Revert "[tests] Changes to the torch.compile() CI and tests"

.github/workflows/benchmark.yml

@@ -23,7 +23,7 @@ jobs:
     runs-on:
       group: aws-g6-4xlarge-plus
     container:
-      image: diffusers/diffusers-pytorch-cuda
+      image: diffusers/diffusers-pytorch-compile-cuda
       options: --shm-size "16gb" --ipc host --gpus 0
     steps:
       - name: Checkout diffusers

.github/workflows/build_docker_images.yml

@@ -41,12 +41,6 @@ jobs:
         run: |
           CHANGED_FILES="${{ steps.file_changes.outputs.all }}"
           for FILE in $CHANGED_FILES; do
-            # skip anything that isn’t still on disk
-            if [[ ! -f "$FILE" ]]; then
-              echo "Skipping removed file $FILE"
-              continue
-            fi
-
             if [[ "$FILE" == docker/*Dockerfile ]]; then
               DOCKER_PATH="${FILE%/Dockerfile}"
               DOCKER_TAG=$(basename "$DOCKER_PATH")
@@ -71,7 +65,7 @@ jobs:
         image-name:
           - diffusers-pytorch-cpu
           - diffusers-pytorch-cuda
-          - diffusers-pytorch-cuda
+          - diffusers-pytorch-compile-cuda
           - diffusers-pytorch-xformers-cuda
           - diffusers-pytorch-minimum-cuda
           - diffusers-flax-cpu

.github/workflows/nightly_tests.yml

@@ -188,7 +188,7 @@ jobs:
       group: aws-g4dn-2xlarge
 
     container:
-      image: diffusers/diffusers-pytorch-cuda
+      image: diffusers/diffusers-pytorch-compile-cuda
       options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:

.github/workflows/push_tests.yml

@@ -262,7 +262,7 @@ jobs:
       group: aws-g4dn-2xlarge
 
     container:
-      image: diffusers/diffusers-pytorch-cuda
+      image: diffusers/diffusers-pytorch-compile-cuda
       options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:

.github/workflows/release_tests_fast.yml

@@ -316,7 +316,7 @@ jobs:
       group: aws-g4dn-2xlarge
 
     container:
-      image: diffusers/diffusers-pytorch-cuda
+      image: diffusers/diffusers-pytorch-compile-cuda
       options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -0,0 +1,50 @@
+FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get -y update \
+    && apt-get install -y software-properties-common \
+    && add-apt-repository ppa:deadsnakes/ppa
+
+RUN apt install -y bash \
+    build-essential \
+    git \
+    git-lfs \
+    curl \
+    ca-certificates \
+    libsndfile1-dev \
+    libgl1 \
+    python3.10 \
+    python3.10-dev \
+    python3-pip \
+    python3.10-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3.10 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
+    python3.10 -m uv pip install --no-cache-dir \
+    torch \
+    torchvision \
+    torchaudio \
+    invisible_watermark && \
+    python3.10 -m pip install --no-cache-dir \
+    accelerate \
+    datasets \
+    hf-doc-builder \
+    huggingface-hub \
+    hf_transfer \
+    Jinja2 \
+    librosa \
+    numpy==1.26.4 \
+    scipy \
+    tensorboard \
+    transformers \
+    hf_transfer
+
+CMD ["/bin/bash"]

tests/models/test_modeling_common.py

@@ -1748,14 +1748,14 @@ class TorchCompileTesterMixin:
     def setUp(self):
         # clean up the VRAM before each test
         super().setUp()
-        torch.compiler.reset()
+        torch._dynamo.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
     def tearDown(self):
         # clean up the VRAM after each test in case of CUDA runtime errors
         super().tearDown()
-        torch.compiler.reset()
+        torch._dynamo.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
@@ -1764,17 +1764,13 @@ def tearDown(self):
     @is_torch_compile
     @slow
     def test_torch_compile_recompilation_and_graph_break(self):
-        torch.compiler.reset()
+        torch._dynamo.reset()
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
 
         model = self.model_class(**init_dict).to(torch_device)
         model = torch.compile(model, fullgraph=True)
 
-        with (
-            torch._inductor.utils.fresh_inductor_cache(),
-            torch._dynamo.config.patch(error_on_recompile=True),
-            torch.no_grad(),
-        ):
+        with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
             _ = model(**inputs_dict)
             _ = model(**inputs_dict)
 
@@ -1802,7 +1798,7 @@ def tearDown(self):
         # It is critical that the dynamo cache is reset for each test. Otherwise, if the test re-uses the same model,
         # there will be recompilation errors, as torch caches the model when run in the same process.
         super().tearDown()
-        torch.compiler.reset()
+        torch._dynamo.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
@@ -1919,7 +1915,7 @@ def test_hotswapping_model(self, rank0, rank1):
     def test_hotswapping_compiled_model_linear(self, rank0, rank1):
         # It's important to add this context to raise an error on recompilation
         target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
-        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
+        with torch._dynamo.config.patch(error_on_recompile=True):
             self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)
 
     @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
@@ -1929,7 +1925,7 @@ def test_hotswapping_compiled_model_conv2d(self, rank0, rank1):
 
         # It's important to add this context to raise an error on recompilation
         target_modules = ["conv", "conv1", "conv2"]
-        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
+        with torch._dynamo.config.patch(error_on_recompile=True):
             self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)
 
     @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
@@ -1939,7 +1935,7 @@ def test_hotswapping_compiled_model_both_linear_and_conv2d(self, rank0, rank1):
 
         # It's important to add this context to raise an error on recompilation
         target_modules = ["to_q", "conv"]
-        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
+        with torch._dynamo.config.patch(error_on_recompile=True):
             self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)
 
     @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa

tests/models/transformers/test_models_transformer_hunyuan_video.py

@@ -19,16 +19,20 @@
 from diffusers import HunyuanVideoTransformer3DModel
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
+    is_torch_compile,
+    require_torch_2,
+    require_torch_gpu,
+    slow,
     torch_device,
 )
 
-from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+from ..test_modeling_common import ModelTesterMixin
 
 
 enable_full_determinism()
 
 
-class HunyuanVideoTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
+class HunyuanVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -92,8 +96,23 @@ def test_gradient_checkpointing_is_applied(self):
         expected_set = {"HunyuanVideoTransformer3DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
+    @require_torch_gpu
+    @require_torch_2
+    @is_torch_compile
+    @slow
+    def test_torch_compile_recompilation_and_graph_break(self):
+        torch._dynamo.reset()
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
 
-class HunyuanSkyreelsImageToVideoTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
+        model = self.model_class(**init_dict).to(torch_device)
+        model = torch.compile(model, fullgraph=True)
+
+        with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)
+
+
+class HunyuanSkyreelsImageToVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -160,8 +179,23 @@ def test_gradient_checkpointing_is_applied(self):
         expected_set = {"HunyuanVideoTransformer3DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
+    @require_torch_gpu
+    @require_torch_2
+    @is_torch_compile
+    @slow
+    def test_torch_compile_recompilation_and_graph_break(self):
+        torch._dynamo.reset()
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        model = torch.compile(model, fullgraph=True)
+
+        with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)
+
 
-class HunyuanVideoImageToVideoTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
+class HunyuanVideoImageToVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -226,10 +260,23 @@ def test_gradient_checkpointing_is_applied(self):
         expected_set = {"HunyuanVideoTransformer3DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
+    @require_torch_gpu
+    @require_torch_2
+    @is_torch_compile
+    @slow
+    def test_torch_compile_recompilation_and_graph_break(self):
+        torch._dynamo.reset()
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
 
-class HunyuanVideoTokenReplaceImageToVideoTransformer3DTests(
-    ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase
-):
+        model = self.model_class(**init_dict).to(torch_device)
+        model = torch.compile(model, fullgraph=True)
+
+        with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)
+
+
+class HunyuanVideoTokenReplaceImageToVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -295,3 +342,18 @@ def test_output(self):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"HunyuanVideoTransformer3DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+    @require_torch_gpu
+    @require_torch_2
+    @is_torch_compile
+    @slow
+    def test_torch_compile_recompilation_and_graph_break(self):
+        torch._dynamo.reset()
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        model = torch.compile(model, fullgraph=True)
+
+        with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)

tests/models/transformers/test_models_transformer_wan.py

@@ -19,16 +19,20 @@
 from diffusers import WanTransformer3DModel
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
+    is_torch_compile,
+    require_torch_2,
+    require_torch_gpu,
+    slow,
     torch_device,
 )
 
-from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+from ..test_modeling_common import ModelTesterMixin
 
 
 enable_full_determinism()
 
 
-class WanTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
+class WanTransformer3DTests(ModelTesterMixin, unittest.TestCase):
     model_class = WanTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -82,3 +86,18 @@ def prepare_init_args_and_inputs_for_common(self):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"WanTransformer3DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+    @require_torch_gpu
+    @require_torch_2
+    @is_torch_compile
+    @slow
+    def test_torch_compile_recompilation_and_graph_break(self):
+        torch._dynamo.reset()
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        model = torch.compile(model, fullgraph=True)
+
+        with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)