

backends/arm/_passes/cast_int64_pass.py

@@ -3,6 +3,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import torch
 from executorch.exir.pass_base import ExportPass, PassResult