Merge branch 'main' into sana

Author: linoytsaban

.github/workflows/benchmark.yml

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

.github/workflows/build_docker_images.yml

@@ -38,9 +38,16 @@ jobs:
           token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Build Changed Docker Images
+        env: 
+          CHANGED_FILES: ${{ steps.file_changes.outputs.all }}
         run: |
-          CHANGED_FILES="${{ steps.file_changes.outputs.all }}"
-          for FILE in $CHANGED_FILES; do
+          echo "$CHANGED_FILES"
+          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")
@@ -65,7 +72,7 @@ jobs:
         image-name:
           - diffusers-pytorch-cpu
           - diffusers-pytorch-cuda
-          - diffusers-pytorch-compile-cuda
+          - diffusers-pytorch-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-compile-cuda
+      image: diffusers/diffusers-pytorch-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-compile-cuda
+      image: diffusers/diffusers-pytorch-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-compile-cuda
+      image: diffusers/diffusers-pytorch-cuda
       options: --gpus 0 --shm-size "16gb" --ipc host
 
     steps:

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -146,21 +146,27 @@ def is_safetensors_compatible(filenames, passed_components=None, folder_names=No
         components[component].append(component_filename)
 
     # If there are no component folders check the main directory for safetensors files
+    filtered_filenames = set()
     if not components:
         if variant is not None:
             filtered_filenames = filter_with_regex(filenames, variant_file_re)
-        else:
+
+        # If no variant filenames exist check if non-variant files are available
+        if not filtered_filenames:
             filtered_filenames = filter_with_regex(filenames, non_variant_file_re)
         return any(".safetensors" in filename for filename in filtered_filenames)
 
     # iterate over all files of a component
     # check if safetensor files exist for that component
-    # if variant is provided check if the variant of the safetensors exists
     for component, component_filenames in components.items():
         matches = []
+        filtered_component_filenames = set()
+        # if variant is provided check if the variant of the safetensors exists
         if variant is not None:
             filtered_component_filenames = filter_with_regex(component_filenames, variant_file_re)
-        else:
+
+        # if variant safetensor files do not exist check for non-variants
+        if not filtered_component_filenames:
             filtered_component_filenames = filter_with_regex(component_filenames, non_variant_file_re)
         for component_filename in filtered_component_filenames:
             filename, extension = os.path.splitext(component_filename)

src/diffusers/pipelines/pipeline_loading_utils.py

@@ -1748,14 +1748,14 @@ class TorchCompileTesterMixin:
     def setUp(self):
         # clean up the VRAM before each test
         super().setUp()
-        torch._dynamo.reset()
+        torch.compiler.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._dynamo.reset()
+        torch.compiler.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
@@ -1764,13 +1764,17 @@ def tearDown(self):
     @is_torch_compile
     @slow
     def test_torch_compile_recompilation_and_graph_break(self):
-        torch._dynamo.reset()
+        torch.compiler.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():
+        with (
+            torch._inductor.utils.fresh_inductor_cache(),
+            torch._dynamo.config.patch(error_on_recompile=True),
+            torch.no_grad(),
+        ):
             _ = model(**inputs_dict)
             _ = model(**inputs_dict)
 
@@ -1798,7 +1802,7 @@ 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._dynamo.reset()
+        torch.compiler.reset()
         gc.collect()
         backend_empty_cache(torch_device)
 
@@ -1915,7 +1919,7 @@ 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):
+        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
             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
@@ -1925,7 +1929,7 @@ 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):
+        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
             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
@@ -1935,7 +1939,7 @@ 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):
+        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
             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/test_modeling_common.py

@@ -19,20 +19,16 @@
 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
+from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
 
 
 enable_full_determinism()
 
 
-class HunyuanVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
+class HunyuanVideoTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -96,23 +92,8 @@ 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 HunyuanSkyreelsImageToVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
+class HunyuanSkyreelsImageToVideoTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -179,23 +160,8 @@ 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, unittest.TestCase):
+class HunyuanVideoImageToVideoTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -260,23 +226,10 @@ 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 HunyuanVideoTokenReplaceImageToVideoTransformer3DTests(ModelTesterMixin, unittest.TestCase):
+class HunyuanVideoTokenReplaceImageToVideoTransformer3DTests(
+    ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase
+):
     model_class = HunyuanVideoTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -342,18 +295,3 @@ 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_hunyuan_video.py

@@ -19,20 +19,16 @@
 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
+from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
 
 
 enable_full_determinism()
 
 
-class WanTransformer3DTests(ModelTesterMixin, unittest.TestCase):
+class WanTransformer3DTests(ModelTesterMixin, TorchCompileTesterMixin, unittest.TestCase):
     model_class = WanTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
@@ -86,18 +82,3 @@ 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)