Merge branch 'main' into ltxv-0.9.1-integration

tests/lora/test_lora_layers_cogvideox.py

@@ -15,8 +15,6 @@
 import sys
 import unittest
 
-import numpy as np
-import pytest
 import torch
 from transformers import AutoTokenizer, T5EncoderModel
 
@@ -29,16 +27,13 @@
 )
 from diffusers.utils.testing_utils import (
     floats_tensor,
-    is_torch_version,
     require_peft_backend,
-    skip_mps,
-    torch_device,
 )
 
 
 sys.path.append(".")
 
-from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set  # noqa: E402
+from utils import PeftLoraLoaderMixinTests  # noqa: E402
 
 
 @require_peft_backend
@@ -123,41 +118,6 @@ def get_dummy_inputs(self, with_generator=True):
 
         return noise, input_ids, pipeline_inputs
 
-    @skip_mps
-    @pytest.mark.xfail(
-        condition=torch.device(torch_device).type == "cpu" and is_torch_version(">=", "2.5"),
-        reason="Test currently fails on CPU and PyTorch 2.5.1 but not on PyTorch 2.4.1.",
-        strict=True,
-    )
-    def test_lora_fuse_nan(self):
-        for scheduler_cls in self.scheduler_classes:
-            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
-            pipe = self.pipeline_class(**components)
-            pipe = pipe.to(torch_device)
-            pipe.set_progress_bar_config(disable=None)
-            _, _, inputs = self.get_dummy_inputs(with_generator=False)
-
-            pipe.transformer.add_adapter(denoiser_lora_config, "adapter-1")
-
-            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
-
-            # corrupt one LoRA weight with `inf` values
-            with torch.no_grad():
-                pipe.transformer.transformer_blocks[0].attn1.to_q.lora_A["adapter-1"].weight += float("inf")
-
-            # with `safe_fusing=True` we should see an Error
-            with self.assertRaises(ValueError):
-                pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
-
-            # without we should not see an error, but every image will be black
-            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
-
-            out = pipe(
-                "test", num_inference_steps=2, max_sequence_length=inputs["max_sequence_length"], output_type="np"
-            )[0]
-
-            self.assertTrue(np.isnan(out).all())
-
     def test_simple_inference_with_text_lora_denoiser_fused_multi(self):
         super().test_simple_inference_with_text_lora_denoiser_fused_multi(expected_atol=9e-3)
 

tests/lora/test_lora_layers_hunyuanvideo.py

@@ -15,8 +15,6 @@
 import sys
 import unittest
 
-import numpy as np
-import pytest
 import torch
 from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
 
@@ -28,16 +26,14 @@
 )
 from diffusers.utils.testing_utils import (
     floats_tensor,
-    is_torch_version,
     require_peft_backend,
     skip_mps,
-    torch_device,
 )
 
 
 sys.path.append(".")
 
-from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set  # noqa: E402
+from utils import PeftLoraLoaderMixinTests  # noqa: E402
 
 
 @require_peft_backend
@@ -144,46 +140,6 @@ def get_dummy_inputs(self, with_generator=True):
 
         return noise, input_ids, pipeline_inputs
 
-    @pytest.mark.xfail(
-        condition=torch.device(torch_device).type == "cpu" and is_torch_version(">=", "2.5"),
-        reason="Test currently fails on CPU and PyTorch 2.5.1 but not on PyTorch 2.4.1.",
-        strict=True,
-    )
-    def test_lora_fuse_nan(self):
-        for scheduler_cls in self.scheduler_classes:
-            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
-            pipe = self.pipeline_class(**components)
-            pipe = pipe.to(torch_device)
-            pipe.set_progress_bar_config(disable=None)
-            _, _, inputs = self.get_dummy_inputs(with_generator=False)
-
-            pipe.transformer.add_adapter(denoiser_lora_config, "adapter-1")
-
-            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
-
-            # corrupt one LoRA weight with `inf` values
-            with torch.no_grad():
-                pipe.transformer.transformer_blocks[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
-
-            # with `safe_fusing=True` we should see an Error
-            with self.assertRaises(ValueError):
-                pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
-
-            # without we should not see an error, but every image will be black
-            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
-
-            out = pipe(
-                prompt=inputs["prompt"],
-                height=inputs["height"],
-                width=inputs["width"],
-                num_frames=inputs["num_frames"],
-                num_inference_steps=inputs["num_inference_steps"],
-                max_sequence_length=inputs["max_sequence_length"],
-                output_type="np",
-            )[0]
-
-            self.assertTrue(np.isnan(out).all())
-
     def test_simple_inference_with_text_lora_denoiser_fused_multi(self):
         super().test_simple_inference_with_text_lora_denoiser_fused_multi(expected_atol=9e-3)
 

tests/lora/test_lora_layers_mochi.py

@@ -15,24 +15,20 @@
 import sys
 import unittest
 
-import numpy as np
-import pytest
 import torch
 from transformers import AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLMochi, FlowMatchEulerDiscreteScheduler, MochiPipeline, MochiTransformer3DModel
 from diffusers.utils.testing_utils import (
     floats_tensor,
-    is_torch_version,
     require_peft_backend,
     skip_mps,
-    torch_device,
 )
 
 
 sys.path.append(".")
 
-from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set  # noqa: E402
+from utils import PeftLoraLoaderMixinTests  # noqa: E402
 
 
 @require_peft_backend
@@ -103,40 +99,6 @@ def get_dummy_inputs(self, with_generator=True):
 
         return noise, input_ids, pipeline_inputs
 
-    @pytest.mark.xfail(
-        condition=torch.device(torch_device).type == "cpu" and is_torch_version(">=", "2.5"),
-        reason="Test currently fails on CPU and PyTorch 2.5.1 but not on PyTorch 2.4.1.",
-        strict=True,
-    )
-    def test_lora_fuse_nan(self):
-        for scheduler_cls in self.scheduler_classes:
-            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
-            pipe = self.pipeline_class(**components)
-            pipe = pipe.to(torch_device)
-            pipe.set_progress_bar_config(disable=None)
-            _, _, inputs = self.get_dummy_inputs(with_generator=False)
-
-            pipe.transformer.add_adapter(denoiser_lora_config, "adapter-1")
-
-            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
-
-            # corrupt one LoRA weight with `inf` values
-            with torch.no_grad():
-                pipe.transformer.transformer_blocks[0].attn1.to_q.lora_A["adapter-1"].weight += float("inf")
-
-            # with `safe_fusing=True` we should see an Error
-            with self.assertRaises(ValueError):
-                pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
-
-            # without we should not see an error, but every image will be black
-            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
-
-            out = pipe(
-                "test", num_inference_steps=2, max_sequence_length=inputs["max_sequence_length"], output_type="np"
-            )[0]
-
-            self.assertTrue(np.isnan(out).all())
-
     def test_simple_inference_with_text_lora_denoiser_fused_multi(self):
         super().test_simple_inference_with_text_lora_denoiser_fused_multi(expected_atol=9e-3)
 

tests/lora/utils.py

@@ -1528,7 +1528,7 @@ def test_simple_inference_with_text_denoiser_multi_adapter_weighted(self):
     @pytest.mark.xfail(
         condition=torch.device(torch_device).type == "cpu" and is_torch_version(">=", "2.5"),
         reason="Test currently fails on CPU and PyTorch 2.5.1 but not on PyTorch 2.4.1.",
-        strict=True,
+        strict=False,
     )
     def test_lora_fuse_nan(self):
         for scheduler_cls in self.scheduler_classes:

tests/models/test_attention_processor.py

@@ -2,10 +2,12 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 
 from diffusers import DiffusionPipeline
 from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor
+from diffusers.utils.testing_utils import torch_device
 
 
 class AttnAddedKVProcessorTests(unittest.TestCase):
@@ -79,6 +81,15 @@ def test_only_cross_attention(self):
 
 
 class DeprecatedAttentionBlockTests(unittest.TestCase):
+    @pytest.fixture(scope="session")
+    def is_dist_enabled(pytestconfig):
+        return pytestconfig.getoption("dist") == "loadfile"
+
+    @pytest.mark.xfail(
+        condition=torch.device(torch_device).type == "cuda" and is_dist_enabled,
+        reason="Test currently fails on our GPU  CI because of `loadfile`. Note that it only fails when the tests are distributed from `pytest ... tests/models`. If the tests are run individually, even with `loadfile` it won't fail.",
+        strict=True,
+    )
     def test_conversion_when_using_device_map(self):
         pipe = DiffusionPipeline.from_pretrained(
             "hf-internal-testing/tiny-stable-diffusion-torch", safety_checker=None

tests/models/transformers/test_models_transformer_mochi.py

@@ -30,6 +30,8 @@ class MochiTransformerTests(ModelTesterMixin, unittest.TestCase):
     model_class = MochiTransformer3DModel
     main_input_name = "hidden_states"
     uses_custom_attn_processor = True
+    # Overriding it because of the transformer size.
+    model_split_percents = [0.7, 0.6, 0.6]
 
     @property
     def dummy_input(self):

tests/models/transformers/test_models_transformer_sana.py

@@ -14,6 +14,7 @@
 
 import unittest
 
+import pytest
 import torch
 
 from diffusers import SanaTransformer2DModel
@@ -80,3 +81,27 @@ def prepare_init_args_and_inputs_for_common(self):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"SanaTransformer2DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+    @pytest.mark.xfail(
+        condition=torch.device(torch_device).type == "cuda",
+        reason="Test currently fails.",
+        strict=True,
+    )
+    def test_cpu_offload(self):
+        return super().test_cpu_offload()
+
+    @pytest.mark.xfail(
+        condition=torch.device(torch_device).type == "cuda",
+        reason="Test currently fails.",
+        strict=True,
+    )
+    def test_disk_offload_with_safetensors(self):
+        return super().test_disk_offload_with_safetensors()
+
+    @pytest.mark.xfail(
+        condition=torch.device(torch_device).type == "cuda",
+        reason="Test currently fails.",
+        strict=True,
+    )
+    def test_disk_offload_without_safetensors(self):
+        return super().test_disk_offload_without_safetensors()