fix copies

.github/workflows/pr_tests.yml

@@ -114,7 +114,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
@@ -191,7 +191,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
 
     - name: Environment
       run: |
@@ -242,7 +242,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality,test]"
+        uv pip install -e ".[quality]"
         # TODO (sayakpaul, DN6): revisit `--no-deps`
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers

setup.py

@@ -101,7 +101,6 @@ _deps = [
     "datasets",
     "filelock",
     "flax>=0.4.1",
-    "ftfy",
     "hf-doc-builder>=0.3.0",
     "httpx<1.0.0",
     "huggingface-hub>=0.34.0,<2.0",
@@ -222,14 +221,12 @@ extras["docs"] = deps_list("hf-doc-builder")
 extras["training"] = deps_list("accelerate", "datasets", "protobuf", "tensorboard", "Jinja2", "peft", "timm")
 extras["test"] = deps_list(
     "compel",
-    "ftfy",
     "GitPython",
     "datasets",
     "Jinja2",
     "invisible-watermark",
     "librosa",
     "parameterized",
-    "protobuf",
     "pytest",
     "pytest-timeout",
     "pytest-xdist",
@@ -238,7 +235,6 @@ extras["test"] = deps_list(
     "sentencepiece",
     "scipy",
     "tiktoken",
-    "torchsde",
     "torchvision",
     "transformers",
     "phonemizer",

src/diffusers/dependency_versions_table.py

@@ -8,7 +8,6 @@ deps = {
     "datasets": "datasets",
     "filelock": "filelock",
     "flax": "flax>=0.4.1",
-    "ftfy": "ftfy",
     "hf-doc-builder": "hf-doc-builder>=0.3.0",
     "httpx": "httpx<1.0.0",
     "huggingface-hub": "huggingface-hub>=0.34.0,<2.0",

src/diffusers/models/attention_dispatch.py

@@ -1117,26 +1117,6 @@ def _sage_attention_backward_op(
     raise NotImplementedError("Backward pass is not implemented for Sage attention.")
 
 
-def _maybe_modify_attn_mask_npu(query: torch.Tensor, key: torch.Tensor, attn_mask: torch.Tensor | None = None):
-    # Skip Attention Mask if all values are 1, `None` mask can speedup the computation
-    if attn_mask is not None and torch.all(attn_mask != 0):
-        attn_mask = None
-
-    # Reshape Attention Mask: [batch_size, seq_len_k] -> [batch_size, 1, sqe_len_q, seq_len_k]
-    # https://www.hiascend.com/document/detail/zh/Pytorch/730/apiref/torchnpuCustomsapi/docs/context/torch_npu-npu_fusion_attention.md
-    if (
-        attn_mask is not None
-        and attn_mask.ndim == 2
-        and attn_mask.shape[0] == query.shape[0]
-        and attn_mask.shape[1] == key.shape[1]
-    ):
-        B, Sq, Skv = attn_mask.shape[0], query.shape[1], key.shape[1]
-        attn_mask = ~attn_mask.to(torch.bool)
-        attn_mask = attn_mask.unsqueeze(1).expand(B, Sq, Skv).unsqueeze(1).contiguous()
-
-    return attn_mask
-
-
 def _npu_attention_forward_op(
     ctx: torch.autograd.function.FunctionCtx,
     query: torch.Tensor,
@@ -1154,14 +1134,11 @@ def _npu_attention_forward_op(
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
 
-    attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
-
     out = npu_fusion_attention(
         query,
         key,
         value,
         query.size(2),  # num_heads
-        atten_mask=attn_mask,
         input_layout="BSND",
         pse=None,
         scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -2691,17 +2668,16 @@ def _native_npu_attention(
     return_lse: bool = False,
     _parallel_config: "ParallelConfig" | None = None,
 ) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for NPU attention")
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
     if _parallel_config is None:
-        attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
-
         out = npu_fusion_attention(
             query,
             key,
             value,
             query.size(2),  # num_heads
-            atten_mask=attn_mask,
             input_layout="BSND",
             pse=None,
             scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -2716,7 +2692,7 @@ def _native_npu_attention(
             query,
             key,
             value,
-            attn_mask,
+            None,
             dropout_p,
             None,
             scale,

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -164,11 +164,7 @@ def compute_text_seq_len_from_mask(
     position_ids = torch.arange(text_seq_len, device=encoder_hidden_states.device, dtype=torch.long)
     active_positions = torch.where(encoder_hidden_states_mask, position_ids, position_ids.new_zeros(()))
     has_active = encoder_hidden_states_mask.any(dim=1)
-    per_sample_len = torch.where(
-        has_active,
-        active_positions.max(dim=1).values + 1,
-        torch.as_tensor(text_seq_len, device=encoder_hidden_states.device),
-    )
+    per_sample_len = torch.where(has_active, active_positions.max(dim=1).values + 1, torch.as_tensor(text_seq_len))
     return text_seq_len, per_sample_len, encoder_hidden_states_mask
 
 

src/diffusers/pipelines/pipeline_utils.py

@@ -112,7 +112,7 @@ LIBRARIES = []
 for library in LOADABLE_CLASSES:
     LIBRARIES.append(library)
 
-SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"]
+SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device()]
 
 logger = logging.get_logger(__name__)
 
@@ -468,7 +468,8 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         pipeline_is_sequentially_offloaded = any(
             module_is_sequentially_offloaded(module) for _, module in self.components.items()
         )
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline."
@@ -1187,7 +1188,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         """
         self._maybe_raise_error_if_group_offload_active(raise_error=True)
 
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`."
@@ -1311,7 +1312,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
         self.remove_all_hooks()
 
-        is_pipeline_device_mapped = self._is_pipeline_device_mapped()
+        is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1
         if is_pipeline_device_mapped:
             raise ValueError(
                 "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`."
@@ -2227,21 +2228,6 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
                 return True
         return False
 
-    def _is_pipeline_device_mapped(self):
-        # We support passing `device_map="cuda"`, for example. This is helpful, in case
-        # users want to pass `device_map="cpu"` when initializing a pipeline. This explicit declaration is desirable
-        # in limited VRAM environments because quantized models often initialize directly on the accelerator.
-        device_map = self.hf_device_map
-        is_device_type_map = False
-        if isinstance(device_map, str):
-            try:
-                torch.device(device_map)
-                is_device_type_map = True
-            except RuntimeError:
-                pass
-
-        return not is_device_type_map and isinstance(device_map, dict) and len(device_map) > 1
-
 
 class StableDiffusionMixin:
     r"""

tests/models/testing_utils/compile.py

@@ -81,7 +81,7 @@ class TorchCompileTesterMixin:
             _ = model(**inputs_dict)
 
     @torch.no_grad()
-    def test_torch_compile_repeated_blocks(self, recompile_limit=1):
+    def test_torch_compile_repeated_blocks(self):
         if self.model_class._repeated_blocks is None:
             pytest.skip("Skipping test as the model class doesn't have `_repeated_blocks` set.")
 
@@ -92,6 +92,7 @@ class TorchCompileTesterMixin:
         model.eval()
         model.compile_repeated_blocks(fullgraph=True)
 
+        recompile_limit = 1
         if self.model_class.__name__ == "UNet2DConditionModel":
             recompile_limit = 2
 

tests/models/testing_utils/quantization.py

@@ -628,21 +628,6 @@ class BitsAndBytesTesterMixin(BitsAndBytesConfigMixin, QuantizationTesterMixin):
         """Test that quantized models can be used for training with adapters."""
         self._test_quantization_training(BitsAndBytesConfigMixin.BNB_CONFIGS["4bit_nf4"])
 
-    @pytest.mark.parametrize(
-        "config_name",
-        list(BitsAndBytesConfigMixin.BNB_CONFIGS.keys()),
-        ids=list(BitsAndBytesConfigMixin.BNB_CONFIGS.keys()),
-    )
-    def test_cpu_device_map(self, config_name):
-        config_kwargs = BitsAndBytesConfigMixin.BNB_CONFIGS[config_name]
-        model_quantized = self._create_quantized_model(config_kwargs, device_map="cpu")
-
-        assert hasattr(model_quantized, "hf_device_map"), "Model should have hf_device_map attribute"
-        assert model_quantized.hf_device_map is not None, "hf_device_map should not be None"
-        assert model_quantized.device == torch.device("cpu"), (
-            f"Model should be on CPU, but is on {model_quantized.device}"
-        )
-
 
 @is_quantization
 @is_quanto

tests/models/transformers/test_models_transformer_wan_vace.py

@@ -147,7 +147,22 @@ class TestWanVACETransformer3DCompile(WanVACETransformer3DTesterConfig, TorchCom
     def test_torch_compile_repeated_blocks(self):
         # WanVACE has two block types (WanTransformerBlock and WanVACETransformerBlock),
         # so we need recompile_limit=2 instead of the default 1.
-        super().test_torch_compile_repeated_blocks(recompile_limit=2)
+        import torch._dynamo
+        import torch._inductor.utils
+
+        init_dict = self.get_init_dict()
+        inputs_dict = self.get_dummy_inputs()
+
+        model = self.model_class(**init_dict).to(torch_device)
+        model.eval()
+        model.compile_repeated_blocks(fullgraph=True)
+
+        with (
+            torch._inductor.utils.fresh_inductor_cache(),
+            torch._dynamo.config.patch(recompile_limit=2),
+        ):
+            _ = model(**inputs_dict)
+            _ = model(**inputs_dict)
 
 
 class TestWanVACETransformer3DBitsAndBytes(WanVACETransformer3DTesterConfig, BitsAndBytesTesterMixin):

tests/pipelines/allegro/test_allegro.py

@@ -158,10 +158,6 @@ class AllegroPipelineFastTests(PipelineTesterMixin, PyramidAttentionBroadcastTes
     def test_save_load_optional_components(self):
         pass
 
-    @unittest.skip("Decoding without tiling is not yet implemented")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass
-
     def test_inference(self):
         device = "cpu"
 

tests/pipelines/kandinsky/test_kandinsky_combined.py

@@ -34,7 +34,9 @@ enable_full_determinism()
 
 class KandinskyPipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyCombinedPipeline
-    params = ["prompt"]
+    params = [
+        "prompt",
+    ]
     batch_params = ["prompt", "negative_prompt"]
     required_optional_params = [
         "generator",
@@ -146,10 +148,6 @@ class KandinskyPipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCase)
     def test_dict_tuple_outputs_equivalent(self):
         super().test_dict_tuple_outputs_equivalent(expected_max_difference=5e-4)
 
-    @unittest.skip("Test not supported.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass
-
 
 class KandinskyPipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyImg2ImgCombinedPipeline
@@ -266,10 +264,6 @@ class KandinskyPipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest.Te
     def test_save_load_optional_components(self):
         super().test_save_load_optional_components(expected_max_difference=5e-4)
 
-    @unittest.skip("Test not supported.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass
-
 
 class KandinskyPipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyInpaintCombinedPipeline
@@ -390,7 +384,3 @@ class KandinskyPipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest.Te
 
     def test_save_load_local(self):
         super().test_save_load_local(expected_max_difference=5e-3)
-
-    @unittest.skip("Test not supported.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass

tests/pipelines/kandinsky2_2/test_kandinsky_combined.py

@@ -36,7 +36,9 @@ enable_full_determinism()
 
 class KandinskyV22PipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyV22CombinedPipeline
-    params = ["prompt"]
+    params = [
+        "prompt",
+    ]
     batch_params = ["prompt", "negative_prompt"]
     required_optional_params = [
         "generator",
@@ -68,7 +70,12 @@ class KandinskyV22PipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCa
     def get_dummy_inputs(self, device, seed=0):
         prior_dummy = PriorDummies()
         inputs = prior_dummy.get_dummy_inputs(device=device, seed=seed)
-        inputs.update({"height": 64, "width": 64})
+        inputs.update(
+            {
+                "height": 64,
+                "width": 64,
+            }
+        )
         return inputs
 
     def test_kandinsky(self):
@@ -148,18 +155,12 @@ class KandinskyV22PipelineCombinedFastTests(PipelineTesterMixin, unittest.TestCa
     def test_save_load_optional_components(self):
         super().test_save_load_optional_components(expected_max_difference=5e-3)
 
-    @unittest.skip("Test not supported.")
     def test_callback_inputs(self):
         pass
 
-    @unittest.skip("Test not supported.")
     def test_callback_cfg(self):
         pass
 
-    @unittest.skip("Test not supported.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass
-
 
 class KandinskyV22PipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyV22Img2ImgCombinedPipeline
@@ -278,18 +279,12 @@ class KandinskyV22PipelineImg2ImgCombinedFastTests(PipelineTesterMixin, unittest
     def save_load_local(self):
         super().test_save_load_local(expected_max_difference=5e-3)
 
-    @unittest.skip("Test not supported.")
     def test_callback_inputs(self):
         pass
 
-    @unittest.skip("Test not supported.")
     def test_callback_cfg(self):
         pass
 
-    @unittest.skip("Test not supported.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass
-
 
 class KandinskyV22PipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = KandinskyV22InpaintCombinedPipeline
@@ -416,7 +411,3 @@ class KandinskyV22PipelineInpaintCombinedFastTests(PipelineTesterMixin, unittest
 
     def test_callback_cfg(self):
         pass
-
-    @unittest.skip("`device_map` is not yet supported for connected pipelines.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass

tests/pipelines/kandinsky2_2/test_kandinsky_inpaint.py

@@ -296,9 +296,6 @@ class KandinskyV22InpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCas
         output = pipe(**inputs)[0]
         assert output.abs().sum() == 0
 
-    def test_pipeline_with_accelerator_device_map(self):
-        super().test_pipeline_with_accelerator_device_map(expected_max_difference=5e-3)
-
 
 @slow
 @require_torch_accelerator

tests/pipelines/kandinsky3/test_kandinsky3_img2img.py

@@ -194,9 +194,6 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
     def test_save_load_dduf(self):
         super().test_save_load_dduf(atol=1e-3, rtol=1e-3)
 
-    def test_pipeline_with_accelerator_device_map(self):
-        super().test_pipeline_with_accelerator_device_map(expected_max_difference=5e-3)
-
 
 @slow
 @require_torch_accelerator

tests/pipelines/test_pipelines_common.py

@@ -2355,6 +2355,7 @@ class PipelineTesterMixin:
                     f"Component '{name}' has dtype {component.dtype} but expected {expected_dtype}",
                 )
 
+    @require_torch_accelerator
     def test_pipeline_with_accelerator_device_map(self, expected_max_difference=1e-4):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)

tests/pipelines/visualcloze/test_pipeline_visualcloze_combined.py

@@ -342,7 +342,3 @@ class VisualClozePipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         self.assertLess(
             max_diff, expected_max_diff, "The output of the fp16 pipeline changed after saving and loading."
         )
-
-    @unittest.skip("Test not supported.")
-    def test_pipeline_with_accelerator_device_map(self):
-        pass

tests/pipelines/visualcloze/test_pipeline_visualcloze_generation.py

@@ -310,7 +310,3 @@ class VisualClozeGenerationPipelineFastTests(unittest.TestCase, PipelineTesterMi
     @unittest.skip("Skipped due to missing layout_prompt. Needs further investigation.")
     def test_encode_prompt_works_in_isolation(self, extra_required_param_value_dict=None, atol=0.0001, rtol=0.0001):
         pass
-
-    @unittest.skip("Needs to be revisited later.")
-    def test_pipeline_with_accelerator_device_map(self, expected_max_difference=0.0001):
-        pass