Merge branch 'pytorch:main' into pr_model_improve

Author: Michiel-Olieslagers

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-7ae0ce6360b6e4f944906502d20da24c04debee5
+59d5cf083b4f860dea76fe8936076177f9367f10

.ci/scripts/setup-openvino.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# 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.
+
+set -ex
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+git clone https://github.com/openvinotoolkit/openvino.git
+cd openvino && git checkout releases/2025/1
+git submodule update --init --recursive
+sudo ./install_build_dependencies.sh
+mkdir build && cd build
+cmake .. -DCMAKE_BUILD_TYPE=Release -DENABLE_PYTHON=ON
+make -j$(nproc)
+
+cd ..
+cmake --install build --prefix dist
+
+source dist/setupvars.sh
+cd ../backends/openvino
+pip install -r requirements.txt
+cd scripts
+./openvino_build.sh --enable_python

.ci/scripts/test_llama.sh

@@ -269,7 +269,7 @@ $PYTHON_EXECUTABLE -m examples.models.llama.export_llama ${EXPORT_ARGS}
 
 # Create tokenizer.bin.
 echo "Creating tokenizer.bin"
-$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 
 
 RUNTIME_ARGS="--model_path=${EXPORTED_MODEL_NAME} --tokenizer_path=tokenizer.bin --prompt=Once --temperature=0 --seq_len=10 --warmup=1"

.ci/scripts/test_llama_torchao_lowbit.sh

@@ -55,7 +55,7 @@ cmake --build cmake-out/examples/models/llama -j16 --config Release
 download_stories_model_artifacts
 
 echo "Creating tokenizer.bin"
-$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 
 # Export model
 LLAMA_CHECKPOINT=stories110M.pt

.ci/scripts/test_openvino.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+# 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.
+
+set -ex
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+source openvino/dist/setupvars.sh
+cd backends/openvino/tests
+python test_runner.py --test_type ops
+python test_runner.py --test_type models

.ci/scripts/test_phi_3_mini.sh

@@ -56,7 +56,7 @@ cmake_build_phi_3_mini() {
 prepare_tokenizer() {
   echo "Downloading and converting tokenizer.model"
   wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/tokenizer.model?download=true"
-  $PYTHON_EXECUTABLE -m executorch.extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+  $PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 }
 
 # Export phi-3-mini model to pte

.ci/scripts/test_qnn_static_llama.sh

@@ -5,7 +5,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-set -exu
+set -euxo pipefail
 
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
@@ -30,7 +30,7 @@ pip install graphviz
 # Download stories llama110m artifacts
 download_stories_model_artifacts
 echo "Creating tokenizer.bin"
-$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
 
 set +e
 # Compile only as weight sharing is not applicable on x86
@@ -56,4 +56,3 @@ if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ]; then
 else
     exit 0
 fi
-set -e

.ci/scripts/utils.sh

@@ -32,7 +32,7 @@ install_executorch() {
   which pip
   # Install executorch, this assumes that Executorch is checked out in the
   # current directory.
-  ./install_executorch.sh --pybind xnnpack "$@"
+  ./install_executorch.sh "$@"
   # Just print out the list of packages for debugging
   pip list
 }

.github/release.yml

@@ -0,0 +1,96 @@
+# .github/release.yml
+
+changelog:
+  exclude:
+    labels:
+      - ignore-for-release
+  categories:
+    - title: Breaking Changes 🛠
+      labels:
+        - Semver-Major
+        - breaking-change
+    - title: API
+      labels:
+        - "release notes: api"
+    - title: ARM
+      labels:
+        - "release notes: arm"
+        - "module: arm"
+        - "partner: arm"
+    - title: NXP
+      labels:
+        - "release notes: nxp"
+        - "module: nxp"
+    - title: Exir
+      labels:
+        - "release notes: exir"
+        - "module: exir"
+    - title: Misc
+      labels:
+        - "release notes: misc"
+    - title: Apple
+      labels:
+        - "release notes: apple"
+        - "module: coreml"
+        - "module: mps"
+    - title: Android
+      labels:
+        - "module: android"
+    - title: IOS
+      labels:
+        - "module: ios"
+    - title: Build
+      labels:
+        - "release notes: build"
+    - title: Vulkan
+      labels:
+        - "release notes: vulkan"
+        - "module: vulkan"
+    - title: Cadence
+      labels:
+        - "release notes: cadence"
+        - "module: cadence"
+    - title: Runtime
+      labels:
+        - "release notes: runtime"
+        - "module: runtime"
+    - title: XNNPACK
+      labels:
+        - "release notes: xnnpack"
+        - "module: xnnpack"
+    - title: Devtools
+      labels:
+        - "release notes: devtools"   
+        - "module: devtools"
+    - title: Examples
+      labels:
+        - "release notes: examples"
+    - title: LLM
+      labels:
+        - "module: llm"
+    - title: Mediatek
+      labels:
+        - "release notes: mediatek"
+        - "partner: mediatek"
+    - title: Openvino
+      labels:
+        - "release notes: openvino"
+    - title: Qualcomm
+      labels:
+        - "release notes: qualcomm"
+        - "partner: qualcomm"
+        - "module: qnn"
+    - title: Training
+      labels:
+        - "release notes: training"
+        - "module: training"
+    - title: Quantization
+      labels:
+        - "release notes: quantization" 
+    - title: Ops & kernels
+      labels:
+        - "release notes: ops & kernels"
+        - "module: kernels"
+    - title: Other Changes
+      labels:
+        - "*"

.github/scripts/extract_benchmark_results.py

@@ -360,6 +360,7 @@ def transform(
                     "app_type": app_type,
                     # Just keep a copy of the benchmark config here
                     "benchmark_config": json.dumps(benchmark_config),
+                    "job_conclusion": "SUCCESS",
                 },
             },
             "model": {
@@ -455,7 +456,7 @@ def transform_failure_record(
         },
         "metric": {
             "name": "FAILURE_REPORT",
-            "benchmark_values": 0,
+            "benchmark_values": [0],
             "target_value": 0,
             "extra_info": {
                 "method": "",

.github/scripts/run_nm.py

@@ -0,0 +1,171 @@
+# 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 re
+import subprocess
+import sys
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Union
+
+
+@dataclass
+class Symbol:
+    name: str
+    addr: int
+    size: int
+    symbol_type: str
+
+
+class Parser:
+    def __init__(self, elf: str, toolchain_prefix: str = "", filter=None):
+        self.elf = elf
+        self.toolchain_prefix = toolchain_prefix
+        self.symbols: Dict[str, Symbol] = self._get_nm_output()
+        self.filter = filter
+
+    @staticmethod
+    def run_nm(
+        elf_file_path: str, args: Optional[List[str]] = None, nm: str = "nm"
+    ) -> str:
+        """
+        Run the nm command on the specified ELF file.
+        """
+        args = [] if args is None else args
+        cmd = [nm] + args + [elf_file_path]
+        try:
+            result = subprocess.run(cmd, check=True, capture_output=True, text=True)
+            return result.stdout
+        except FileNotFoundError:
+            print(f"Error: 'nm' command not found. Please ensure it's installed.")
+            sys.exit(1)
+        except subprocess.CalledProcessError as e:
+            print(f"Error running nm on {elf_file_path}: {e}")
+            print(f"stderr: {e.stderr}")
+            sys.exit(1)
+
+    def _get_nm_output(self) -> Dict[str, Symbol]:
+        args = [
+            "--print-size",
+            "--size-sort",
+            "--reverse-sort",
+            "--demangle",
+            "--format=bsd",
+        ]
+        output = Parser.run_nm(
+            self.elf,
+            args,
+            nm=self.toolchain_prefix + "nm" if self.toolchain_prefix else "nm",
+        )
+        lines = output.splitlines()
+        symbols = []
+        symbol_pattern = re.compile(
+            r"(?P<addr>[0-9a-fA-F]+)\s+(?P<size>[0-9a-fA-F]+)\s+(?P<type>\w)\s+(?P<name>.+)"
+        )
+
+        def parse_line(line: str) -> Optional[Symbol]:
+
+            match = symbol_pattern.match(line)
+            if match:
+                addr = int(match.group("addr"), 16)
+                size = int(match.group("size"), 16)
+                type_ = match.group("type").strip().strip("\n")
+                name = match.group("name").strip().strip("\n")
+                return Symbol(name=name, addr=addr, size=size, symbol_type=type_)
+            return None
+
+        for line in lines:
+            symbol = parse_line(line)
+            if symbol:
+                symbols.append(symbol)
+
+        assert len(symbols) > 0, "No symbols found in nm output"
+        if len(symbols) != len(lines):
+            print(
+                "** Warning: Not all lines were parsed, check the output of nm. Parsed {len(symbols)} lines, given {len(lines)}"
+            )
+        if any(symbol.size == 0 for symbol in symbols):
+            print("** Warning: Some symbols have zero size, check the output of nm.")
+
+        # TODO: Populate the section and module fields from the linker map if available (-Wl,-Map=linker.map)
+        return {symbol.name: symbol for symbol in symbols}
+
+    def print(self):
+        print(f"Elf: {self.elf}")
+
+        def print_table(filter=None, filter_name=None):
+            print("\nAddress\t\tSize\tType\tName")
+            # Apply filter and sort symbols
+            symbols_to_print = {
+                name: sym
+                for name, sym in self.symbols.items()
+                if not filter or filter(sym)
+            }
+            sorted_symbols = sorted(
+                symbols_to_print.items(), key=lambda x: x[1].size, reverse=True
+            )
+
+            # Print symbols and calculate total size
+            size_total = 0
+            for name, sym in sorted_symbols:
+                print(f"{hex(sym.addr)}\t\t{sym.size}\t{sym.symbol_type}\t{sym.name}")
+                size_total += sym.size
+
+            # Print summary
+            symbol_percent = len(symbols_to_print) / len(self.symbols) * 100
+            print("-----")
+            print(f"> Total bytes: {size_total}")
+            print(
+                f"Counted: {len(symbols_to_print)}/{len(self.symbols)}, {symbol_percent:0.2f}% (filter: '{filter_name}')"
+            )
+            print("=====\n")
+
+        # Print tables with different filters
+        def is_executorch_symbol(s):
+            return "executorch" in s.name or s.name.startswith("et")
+
+        FILTER_NAME_TO_FILTER_AND_LABEL = {
+            "all": (None, "All"),
+            "executorch": (is_executorch_symbol, "ExecuTorch"),
+            "executorch_text": (
+                lambda s: is_executorch_symbol(s) and s.symbol_type.lower() == "t",
+                "ExecuTorch .text",
+            ),
+        }
+
+        filter_func, label = FILTER_NAME_TO_FILTER_AND_LABEL.get(
+            self.filter, FILTER_NAME_TO_FILTER_AND_LABEL["all"]
+        )
+        print_table(filter_func, label)
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser(
+        description="Process ELF file and linker map file."
+    )
+    parser.add_argument(
+        "-e", "--elf-file-path", required=True, help="Path to the ELF file"
+    )
+    parser.add_argument(
+        "-f",
+        "--filter",
+        required=False,
+        default="all",
+        help="Filter symbols by pre-defined filters",
+        choices=["all", "executorch", "executorch_text"],
+    )
+    parser.add_argument(
+        "-p",
+        "--toolchain-prefix",
+        required=False,
+        default="",
+        help="Optional toolchain prefix for nm",
+    )
+
+    args = parser.parse_args()
+    p = Parser(args.elf_file_path, args.toolchain_prefix, filter=args.filter)
+    p.print()