Update on "Fix Cuda out of memory issue for eager runner"

This PR updates the eager runner to disable grad and save memory usage.

It also update the prompt format to not include bos.

Differential Revision: [D65962743](https://our.internmc.facebook.com/intern/diff/D65962743/)

[ghstack-poisoned]

Author: helunwencser

.ci/scripts/gather_test_models.py

@@ -25,6 +25,7 @@
         "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

@@ -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/pull.yml

@@ -72,10 +72,6 @@ jobs:
         conda activate "${CONDA_ENV}"
 
         MODEL_NAME=${{ matrix.model }}
-        # Install requirements for llama vision
-        if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" ]]; then
-          bash examples/models/llama3_2_vision/install_requirements.sh
-        fi
         BUILD_TOOL=${{ matrix.build-tool }}
         BACKEND=${{ matrix.backend }}
         DEMO_BACKEND_DELEGATION=${{ matrix.demo_backend_delegation }}

.github/workflows/trunk.yml

@@ -58,10 +58,6 @@ 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}"
-        # Install requirements for llama vision
-        if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" ]]; then
-          ${CONDA_RUN} bash examples/models/llama3_2_vision/install_requirements.sh
-        fi
         # Build and test executorch
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "${BACKEND}" "${DEMO_BACKEND_DELEGATION}"
 
@@ -415,8 +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::"
 

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
 

backends/arm/_passes/conv1d_unsqueeze_pass.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
+
 
 import torch
 from executorch.backends.arm._passes.arm_pass_utils import (

backends/arm/_passes/decompose_div_pass.py

@@ -4,6 +4,8 @@
 # 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.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass

backends/arm/_passes/decompose_layernorm_pass.py

@@ -4,6 +4,8 @@
 # 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 operator
 
 import torch