Merge branch 'main' into save-automodel

* update

* update

.github/workflows/benchmark.yml

@@ -62,20 +62,6 @@ jobs:
         with:
           name: benchmark_test_reports
           path: benchmarks/${{ env.BASE_PATH }}
-      
-      # TODO: enable this once the connection problem has been resolved.
-      - name: Update benchmarking results to DB
-        env:
-          PGDATABASE: metrics
-          PGHOST: ${{ secrets.DIFFUSERS_BENCHMARKS_PGHOST }}
-          PGUSER: transformers_benchmarks
-          PGPASSWORD: ${{ secrets.DIFFUSERS_BENCHMARKS_PGPASSWORD }}
-          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-        run: |
-          git config --global --add safe.directory /__w/diffusers/diffusers
-          commit_id=$GITHUB_SHA
-          commit_msg=$(git show -s --format=%s "$commit_id" | cut -c1-70)
-          cd benchmarks && python populate_into_db.py "$BRANCH_NAME" "$commit_id" "$commit_msg"
 
       - name: Report success status
         if: ${{ success() }}

.github/workflows/pypi_publish.yaml

@@ -54,7 +54,6 @@ jobs:
           python -m pip install --upgrade pip
           pip install -U setuptools wheel twine
           pip install -U torch --index-url https://download.pytorch.org/whl/cpu
-          pip install -U transformers
 
       - name: Build the dist files
         run: python setup.py bdist_wheel && python setup.py sdist
@@ -69,6 +68,8 @@ jobs:
         run: |
           pip install diffusers && pip uninstall diffusers -y
           pip install -i https://test.pypi.org/simple/ diffusers
+          pip install -U transformers
+          python utils/print_env.py
           python -c "from diffusers import __version__; print(__version__)"
           python -c "from diffusers import DiffusionPipeline; pipe = DiffusionPipeline.from_pretrained('fusing/unet-ldm-dummy-update'); pipe()"
           python -c "from diffusers import DiffusionPipeline; pipe = DiffusionPipeline.from_pretrained('hf-internal-testing/tiny-stable-diffusion-pipe', safety_checker=None); pipe('ah suh du')"

benchmarks/populate_into_db.py

@@ -1,166 +0,0 @@
-import argparse
-import os
-import sys
-
-import gpustat
-import pandas as pd
-import psycopg2
-import psycopg2.extras
-from psycopg2.extensions import register_adapter
-from psycopg2.extras import Json
-
-
-register_adapter(dict, Json)
-
-FINAL_CSV_FILENAME = "collated_results.csv"
-# https://github.com/huggingface/transformers/blob/593e29c5e2a9b17baec010e8dc7c1431fed6e841/benchmark/init_db.sql#L27
-BENCHMARKS_TABLE_NAME = "benchmarks"
-MEASUREMENTS_TABLE_NAME = "model_measurements"
-
-
-def _init_benchmark(conn, branch, commit_id, commit_msg):
-    gpu_stats = gpustat.GPUStatCollection.new_query()
-    metadata = {"gpu_name": gpu_stats[0]["name"]}
-    repository = "huggingface/diffusers"
-    with conn.cursor() as cur:
-        cur.execute(
-            f"INSERT INTO {BENCHMARKS_TABLE_NAME} (repository, branch, commit_id, commit_message, metadata) VALUES (%s, %s, %s, %s, %s) RETURNING benchmark_id",
-            (repository, branch, commit_id, commit_msg, metadata),
-        )
-        benchmark_id = cur.fetchone()[0]
-        print(f"Initialised benchmark #{benchmark_id}")
-        return benchmark_id
-
-
-def parse_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "branch",
-        type=str,
-        help="The branch name on which the benchmarking is performed.",
-    )
-
-    parser.add_argument(
-        "commit_id",
-        type=str,
-        help="The commit hash on which the benchmarking is performed.",
-    )
-
-    parser.add_argument(
-        "commit_msg",
-        type=str,
-        help="The commit message associated with the commit, truncated to 70 characters.",
-    )
-    args = parser.parse_args()
-    return args
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    try:
-        conn = psycopg2.connect(
-            host=os.getenv("PGHOST"),
-            database=os.getenv("PGDATABASE"),
-            user=os.getenv("PGUSER"),
-            password=os.getenv("PGPASSWORD"),
-        )
-        print("DB connection established successfully.")
-    except Exception as e:
-        print(f"Problem during DB init: {e}")
-        sys.exit(1)
-
-    try:
-        benchmark_id = _init_benchmark(
-            conn=conn,
-            branch=args.branch,
-            commit_id=args.commit_id,
-            commit_msg=args.commit_msg,
-        )
-    except Exception as e:
-        print(f"Problem during initializing benchmark: {e}")
-        sys.exit(1)
-
-    cur = conn.cursor()
-
-    df = pd.read_csv(FINAL_CSV_FILENAME)
-
-    # Helper to cast values (or None) given a dtype
-    def _cast_value(val, dtype: str):
-        if pd.isna(val):
-            return None
-
-        if dtype == "text":
-            return str(val).strip()
-
-        if dtype == "float":
-            try:
-                return float(val)
-            except ValueError:
-                return None
-
-        if dtype == "bool":
-            s = str(val).strip().lower()
-            if s in ("true", "t", "yes", "1"):
-                return True
-            if s in ("false", "f", "no", "0"):
-                return False
-            if val in (1, 1.0):
-                return True
-            if val in (0, 0.0):
-                return False
-            return None
-
-        return val
-
-    try:
-        rows_to_insert = []
-        for _, row in df.iterrows():
-            scenario = _cast_value(row.get("scenario"), "text")
-            model_cls = _cast_value(row.get("model_cls"), "text")
-            num_params_B = _cast_value(row.get("num_params_B"), "float")
-            flops_G = _cast_value(row.get("flops_G"), "float")
-            time_plain_s = _cast_value(row.get("time_plain_s"), "float")
-            mem_plain_GB = _cast_value(row.get("mem_plain_GB"), "float")
-            time_compile_s = _cast_value(row.get("time_compile_s"), "float")
-            mem_compile_GB = _cast_value(row.get("mem_compile_GB"), "float")
-            fullgraph = _cast_value(row.get("fullgraph"), "bool")
-            mode = _cast_value(row.get("mode"), "text")
-
-            # If "github_sha" column exists in the CSV, cast it; else default to None
-            if "github_sha" in df.columns:
-                github_sha = _cast_value(row.get("github_sha"), "text")
-            else:
-                github_sha = None
-
-            measurements = {
-                "scenario": scenario,
-                "model_cls": model_cls,
-                "num_params_B": num_params_B,
-                "flops_G": flops_G,
-                "time_plain_s": time_plain_s,
-                "mem_plain_GB": mem_plain_GB,
-                "time_compile_s": time_compile_s,
-                "mem_compile_GB": mem_compile_GB,
-                "fullgraph": fullgraph,
-                "mode": mode,
-                "github_sha": github_sha,
-            }
-            rows_to_insert.append((benchmark_id, measurements))
-
-        # Batch-insert all rows
-        insert_sql = f"""
-        INSERT INTO {MEASUREMENTS_TABLE_NAME} (
-            benchmark_id,
-            measurements
-        )
-        VALUES (%s, %s);
-        """
-
-        psycopg2.extras.execute_batch(cur, insert_sql, rows_to_insert)
-        conn.commit()
-
-        cur.close()
-        conn.close()
-    except Exception as e:
-        print(f"Exception: {e}")
-        sys.exit(1)

docs/source/en/using-diffusers/automodel.md

@@ -97,5 +97,32 @@ If the custom model inherits from the [`ModelMixin`] class, it gets access to th
 > )
 > ```
 
+### Saving custom models
+
+Use [`~ConfigMixin.register_for_auto_class`] to add the `auto_map` entry to `config.json` automatically when saving. This avoids having to manually edit the config file.
+
+```py
+# my_model.py
+from diffusers import ModelMixin, ConfigMixin
+
+class MyCustomModel(ModelMixin, ConfigMixin):
+    ...
+
+MyCustomModel.register_for_auto_class("AutoModel")
+
+model = MyCustomModel(...)
+model.save_pretrained("./my_model")
+```
+
+The saved `config.json` will include the `auto_map` field.
+
+```json
+{
+  "auto_map": {
+    "AutoModel": "my_model.MyCustomModel"
+  }
+}
+```
+
 > [!NOTE]
 > Learn more about implementing custom models in the [Community components](../using-diffusers/custom_pipeline_overview#community-components) guide.

src/diffusers/configuration_utils.py

@@ -107,6 +107,38 @@ class ConfigMixin:
     has_compatibles = False
 
     _deprecated_kwargs = []
+    _auto_class = None
+
+    @classmethod
+    def register_for_auto_class(cls, auto_class="AutoModel"):
+        """
+        Register this class with the given auto class so that it can be loaded with `AutoModel.from_pretrained(...,
+        trust_remote_code=True)`.
+
+        When the config is saved, the resulting `config.json` will include an `auto_map` entry mapping the auto class
+        to this class's module and class name.
+
+        Args:
+            auto_class (`str` or type, *optional*, defaults to `"AutoModel"`):
+                The auto class to register this class with. Can be a string (e.g. `"AutoModel"`) or the class itself.
+                Currently only `"AutoModel"` is supported.
+
+        Example:
+
+        ```python
+        from diffusers import ModelMixin, ConfigMixin
+
+
+        class MyCustomModel(ModelMixin, ConfigMixin): ...
+
+
+        MyCustomModel.register_for_auto_class("AutoModel")
+        ```
+        """
+        if auto_class != "AutoModel":
+            raise ValueError(f"Only 'AutoModel' is supported, got '{auto_class}'.")
+
+        cls._auto_class = auto_class
 
     def register_to_config(self, **kwargs):
         if self.config_name is None:
@@ -621,6 +653,12 @@ class ConfigMixin:
         # pop the `_pre_quantization_dtype` as torch.dtypes are not serializable.
         _ = config_dict.pop("_pre_quantization_dtype", None)
 
+        if getattr(self, "_auto_class", None) is not None:
+            module = self.__class__.__module__.split(".")[-1]
+            auto_map = config_dict.get("auto_map", {})
+            auto_map[self._auto_class] = f"{module}.{self.__class__.__name__}"
+            config_dict["auto_map"] = auto_map
+
         return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
 
     def to_json_file(self, json_file_path: str | os.PathLike):

src/diffusers/utils/dynamic_modules_utils.py

@@ -299,7 +299,10 @@ def get_cached_module_file(
     # Download and cache module_file from the repo `pretrained_model_name_or_path` of grab it if it's a local file.
     pretrained_model_name_or_path = str(pretrained_model_name_or_path)
 
-    module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file)
+    if subfolder is not None:
+        module_file_or_url = os.path.join(pretrained_model_name_or_path, subfolder, module_file)
+    else:
+        module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file)
 
     if os.path.isfile(module_file_or_url):
         resolved_module_file = module_file_or_url
@@ -384,7 +387,11 @@ def get_cached_module_file(
                 if not os.path.exists(submodule_path / module_folder):
                     os.makedirs(submodule_path / module_folder)
             module_needed = f"{module_needed}.py"
-            shutil.copyfile(os.path.join(pretrained_model_name_or_path, module_needed), submodule_path / module_needed)
+            if subfolder is not None:
+                source_path = os.path.join(pretrained_model_name_or_path, subfolder, module_needed)
+            else:
+                source_path = os.path.join(pretrained_model_name_or_path, module_needed)
+            shutil.copyfile(source_path, submodule_path / module_needed)
     else:
         # Get the commit hash
         # TODO: we will get this info in the etag soon, so retrieve it from there and not here.

tests/models/test_models_auto.py

@@ -1,9 +1,15 @@
+import json
+import os
+import tempfile
 import unittest
 from unittest.mock import MagicMock, patch
 
+import torch
 from transformers import CLIPTextModel, LongformerModel
 
+from diffusers import ConfigMixin
 from diffusers.models import AutoModel, UNet2DConditionModel
+from diffusers.models.modeling_utils import ModelMixin
 
 
 class TestAutoModel(unittest.TestCase):
@@ -35,6 +41,45 @@ class TestAutoModel(unittest.TestCase):
         )
         assert isinstance(model, CLIPTextModel)
 
+    def test_load_dynamic_module_from_local_path_with_subfolder(self):
+        CUSTOM_MODEL_CODE = (
+            "import torch\n"
+            "from diffusers import ModelMixin, ConfigMixin\n"
+            "from diffusers.configuration_utils import register_to_config\n"
+            "\n"
+            "class CustomModel(ModelMixin, ConfigMixin):\n"
+            "    @register_to_config\n"
+            "    def __init__(self, hidden_size=8):\n"
+            "        super().__init__()\n"
+            "        self.linear = torch.nn.Linear(hidden_size, hidden_size)\n"
+            "\n"
+            "    def forward(self, x):\n"
+            "        return self.linear(x)\n"
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            subfolder = "custom_model"
+            model_dir = os.path.join(tmpdir, subfolder)
+            os.makedirs(model_dir)
+
+            with open(os.path.join(model_dir, "modeling.py"), "w") as f:
+                f.write(CUSTOM_MODEL_CODE)
+
+            config = {
+                "_class_name": "CustomModel",
+                "_diffusers_version": "0.0.0",
+                "auto_map": {"AutoModel": "modeling.CustomModel"},
+                "hidden_size": 8,
+            }
+            with open(os.path.join(model_dir, "config.json"), "w") as f:
+                json.dump(config, f)
+
+            torch.save({}, os.path.join(model_dir, "diffusion_pytorch_model.bin"))
+
+            model = AutoModel.from_pretrained(tmpdir, subfolder=subfolder, trust_remote_code=True)
+            assert model.__class__.__name__ == "CustomModel"
+            assert model.config["hidden_size"] == 8
+
 
 class TestAutoModelFromConfig(unittest.TestCase):
     @patch(
@@ -100,3 +145,51 @@ class TestAutoModelFromConfig(unittest.TestCase):
     def test_from_config_raises_on_none(self):
         with self.assertRaises(ValueError, msg="Please provide a `pretrained_model_name_or_path_or_dict`"):
             AutoModel.from_config(None)
+
+
+class TestRegisterForAutoClass(unittest.TestCase):
+    def test_register_for_auto_class_sets_attribute(self):
+        class DummyModel(ModelMixin, ConfigMixin):
+            config_name = "config.json"
+
+        DummyModel.register_for_auto_class("AutoModel")
+        self.assertEqual(DummyModel._auto_class, "AutoModel")
+
+    def test_register_for_auto_class_rejects_unsupported(self):
+        class DummyModel(ModelMixin, ConfigMixin):
+            config_name = "config.json"
+
+        with self.assertRaises(ValueError, msg="Only 'AutoModel' is supported"):
+            DummyModel.register_for_auto_class("AutoPipeline")
+
+    def test_auto_map_in_saved_config(self):
+        class DummyModel(ModelMixin, ConfigMixin):
+            config_name = "config.json"
+
+        DummyModel.register_for_auto_class("AutoModel")
+        model = DummyModel()
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            model.save_config(tmpdir)
+            config_path = os.path.join(tmpdir, "config.json")
+            with open(config_path, "r") as f:
+                config = json.load(f)
+
+        self.assertIn("auto_map", config)
+        self.assertIn("AutoModel", config["auto_map"])
+        module_name = DummyModel.__module__.split(".")[-1]
+        self.assertEqual(config["auto_map"]["AutoModel"], f"{module_name}.DummyModel")
+
+    def test_no_auto_map_without_register(self):
+        class DummyModel(ModelMixin, ConfigMixin):
+            config_name = "config.json"
+
+        model = DummyModel()
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            model.save_config(tmpdir)
+            config_path = os.path.join(tmpdir, "config.json")
+            with open(config_path, "r") as f:
+                config = json.load(f)
+
+        self.assertNotIn("auto_map", config)

tests/quantization/torchao/test_torchao.py

@@ -74,7 +74,7 @@ if is_torchao_available():
 
 @require_torch
 @require_torch_accelerator
-@require_torchao_version_greater_or_equal("0.7.0")
+@require_torchao_version_greater_or_equal("0.14.0")
 class TorchAoConfigTest(unittest.TestCase):
     def test_to_dict(self):
         """
@@ -132,7 +132,7 @@ class TorchAoConfigTest(unittest.TestCase):
 # Slices for these tests have been obtained on our aws-g6e-xlarge-plus runners
 @require_torch
 @require_torch_accelerator
-@require_torchao_version_greater_or_equal("0.7.0")
+@require_torchao_version_greater_or_equal("0.14.0")
 class TorchAoTest(unittest.TestCase):
     def tearDown(self):
         gc.collect()
@@ -587,7 +587,7 @@ class TorchAoTest(unittest.TestCase):
 # Slices for these tests have been obtained on our aws-g6e-xlarge-plus runners
 @require_torch
 @require_torch_accelerator
-@require_torchao_version_greater_or_equal("0.7.0")
+@require_torchao_version_greater_or_equal("0.14.0")
 class TorchAoSerializationTest(unittest.TestCase):
     model_name = "hf-internal-testing/tiny-flux-pipe"
 
@@ -698,23 +698,22 @@ class TorchAoSerializationTest(unittest.TestCase):
         self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device)
 
 
-@require_torchao_version_greater_or_equal("0.7.0")
+@require_torchao_version_greater_or_equal("0.14.0")
 class TorchAoCompileTest(QuantCompileTests, unittest.TestCase):
     @property
     def quantization_config(self):
         return PipelineQuantizationConfig(
-            quant_mapping={
-                "transformer": TorchAoConfig(quant_type="int8_weight_only"),
-            },
+            quant_mapping={"transformer": TorchAoConfig(Int8WeightOnlyConfig())},
         )
 
-    @unittest.skip(
-        "Changing the device of AQT tensor with module._apply (called from doing module.to() in accelerate) does not work "
-        "when compiling."
-    )
     def test_torch_compile_with_cpu_offload(self):
+        pipe = self._init_pipeline(self.quantization_config, torch.bfloat16)
+        pipe.enable_model_cpu_offload()
+        # No compilation because it fails with:
         # RuntimeError: _apply(): Couldn't swap Linear.weight
-        super().test_torch_compile_with_cpu_offload()
+
+        # small resolutions to ensure speedy execution.
+        pipe("a dog", num_inference_steps=2, max_sequence_length=16, height=256, width=256)
 
     @parameterized.expand([False, True])
     @unittest.skip(
@@ -745,7 +744,7 @@ class TorchAoCompileTest(QuantCompileTests, unittest.TestCase):
 # Slices for these tests have been obtained on our aws-g6e-xlarge-plus runners
 @require_torch
 @require_torch_accelerator
-@require_torchao_version_greater_or_equal("0.7.0")
+@require_torchao_version_greater_or_equal("0.14.0")
 @slow
 @nightly
 class SlowTorchAoTests(unittest.TestCase):
@@ -907,7 +906,7 @@ class SlowTorchAoTests(unittest.TestCase):
 
 @require_torch
 @require_torch_accelerator
-@require_torchao_version_greater_or_equal("0.7.0")
+@require_torchao_version_greater_or_equal("0.14.0")
 @slow
 @nightly
 class SlowTorchAoPreserializedModelTests(unittest.TestCase):