Update on "Reuse GELU implementation from PyTorch core"

kernels/optimized doesn't need to support embedded systems, so it can just take a header-only dep on PyTorch.

Note that, because we will pick up Sleef internally and ignore it
externally thanks to ATen vec, this PR gets to enable optimized GELU in OSS.

Testing: CI to make sure this doesn't break mobile build modes; happy to take advice on anything not currently covered that might break.

Differential Revision: [D66335522](https://our.internmc.facebook.com/intern/diff/D66335522/)

[ghstack-poisoned]

Author: swolchok

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-cb4af2b4752220c3ca3de6e7e95b3a6fdc31f794
+4eeb6f34ef415aa8701a36453380e9c1ba2c8f3a

.ci/docker/requirements-ci.txt

@@ -1,5 +1,5 @@
 mpmath==1.3.0
-numpy==2.0.0; python_version >= '3.10'
+numpy>=2.0.0; python_version >= '3.10'
 PyYAML==6.0.1
 ruamel.yaml==0.17.32
 sympy==1.12
@@ -8,7 +8,7 @@ tomli==2.0.1
 torchsr==1.0.4
 transformers==4.47.1
 zstd==1.5.5.1
-pandas==2.2.2; python_version >= '3.10'
+pandas>=2.2.2; python_version >= '3.10'
 pytest==7.2.0
 pytest-cov==4.1.0
 expecttest==0.1.6
@@ -21,7 +21,7 @@ sphinx-gallery==0.14.0
 breathe==4.34.0
 exhale==0.2.3
 docutils==0.16
-matplotlib==3.9.4
+matplotlib>=3.9.4
 # PyTorch Theme
 -e git+https://github.com/pytorch/pytorch_sphinx_theme.git#egg=pytorch_sphinx_theme
 myst-parser==0.18.1

.ci/scripts/gather_benchmark_configs.py

@@ -10,7 +10,7 @@
 import os
 import re
 import sys
-from typing import Any, Dict, List
+from typing import Any, Dict, List, NamedTuple
 
 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "../..")))
 from examples.models import MODEL_NAME_TO_MODEL
@@ -47,6 +47,50 @@
 }
 
 
+class DisabledConfig(NamedTuple):
+    config_name: str
+    github_issue: str  # Link to the GitHub issue
+
+
+# Updated DISABLED_CONFIGS
+DISABLED_CONFIGS: Dict[str, List[DisabledConfig]] = {
+    "resnet50": [
+        DisabledConfig(
+            config_name="qnn_q8",
+            github_issue="https://github.com/pytorch/executorch/issues/7892",
+        ),
+    ],
+    "w2l": [
+        DisabledConfig(
+            config_name="qnn_q8",
+            github_issue="https://github.com/pytorch/executorch/issues/7634",
+        ),
+    ],
+    "mobilebert": [
+        DisabledConfig(
+            config_name="mps",
+            github_issue="https://github.com/pytorch/executorch/issues/7904",
+        ),
+        DisabledConfig(
+            config_name="qnn_q8",
+            github_issue="https://github.com/pytorch/executorch/issues/7946",
+        ),
+    ],
+    "edsr": [
+        DisabledConfig(
+            config_name="mps",
+            github_issue="https://github.com/pytorch/executorch/issues/7905",
+        ),
+    ],
+    "llama": [
+        DisabledConfig(
+            config_name="mps",
+            github_issue="https://github.com/pytorch/executorch/issues/7907",
+        ),
+    ],
+}
+
+
 def extract_all_configs(data, target_os=None):
     if isinstance(data, dict):
         # If target_os is specified, include "xplat" and the specified branch
@@ -117,6 +161,14 @@ def generate_compatible_configs(model_name: str, target_os=None) -> List[str]:
         # Skip unknown models with a warning
         logging.warning(f"Unknown or invalid model name '{model_name}'. Skipping.")
 
+    # Remove disabled configs for the given model
+    disabled_configs = DISABLED_CONFIGS.get(model_name, [])
+    disabled_config_names = {disabled.config_name for disabled in disabled_configs}
+    for disabled in disabled_configs:
+        print(
+            f"Excluding disabled config: '{disabled.config_name}' for model '{model_name}' on '{target_os}'. Linked GitHub issue: {disabled.github_issue}"
+        )
+    configs = [config for config in configs if config not in disabled_config_names]
     return configs
 
 

.ci/scripts/test_llama.sh

@@ -112,6 +112,8 @@ fi
 
 if [[ "${MODE}" =~ .*quantize_kv.* ]]; then
   QUANTIZE_KV_CACHE=ON
+  # quantize_kv cache transform uses custom kv cache update op
+  CUSTOM=ON
 else
   QUANTIZE_KV_CACHE=OFF
 fi

.ci/scripts/test_model.sh

@@ -170,6 +170,13 @@ test_model_with_qnn() {
     EXPORT_SCRIPT=inception_v3
   elif [[ "${MODEL_NAME}" == "vit" ]]; then
     EXPORT_SCRIPT=torchvision_vit
+  elif [[ "${MODEL_NAME}" == "edsr" ]]; then
+    EXPORT_SCRIPT=edsr
+    # Additional deps for edsr
+    pip install piq
+  else
+    echo "Unsupported model $MODEL_NAME"
+    exit 1
   fi
 
   # Use SM8450 for S22, SM8550 for S23, and SM8560 for S24

.ci/scripts/tests/test_gather_benchmark_configs.py

@@ -1,36 +1,41 @@
 import importlib.util
 import os
+import re
 import subprocess
 import sys
 import unittest
 from unittest.mock import mock_open, patch
 
 import pytest
 
-# Dynamically import the script
-script_path = os.path.join(".ci", "scripts", "gather_benchmark_configs.py")
-spec = importlib.util.spec_from_file_location("gather_benchmark_configs", script_path)
-gather_benchmark_configs = importlib.util.module_from_spec(spec)
-spec.loader.exec_module(gather_benchmark_configs)
-
 
 @pytest.mark.skipif(
     sys.platform != "linux", reason="The script under test runs on Linux runners only"
 )
 class TestGatehrBenchmarkConfigs(unittest.TestCase):
 
+    @classmethod
+    def setUpClass(cls):
+        # Dynamically import the script
+        script_path = os.path.join(".ci", "scripts", "gather_benchmark_configs.py")
+        spec = importlib.util.spec_from_file_location(
+            "gather_benchmark_configs", script_path
+        )
+        cls.gather_benchmark_configs = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(cls.gather_benchmark_configs)
+
     def test_extract_all_configs_android(self):
-        android_configs = gather_benchmark_configs.extract_all_configs(
-            gather_benchmark_configs.BENCHMARK_CONFIGS, "android"
+        android_configs = self.gather_benchmark_configs.extract_all_configs(
+            self.gather_benchmark_configs.BENCHMARK_CONFIGS, "android"
         )
         self.assertIn("xnnpack_q8", android_configs)
         self.assertIn("qnn_q8", android_configs)
         self.assertIn("llama3_spinquant", android_configs)
         self.assertIn("llama3_qlora", android_configs)
 
     def test_extract_all_configs_ios(self):
-        ios_configs = gather_benchmark_configs.extract_all_configs(
-            gather_benchmark_configs.BENCHMARK_CONFIGS, "ios"
+        ios_configs = self.gather_benchmark_configs.extract_all_configs(
+            self.gather_benchmark_configs.BENCHMARK_CONFIGS, "ios"
         )
 
         self.assertIn("xnnpack_q8", ios_configs)
@@ -40,51 +45,114 @@ def test_extract_all_configs_ios(self):
         self.assertIn("llama3_spinquant", ios_configs)
         self.assertIn("llama3_qlora", ios_configs)
 
+    def test_skip_disabled_configs(self):
+        # Use patch as a context manager to avoid modifying DISABLED_CONFIGS and BENCHMARK_CONFIGS
+        with patch.dict(
+            self.gather_benchmark_configs.DISABLED_CONFIGS,
+            {
+                "mv3": [
+                    self.gather_benchmark_configs.DisabledConfig(
+                        config_name="disabled_config1",
+                        github_issue="https://github.com/org/repo/issues/123",
+                    ),
+                    self.gather_benchmark_configs.DisabledConfig(
+                        config_name="disabled_config2",
+                        github_issue="https://github.com/org/repo/issues/124",
+                    ),
+                ]
+            },
+        ), patch.dict(
+            self.gather_benchmark_configs.BENCHMARK_CONFIGS,
+            {
+                "ios": [
+                    "disabled_config1",
+                    "disabled_config2",
+                    "enabled_config1",
+                    "enabled_config2",
+                ]
+            },
+        ):
+            result = self.gather_benchmark_configs.generate_compatible_configs(
+                "mv3", target_os="ios"
+            )
+
+            # Assert that disabled configs are excluded
+            self.assertNotIn("disabled_config1", result)
+            self.assertNotIn("disabled_config2", result)
+            # Assert enabled configs are included
+            self.assertIn("enabled_config1", result)
+            self.assertIn("enabled_config2", result)
+
+    def test_disabled_configs_have_github_links(self):
+        github_issue_regex = re.compile(r"https://github\.com/.+/.+/issues/\d+")
+
+        for (
+            model_name,
+            disabled_configs,
+        ) in self.gather_benchmark_configs.DISABLED_CONFIGS.items():
+            for disabled in disabled_configs:
+                with self.subTest(model_name=model_name, config=disabled.config_name):
+                    # Assert that disabled is an instance of DisabledConfig
+                    self.assertIsInstance(
+                        disabled, self.gather_benchmark_configs.DisabledConfig
+                    )
+
+                    # Assert that github_issue is provided and matches the expected pattern
+                    self.assertTrue(
+                        disabled.github_issue
+                        and github_issue_regex.match(disabled.github_issue),
+                        f"Invalid or missing GitHub issue link for '{disabled.config_name}' in model '{model_name}'.",
+                    )
+
     def test_generate_compatible_configs_llama_model(self):
         model_name = "meta-llama/Llama-3.2-1B"
         target_os = "ios"
-        result = gather_benchmark_configs.generate_compatible_configs(
+        result = self.gather_benchmark_configs.generate_compatible_configs(
             model_name, target_os
         )
         expected = ["llama3_fb16", "llama3_coreml_ane"]
         self.assertEqual(result, expected)
 
         target_os = "android"
-        result = gather_benchmark_configs.generate_compatible_configs(
+        result = self.gather_benchmark_configs.generate_compatible_configs(
             model_name, target_os
         )
         expected = ["llama3_fb16"]
         self.assertEqual(result, expected)
 
     def test_generate_compatible_configs_quantized_llama_model(self):
         model_name = "meta-llama/Llama-3.2-1B-Instruct-SpinQuant_INT4_EO8"
-        result = gather_benchmark_configs.generate_compatible_configs(model_name, None)
+        result = self.gather_benchmark_configs.generate_compatible_configs(
+            model_name, None
+        )
         expected = ["llama3_spinquant"]
         self.assertEqual(result, expected)
 
         model_name = "meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8"
-        result = gather_benchmark_configs.generate_compatible_configs(model_name, None)
+        result = self.gather_benchmark_configs.generate_compatible_configs(
+            model_name, None
+        )
         expected = ["llama3_qlora"]
         self.assertEqual(result, expected)
 
     def test_generate_compatible_configs_non_genai_model(self):
         model_name = "mv2"
         target_os = "xplat"
-        result = gather_benchmark_configs.generate_compatible_configs(
+        result = self.gather_benchmark_configs.generate_compatible_configs(
             model_name, target_os
         )
         expected = ["xnnpack_q8"]
         self.assertEqual(result, expected)
 
         target_os = "android"
-        result = gather_benchmark_configs.generate_compatible_configs(
+        result = self.gather_benchmark_configs.generate_compatible_configs(
             model_name, target_os
         )
         expected = ["xnnpack_q8", "qnn_q8"]
         self.assertEqual(result, expected)
 
         target_os = "ios"
-        result = gather_benchmark_configs.generate_compatible_configs(
+        result = self.gather_benchmark_configs.generate_compatible_configs(
             model_name, target_os
         )
         expected = ["xnnpack_q8", "coreml_fp16", "mps"]
@@ -93,22 +161,22 @@ def test_generate_compatible_configs_non_genai_model(self):
     def test_generate_compatible_configs_unknown_model(self):
         model_name = "unknown_model"
         target_os = "ios"
-        result = gather_benchmark_configs.generate_compatible_configs(
+        result = self.gather_benchmark_configs.generate_compatible_configs(
             model_name, target_os
         )
         self.assertEqual(result, [])
 
     def test_is_valid_huggingface_model_id_valid(self):
         valid_model = "meta-llama/Llama-3.2-1B"
         self.assertTrue(
-            gather_benchmark_configs.is_valid_huggingface_model_id(valid_model)
+            self.gather_benchmark_configs.is_valid_huggingface_model_id(valid_model)
         )
 
     @patch("builtins.open", new_callable=mock_open)
     @patch("os.getenv", return_value=None)
     def test_set_output_no_github_env(self, mock_getenv, mock_file):
         with patch("builtins.print") as mock_print:
-            gather_benchmark_configs.set_output("test_name", "test_value")
+            self.gather_benchmark_configs.set_output("test_name", "test_value")
             mock_print.assert_called_with("::set-output name=test_name::test_value")
 
     def test_device_pools_contains_all_devices(self):
@@ -120,7 +188,7 @@ def test_device_pools_contains_all_devices(self):
             "google_pixel_8_pro",
         ]
         for device in expected_devices:
-            self.assertIn(device, gather_benchmark_configs.DEVICE_POOLS)
+            self.assertIn(device, self.gather_benchmark_configs.DEVICE_POOLS)
 
     def test_gather_benchmark_configs_cli(self):
         args = {

.github/workflows/_android.yml

@@ -7,7 +7,10 @@ on:
 jobs:
   build-llm-demo:
     name: build-llm-demo
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-clang12-android

.github/workflows/_unittest.yml

@@ -14,7 +14,10 @@ on:
 
 jobs:
   linux:
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     with:
       runner: linux.2xlarge
       docker-image: ${{ inputs.docker-image }}

.github/workflows/android-perf.yml

@@ -155,7 +155,10 @@ jobs:
 
   export-models:
     name: export-models
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     needs: set-parameters
     secrets: inherit
     strategy:
@@ -332,7 +335,10 @@ jobs:
 
   build-benchmark-app:
     name: build-benchmark-app
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     needs: set-parameters
     with:
       runner: linux.2xlarge

.github/workflows/android-release-artifacts.yml

@@ -31,7 +31,10 @@ jobs:
   build-aar:
     name: build-aar
     needs: check-if-aar-exists
-    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-clang12-android