Update executorch to latest

.ci/docker/build.sh

@@ -37,6 +37,10 @@ case "${IMAGE_NAME}" in
     ARM_SDK=yes
     CLANG_VERSION=12
     ;;
+  executorch-ubuntu-22.04-qnn-sdk)
+    QNN_SDK=yes
+    CLANG_VERSION=12
+    ;;
   executorch-ubuntu-22.04-clang12-android)
     LINTRUNNER=""
     CLANG_VERSION=12
@@ -72,6 +76,7 @@ docker build \
   --build-arg "LINTRUNNER=${LINTRUNNER:-}" \
   --build-arg "BUILD_DOCS=${BUILD_DOCS}" \
   --build-arg "ARM_SDK=${ARM_SDK:-}" \
+  --build-arg "QNN_SDK=${QNN_SDK:-}" \
   --build-arg "ANDROID_NDK_VERSION=${ANDROID_NDK_VERSION:-}" \
   -f "${OS}"/Dockerfile \
   "$@" \

.ci/docker/ci_commit_pins/pytorch.txt

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

.ci/docker/common/install_cache.sh

@@ -12,6 +12,26 @@ set -ex
 # shellcheck source=/dev/null
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
+install_ubuntu() {
+  echo "Preparing to build sccache from source"
+  apt-get update
+  # libssl-dev will not work as it is upgraded to libssl3 in Ubuntu-22.04.
+  # Instead use lib and headers from OpenSSL1.1 installed in `install_openssl.sh``
+  apt-get install -y cargo
+  echo "Checking out sccache repo"
+  git clone https://github.com/mozilla/sccache -b v0.8.2
+
+  cd sccache
+  echo "Building sccache"
+  cargo build --release
+  cp target/release/sccache /opt/cache/bin
+  echo "Cleaning up"
+  cd ..
+  rm -rf sccache
+  apt-get remove -y cargo rustc
+  apt-get autoclean && apt-get clean
+}
+
 install_binary() {
   echo "Downloading sccache binary from S3 repo"
   curl --retry 3 https://s3.amazonaws.com/ossci-linux/sccache -o /opt/cache/bin/sccache
@@ -22,15 +42,33 @@ mkdir -p /opt/cache/bin
 sed -e 's|PATH="\(.*\)"|PATH="/opt/cache/bin:\1"|g' -i /etc/environment
 export PATH="/opt/cache/bin:$PATH"
 
-# NB: Install the pre-built binary from S3 as building from source
-# https://github.com/pytorch/sccache has started failing mysteriously
-# in which sccache server couldn't start with the following error:
-#   sccache: error: Invalid argument (os error 22)
-install_binary
+install_ubuntu
 
 function write_sccache_stub() {
   BINARY=$1
-  printf "#!/bin/sh\nif [ \$(env -u LD_PRELOAD ps -p \$PPID -o comm=) != sccache ]; then\n  exec sccache %s \"\$@\"\nelse\n  exec %s \"\$@\"\nfi" "$(which "${BINARY}")" "$(which "${BINARY}")" > "/opt/cache/bin/${BINARY}"
+  if [ $1 == "gcc" ]; then
+    # Do not call sccache recursively when dumping preprocessor argument
+    # For some reason it's very important for the first cached nvcc invocation
+    cat >"/opt/cache/bin/$1" <<EOF
+#!/bin/sh
+if [ "\$1" = "-E" ] || [ "\$2" = "-E" ]; then
+  exec $(which $1) "\$@"
+elif [ \$(env -u LD_PRELOAD ps -p \$PPID -o comm=) != sccache ]; then
+  exec sccache $(which $1) "\$@"
+else
+  exec $(which $1) "\$@"
+fi
+EOF
+  else
+    cat >"/opt/cache/bin/$1" <<EOF
+#!/bin/sh
+if [ \$(env -u LD_PRELOAD ps -p \$PPID -o comm=) != sccache ]; then
+  exec sccache $(which $1) "\$@"
+else
+  exec $(which $1) "\$@"
+fi
+EOF
+  fi
   chmod a+x "/opt/cache/bin/${BINARY}"
 }
 
@@ -44,7 +82,7 @@ init_sccache() {
 
   # NB: This function is adopted from PyTorch core at
   # https://github.com/pytorch/pytorch/blob/main/.ci/pytorch/common-build.sh
-  as_ci_user sccache --stop-server > /dev/null 2>&1 || true
+  as_ci_user sccache --stop-server >/dev/null 2>&1 || true
   rm -f "${SCCACHE_ERROR_LOG}" || true
 
   # Clear sccache stats before using it

.ci/docker/ubuntu/Dockerfile

@@ -57,6 +57,7 @@ COPY ./common/utils.sh utils.sh
 RUN bash ./install_cache.sh && rm install_cache.sh utils.sh
 ENV SCCACHE_BUCKET ossci-compiler-cache-circleci-v2
 ENV SCCACHE_S3_KEY_PREFIX executorch
+ENV SCCACHE_REGION us-east-1
 
 ARG TORCH_VERSION
 COPY ./common/install_pytorch.sh install_pytorch.sh
@@ -82,5 +83,7 @@ COPY --chown=ci-user:ci-user ./arm /opt/arm
 # Set up ARM SDK if needed
 RUN if [ -n "${ARM_SDK}" ]; then git config --global user.email "ossci@example.com"; git config --global user.name "OSS CI"; bash /opt/arm/setup.sh --i-agree-to-the-contained-eula /opt/arm-sdk; chown -R ci-user:ci-user /opt/arm-sdk; fi
 
+ARG QNN_SDK
+
 USER ci-user
 CMD ["bash"]

.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",  # TODO: re-enable test when Huy's change is in / model gets smaller.
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",
@@ -88,8 +90,8 @@ def model_should_run_on_event(model: str, event: str) -> bool:
     if event == "pull_request":
         return model in ["mv3", "vit"]
     elif event == "push":
-        # 'emformer_predict' is running super slow. Only run it periodically
-        return model not in ["emformer_predict"]
+        # These are super slow. Only run it periodically
+        return model not in ["dl3", "edsr", "emformer_predict"]
     else:
         return True
 

.ci/scripts/setup-linux.sh

@@ -23,3 +23,4 @@ fi
 # of nightly. This allows CI to test against latest commits from PyTorch
 install_executorch "use-pt-pinned-commit"
 build_executorch_runner "${BUILD_TOOL}"
+do_not_use_nightly_on_ci

.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/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

.ci/scripts/utils.sh

@@ -113,10 +113,26 @@ cmake_install_executorch_lib() {
 }
 
 download_stories_model_artifacts() {
-    # Download stories110M.pt and tokenizer from Github
+  # Download stories110M.pt and tokenizer from Github
   curl -Ls "https://huggingface.co/karpathy/tinyllamas/resolve/main/stories110M.pt" --output stories110M.pt
   curl -Ls "https://raw.githubusercontent.com/karpathy/llama2.c/master/tokenizer.model" --output tokenizer.model
   # Create params.json file
   touch params.json
   echo '{"dim": 768, "multiple_of": 32, "n_heads": 12, "n_layers": 12, "norm_eps": 1e-05, "vocab_size": 32000}' > params.json
 }
+
+do_not_use_nightly_on_ci() {
+  # An assert to make sure that we are not using PyTorch nightly on CI to prevent
+  # regression as documented in https://github.com/pytorch/executorch/pull/6564
+  TORCH_VERSION=$(pip list | grep -w 'torch ' | awk -F ' ' {'print $2'} | tr -d '\n')
+
+  # The version of PyTorch building from source looks like 2.6.0a0+gitc8a648d that
+  # includes the commit while nightly (2.6.0.dev20241019+cpu) or release (2.6.0)
+  # won't have that. Note that we couldn't check for the exact commit from the pin
+  # ci_commit_pins/pytorch.txt here because the value will be different when running
+  # this on PyTorch CI
+  if [[ "${TORCH_VERSION}" != *"+git"* ]]; then
+    echo "Unexpected torch version. Expected binary built from source, got ${TORCH_VERSION}"
+    exit 1
+  fi
+}

.github/scripts/check_labels.py

@@ -0,0 +1,62 @@
+#!/usr/bin/env python3
+"""Check whether a PR has required labels."""
+
+import sys
+from typing import Any
+
+from github_utils import gh_delete_comment, gh_post_pr_comment
+from gitutils import get_git_remote_name, get_git_repo_dir, GitRepo
+from label_utils import has_required_labels, is_label_err_comment, LABEL_ERR_MSG
+from trymerge import GitHubPR
+
+
+def delete_all_label_err_comments(pr: "GitHubPR") -> None:
+    for comment in pr.get_comments():
+        if is_label_err_comment(comment):
+            gh_delete_comment(pr.org, pr.project, comment.database_id)
+
+
+def add_label_err_comment(pr: "GitHubPR") -> None:
+    # Only make a comment if one doesn't exist already
+    if not any(is_label_err_comment(comment) for comment in pr.get_comments()):
+        gh_post_pr_comment(pr.org, pr.project, pr.pr_num, LABEL_ERR_MSG)
+
+
+def parse_args() -> Any:
+    from argparse import ArgumentParser
+
+    parser = ArgumentParser("Check PR labels")
+    parser.add_argument("pr_num", type=int)
+    # add a flag to return a non-zero exit code if the PR does not have the required labels
+    parser.add_argument(
+        "--exit-non-zero",
+        action="store_true",
+        help="Return a non-zero exit code if the PR does not have the required labels",
+    )
+
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    repo = GitRepo(get_git_repo_dir(), get_git_remote_name())
+    org, project = repo.gh_owner_and_name()
+    pr = GitHubPR(org, project, args.pr_num)
+
+    try:
+        if not has_required_labels(pr):
+            print(LABEL_ERR_MSG, flush=True)
+            add_label_err_comment(pr)
+            if args.exit_non_zero:
+                raise RuntimeError("PR does not have required labels")
+        else:
+            delete_all_label_err_comments(pr)
+    except Exception as e:
+        if args.exit_non_zero:
+            raise RuntimeError(f"Error checking labels: {e}") from e
+
+    sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()

.github/scripts/github_utils.py

@@ -72,10 +72,10 @@ def gh_fetch_url(
     headers: Optional[Dict[str, str]] = None,
     data: Union[Optional[Dict[str, Any]], str] = None,
     method: Optional[str] = None,
-    reader: Callable[[Any], Any] = lambda x: x.read(),
+    reader: Callable[[Any], Any] = json.load,
 ) -> Any:
     return gh_fetch_url_and_headers(
-        url, headers=headers, data=data, reader=json.load, method=method
+        url, headers=headers, data=data, reader=reader, method=method
     )[1]
 
 
@@ -169,7 +169,7 @@ def gh_post_commit_comment(
 
 def gh_delete_comment(org: str, repo: str, comment_id: int) -> None:
     url = f"{GITHUB_API_URL}/repos/{org}/{repo}/issues/comments/{comment_id}"
-    gh_fetch_url(url, method="DELETE")
+    gh_fetch_url(url, method="DELETE", reader=lambda x: x.read())
 
 
 def gh_fetch_merge_base(org: str, repo: str, base: str, head: str) -> str:

.github/workflows/_android.yml

@@ -73,6 +73,8 @@ jobs:
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug-androidTest.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/fp32-xnnpack-custom/model.zip
+          curl -o android-test-debug.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug.apk
+          curl -o android-test-debug-androidTest.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug-androidTest.apk
           unzip model.zip
           mv *.pte model.pte
 

.github/workflows/android-perf.yml

@@ -136,7 +136,7 @@ jobs:
       fail-fast: false
     with:
       runner: linux.4xlarge
-      docker-image: executorch-ubuntu-22.04-clang12-android
+      docker-image: executorch-ubuntu-22.04-qnn-sdk
       submodules: 'true'
       timeout: 60
       upload-artifact: android-models