Update on "[executorch][serialization] Refactor flatbuffer utils into separate file"

For usage in extension/flat_tensor/serialize.

Differential Revision: [D66854756](https://our.internmc.facebook.com/intern/diff/D66854756/)

[ghstack-poisoned]

Author: lucylq

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-19eff28ff3f19b50da46f5a9ff5f4d4d213806fe
+2ea4b56ec872424e486c4fe2d55da061067a2ed3

.ci/docker/common/install_pytorch.sh

@@ -26,7 +26,7 @@ install_pytorch_and_domains() {
 
   chown -R ci-user .
 
-  export _GLIBCXX_USE_CXX11_ABI=0
+  export _GLIBCXX_USE_CXX11_ABI=1
   # Then build and install PyTorch
   conda_run python setup.py bdist_wheel
   pip_install "$(echo dist/*.whl)"

.ci/docker/ubuntu/Dockerfile

@@ -79,9 +79,6 @@ RUN if [ -n "${ANDROID_NDK_VERSION}" ]; then bash ./install_android.sh; fi
 RUN rm install_android.sh
 
 ARG ARM_SDK
-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 "[email protected]"; 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
 

.ci/scripts/build_llama_android.sh

@@ -12,7 +12,8 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
 install_executorch_and_backend_lib() {
   echo "Installing executorch and xnnpack backend"
-  rm -rf cmake-android-out && mkdir cmake-android-out
+  clean_executorch_install_folders
+  mkdir cmake-android-out
   ANDROID_NDK=/opt/ndk
   BUCK2=buck2
   ANDROID_ABI=arm64-v8a

.ci/scripts/download_hf_hub.sh

@@ -0,0 +1,95 @@
+#!/bin/bash
+
+# Function to download files from the Hugging Face Hub
+# Arguments:
+# 1. model_id: The Hugging Face repository ID (e.g., "organization/model_name")
+# 2. subdir: The optional subdirectory in the repo to look for files (pass "" if not used)
+# 3. file_names: A space-separated list of filenames to be downloaded
+# Returns:
+# The directory containing the downloaded files
+function download_hf_files() {
+  local model_id="$1"
+  local subdir="$2"
+  shift 2
+  local file_names=("$@")  # Capture all remaining arguments as an array
+
+  local download_dir
+
+  # Use the first file to determine the download directory
+  download_dir=$(python3 -c "
+from huggingface_hub import hf_hub_download
+# Download the first file and get its directory
+path = hf_hub_download(
+    repo_id='${model_id}',
+    filename='${subdir:+${subdir}/}${file_names[0]}'
+)
+import os
+print(os.path.dirname(path))")
+
+  if [ $? -ne 0 ]; then
+    echo "Error: Failed to determine download directory from ${file_names[0]}" >&2
+    return 1
+  fi
+
+  # Download remaining files into the same directory
+  for file_name in "${file_names[@]:1}"; do
+    python3 -c "
+from huggingface_hub import hf_hub_download
+# Download the file
+hf_hub_download(
+    repo_id='${model_id}',
+    filename='${subdir:+${subdir}/}${file_name}'
+)"
+
+    if [ $? -ne 0 ]; then
+      echo "Error: Failed to download ${file_name} from ${model_id}" >&2
+      return 1
+    fi
+  done
+
+  # Return the directory containing the downloaded files
+  echo "$download_dir"
+}
+
+# Check if script is called directly
+if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
+  # Parse arguments from CLI
+  while [[ $# -gt 0 ]]; do
+    case $1 in
+      --model_id)
+        MODEL_ID="$2"
+        shift 2
+        ;;
+      --subdir)
+        SUBDIR="$2"
+        shift 2
+        ;;
+      --files)
+        shift
+        FILES_TO_DOWNLOAD=()
+        while [[ $# -gt 0 && $1 != --* ]]; do
+          FILES_TO_DOWNLOAD+=("$1")
+          shift
+        done
+        ;;
+      *)
+        echo "Unknown option: $1" >&2
+        exit 1
+        ;;
+    esac
+  done
+
+  # Validate required arguments
+  if [ -z "$MODEL_ID" ] || [ ${#FILES_TO_DOWNLOAD[@]} -eq 0 ]; then
+    echo "Usage: $0 --model_id <model_id> --subdir <subdir> --files <file1> [<file2> ...]" >&2
+    exit 1
+  fi
+
+  # Call the function
+  DOWNLOAD_DIR=$(download_hf_files "$MODEL_ID" "$SUBDIR" "${FILES_TO_DOWNLOAD[@]}")
+  if [ $? -eq 0 ]; then
+    echo "$DOWNLOAD_DIR"
+  else
+    exit 1
+  fi
+fi

.ci/scripts/gather_benchmark_configs.py

@@ -0,0 +1,224 @@
+#!/usr/bin/env python
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+import logging
+import os
+import re
+from typing import Any, Dict
+
+from examples.models import MODEL_NAME_TO_MODEL
+
+
+# Device pools for AWS Device Farm
+DEVICE_POOLS = {
+    "apple_iphone_15": "arn:aws:devicefarm:us-west-2:308535385114:devicepool:02a2cf0f-6d9b-45ee-ba1a-a086587469e6/3b5acd2e-92e2-4778-b651-7726bafe129d",
+    "apple_iphone_15+ios_18": "arn:aws:devicefarm:us-west-2:308535385114:devicepool:02a2cf0f-6d9b-45ee-ba1a-a086587469e6/12c8b15c-8d03-4e07-950d-0a627e7595b4",
+    "samsung_galaxy_s22": "arn:aws:devicefarm:us-west-2:308535385114:devicepool:02a2cf0f-6d9b-45ee-ba1a-a086587469e6/e59f866a-30aa-4aa1-87b7-4510e5820dfa",
+    "samsung_galaxy_s24": "arn:aws:devicefarm:us-west-2:308535385114:devicepool:02a2cf0f-6d9b-45ee-ba1a-a086587469e6/98f8788c-2e25-4a3c-8bb2-0d1e8897c0db",
+    "google_pixel_8_pro": "arn:aws:devicefarm:us-west-2:308535385114:devicepool:02a2cf0f-6d9b-45ee-ba1a-a086587469e6/d65096ab-900b-4521-be8b-a3619b69236a",
+}
+
+# Predefined benchmark configurations
+BENCHMARK_CONFIGS = {
+    "xplat": [
+        "xnnpack_q8",
+        "hf_xnnpack_fp32",
+        "llama3_fb16",
+        "llama3_spinquant",
+        "llama3_qlora",
+    ],
+    "android": [
+        "qnn_q8",
+        # TODO: Add support for llama3 htp
+        # "llama3_qnn_htp",
+    ],
+    "ios": [
+        "coreml_fp16",
+        "mps",
+        "llama3_coreml_ane",
+    ],
+}
+
+
+def parse_args() -> Any:
+    """
+    Parse command-line arguments.
+
+    Returns:
+        argparse.Namespace: Parsed command-line arguments.
+
+    Example:
+        parse_args() -> Namespace(models=['mv3', 'meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8'],
+                                   os='android',
+                                   devices=['samsung_galaxy_s22'])
+    """
+    from argparse import ArgumentParser
+
+    def comma_separated(value: str):
+        """
+        Parse a comma-separated string into a list.
+        """
+        return value.split(",")
+
+    parser = ArgumentParser("Gather all benchmark configs.")
+    parser.add_argument(
+        "--os",
+        type=str,
+        choices=["android", "ios"],
+        help="The target OS.",
+    )
+    parser.add_argument(
+        "--models",
+        type=comma_separated,  # Use the custom parser for comma-separated values
+        help=f"Comma-separated model IDs or names. Valid values include {MODEL_NAME_TO_MODEL}.",
+    )
+    parser.add_argument(
+        "--devices",
+        type=comma_separated,  # Use the custom parser for comma-separated values
+        help=f"Comma-separated device names. Available devices: {list(DEVICE_POOLS.keys())}",
+    )
+
+    return parser.parse_args()
+
+
+def set_output(name: str, val: Any) -> None:
+    """
+    Set the output value to be used by other GitHub jobs.
+
+    Args:
+        name (str): The name of the output variable.
+        val (Any): The value to set for the output variable.
+
+    Example:
+        set_output("benchmark_configs", {"include": [...]})
+    """
+
+    if os.getenv("GITHUB_OUTPUT"):
+        print(f"Setting {val} to GitHub output")
+        with open(str(os.getenv("GITHUB_OUTPUT")), "a") as env:
+            print(f"{name}={val}", file=env)
+    else:
+        print(f"::set-output name={name}::{val}")
+
+
+def is_valid_huggingface_model_id(model_name: str) -> bool:
+    """
+    Validate if the model name matches the pattern for HuggingFace model IDs.
+
+    Args:
+        model_name (str): The model name to validate.
+
+    Returns:
+        bool: True if the model name matches the valid pattern, False otherwise.
+
+    Example:
+        is_valid_huggingface_model_id('meta-llama/Llama-3.2-1B') -> True
+    """
+    pattern = r"^[a-zA-Z0-9-_]+/[a-zA-Z0-9-_.]+$"
+    return bool(re.match(pattern, model_name))
+
+
+def get_benchmark_configs() -> Dict[str, Dict]:
+    """
+    Gather benchmark configurations for a given set of models on the target operating system and devices.
+
+    Args:
+        None
+
+    Returns:
+        Dict[str, Dict]: A dictionary containing the benchmark configurations.
+
+    Example:
+        get_benchmark_configs() -> {
+            "include": [
+                {
+                    "model": "meta-llama/Llama-3.2-1B",
+                    "config": "llama3_qlora",
+                    "device_name": "apple_iphone_15",
+                    "device_arn": "arn:aws:..."
+                },
+                {
+                    "model": "mv3",
+                    "config": "xnnpack_q8",
+                    "device_name": "samsung_galaxy_s22",
+                    "device_arn": "arn:aws:..."
+                },
+                ...
+            ]
+        }
+    """
+    args = parse_args()
+    target_os = args.os
+    devices = args.devices
+    models = args.models
+
+    benchmark_configs = {"include": []}
+
+    for model_name in models:
+        configs = []
+        if is_valid_huggingface_model_id(model_name):
+            if model_name.startswith("meta-llama/"):
+                # LLaMA models
+                repo_name = model_name.split("meta-llama/")[1]
+                if "qlora" in repo_name.lower():
+                    configs.append("llama3_qlora")
+                elif "spinquant" in repo_name.lower():
+                    configs.append("llama3_spinquant")
+                else:
+                    configs.append("llama3_fb16")
+                    configs.extend(
+                        [
+                            config
+                            for config in BENCHMARK_CONFIGS.get(target_os, [])
+                            if config.startswith("llama")
+                        ]
+                    )
+            else:
+                # Non-LLaMA models
+                configs.append("hf_xnnpack_fp32")
+        elif model_name in MODEL_NAME_TO_MODEL:
+            # ExecuTorch in-tree non-GenAI models
+            configs.append("xnnpack_q8")
+            configs.extend(
+                [
+                    config
+                    for config in BENCHMARK_CONFIGS.get(target_os, [])
+                    if not config.startswith("llama")
+                ]
+            )
+        else:
+            # Skip unknown models with a warning
+            logging.warning(f"Unknown or invalid model name '{model_name}'. Skipping.")
+            continue
+
+        # Add configurations for each valid device
+        for device in devices:
+            for config in configs:
+                if config == "llama3_coreml_ane" and not device.endswith("+ios_18"):
+                    device = f"{device}+ios_18"
+                    logging.info(
+                        f"Benchmark config '{config}' only works on iOS 18+, auto-upgraded device pool to '{device}'"
+                    )
+
+                if device not in DEVICE_POOLS:
+                    logging.warning(f"Unsupported device '{device}'. Skipping.")
+                    continue
+
+                record = {
+                    "model": model_name,
+                    "config": config,
+                    "device_name": device,
+                    "device_arn": DEVICE_POOLS[device],
+                }
+                benchmark_configs["include"].append(record)
+
+    set_output("benchmark_configs", json.dumps(benchmark_configs))
+
+
+if __name__ == "__main__":
+    get_benchmark_configs()

.ci/scripts/setup-vulkan-linux-deps.sh

@@ -27,7 +27,7 @@ install_swiftshader() {
 
 install_vulkan_sdk() {
   VULKAN_SDK_VERSION=$1
-  _vulkan_sdk_url="https://sdk.lunarg.com/sdk/download/${VULKAN_SDK_VERSION}/linux/vulkansdk-linux-x86_64-${VULKAN_SDK_VERSION}.tar.gz"
+  _vulkan_sdk_url="https://sdk.lunarg.com/sdk/download/${VULKAN_SDK_VERSION}/linux/vulkansdk-linux-x86_64-${VULKAN_SDK_VERSION}.tar.xz"
 
   _vulkan_sdk_dir=/tmp/vulkansdk
   mkdir -p $_vulkan_sdk_dir
@@ -37,12 +37,12 @@ install_vulkan_sdk() {
   curl --silent --show-error --location --fail --retry 3 \
     --output "${_tmp_archive}" "${_vulkan_sdk_url}"
 
-  tar -C "${_vulkan_sdk_dir}" -xzf "${_tmp_archive}"
+  tar -C "${_vulkan_sdk_dir}" -xJf "${_tmp_archive}"
 
   export PATH="${PATH}:${_vulkan_sdk_dir}/${VULKAN_SDK_VERSION}/x86_64/bin/"
 }
 
-VULKAN_SDK_VERSION="1.2.198.1"
+VULKAN_SDK_VERSION="1.3.296.0"
 
 install_swiftshader
 install_vulkan_sdk "${VULKAN_SDK_VERSION}"

.ci/scripts/test_llama.sh

@@ -208,7 +208,7 @@ prepare_artifacts_upload() {
 PARAMS="params.json"
 CHECKPOINT_FILE_NAME=""
 touch "${PARAMS}"
-if [[ "${MODEL_NAME}" == "stories110M" ]]; then
+if [[ "${MODEL_NAME}" == "llama" ]] || [[ "${MODEL_NAME}" == "stories"* ]] || [[ "${MODEL_NAME}" == "tinyllama" ]]; then
   CHECKPOINT_FILE_NAME="stories110M.pt"
   download_stories_model_artifacts
 else

.ci/scripts/test_llava.sh

@@ -41,7 +41,6 @@ EXECUTORCH_COMMON_CMAKE_ARGS="                      \
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON     \
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON     \
         -DEXECUTORCH_BUILD_XNNPACK=ON               \
-        -DEXECUTORCH_DO_NOT_USE_CXX11_ABI=ON        \
         -DEXECUTORCH_XNNPACK_SHARED_WORKSPACE=ON"
 
 cmake_install_executorch_libraries() {