update

up

Co-authored-by: yiyi@huggingface.co <yiyi@ip-26-0-161-123.ec2.internal>

.github/workflows/pr_tests.yml

@@ -92,7 +92,6 @@ jobs:
             runner: aws-general-8-plus
             image: diffusers/diffusers-pytorch-cpu
             report: torch_example_cpu
-
     name: ${{ matrix.config.name }}
 
     runs-on:
@@ -115,8 +114,7 @@ jobs:
     - name: Install dependencies
       run: |
         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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
 
     - name: Environment
@@ -218,8 +216,6 @@ jobs:
 
   run_lora_tests:
     needs: [check_code_quality, check_repository_consistency]
-    strategy:
-      fail-fast: false
 
     name: LoRA tests with PEFT main
 
@@ -247,9 +243,8 @@ jobs:
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
-        #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
+        
     - name: Environment
       run: |
         python utils/print_env.py
@@ -275,6 +270,6 @@ jobs:
       if: ${{ always() }}
       uses: actions/upload-artifact@v6
       with:
-        name: pr_main_test_reports
+        name: pr_lora_test_reports
         path: reports
 

.github/workflows/pr_tests_gpu.yml

@@ -131,8 +131,7 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-          #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
       - name: Environment
         run: |
@@ -202,8 +201,7 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-        #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
     - name: Environment
       run: |
@@ -264,8 +262,7 @@ jobs:
         nvidia-smi
     - name: Install dependencies
       run: |
-        #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip install -e ".[quality,training]"
 
     - name: Environment

.github/workflows/push_tests.yml

@@ -76,8 +76,7 @@ jobs:
         run: |
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-          #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
       - name: Environment
         run: |
           python utils/print_env.py
@@ -129,8 +128,7 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
-        #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
     - name: Environment
       run: |
@@ -182,8 +180,7 @@ jobs:
     - name: Install dependencies
       run: |
         uv pip install -e ".[quality,training]"
-        #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 transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
     - name: Environment
       run: |
         python utils/print_env.py

docs/source/en/api/modular_diffusers/pipeline_blocks.md

@@ -14,4 +14,8 @@
 
 ## AutoPipelineBlocks
 
-[[autodoc]] diffusers.modular_pipelines.modular_pipeline.AutoPipelineBlocks
+[[autodoc]] diffusers.modular_pipelines.modular_pipeline.AutoPipelineBlocks
+
+## ConditionalPipelineBlocks
+
+[[autodoc]] diffusers.modular_pipelines.modular_pipeline.ConditionalPipelineBlocks

docs/source/en/api/pipelines/cosmos.md

@@ -46,6 +46,20 @@ output = pipe(
 output.save("output.png")
 ```
 
+## Cosmos2_5_TransferPipeline
+
+[[autodoc]] Cosmos2_5_TransferPipeline
+  - all
+  - __call__
+
+
+## Cosmos2_5_PredictBasePipeline
+
+[[autodoc]] Cosmos2_5_PredictBasePipeline
+  - all
+  - __call__
+
+
 ## CosmosTextToWorldPipeline
 
 [[autodoc]] CosmosTextToWorldPipeline
@@ -70,12 +84,6 @@ output.save("output.png")
   - all
   - __call__
 
-## Cosmos2_5_PredictBasePipeline
-
-[[autodoc]] Cosmos2_5_PredictBasePipeline
-  - all
-  - __call__
-
 ## CosmosPipelineOutput
 
 [[autodoc]] pipelines.cosmos.pipeline_output.CosmosPipelineOutput

docs/source/en/modular_diffusers/auto_pipeline_blocks.md

@@ -121,7 +121,7 @@ from diffusers.modular_pipelines import AutoPipelineBlocks
 
 class AutoImageBlocks(AutoPipelineBlocks):
     # List of sub-block classes to choose from
-    block_classes = [block_inpaint_cls, block_i2i_cls, block_t2i_cls]
+    block_classes = [InpaintBlock, ImageToImageBlock, TextToImageBlock]
     # Names for each block in the same order
     block_names = ["inpaint", "img2img", "text2img"]
     # Trigger inputs that determine which block to run
@@ -129,8 +129,8 @@ class AutoImageBlocks(AutoPipelineBlocks):
     # - "image" triggers img2img workflow (but only if mask is not provided)
     # - if none of above, runs the text2img workflow (default)
     block_trigger_inputs = ["mask", "image", None]
-    # Description is extremely important for AutoPipelineBlocks
 
+    @property
     def description(self):
         return (
             "Pipeline generates images given different types of conditions!\n"
@@ -141,7 +141,7 @@ class AutoImageBlocks(AutoPipelineBlocks):
         )
 ```
 
-It is **very** important to include a `description` to avoid any confusion over how to run a block and what inputs are required. While [`~modular_pipelines.AutoPipelineBlocks`] are convenient, it's conditional logic may be difficult to figure out if it isn't properly explained.
+It is **very** important to include a `description` to avoid any confusion over how to run a block and what inputs are required. While [`~modular_pipelines.AutoPipelineBlocks`] are convenient, its conditional logic may be difficult to figure out if it isn't properly explained.
 
 Create an instance of `AutoImageBlocks`.
 
@@ -152,5 +152,74 @@ auto_blocks = AutoImageBlocks()
 For more complex compositions, such as nested [`~modular_pipelines.AutoPipelineBlocks`] blocks when they're used as sub-blocks in larger pipelines, use the [`~modular_pipelines.SequentialPipelineBlocks.get_execution_blocks`] method to extract the a block that is actually run based on your input.
 
 ```py
-auto_blocks.get_execution_blocks("mask")
+auto_blocks.get_execution_blocks(mask=True)
+```
+
+## ConditionalPipelineBlocks
+
+[`~modular_pipelines.AutoPipelineBlocks`] is a special case of [`~modular_pipelines.ConditionalPipelineBlocks`]. While [`~modular_pipelines.AutoPipelineBlocks`] selects blocks based on whether a trigger input is provided or not, [`~modular_pipelines.ConditionalPipelineBlocks`] is able to select a block based on custom selection logic provided in the `select_block` method.
+
+Here is the same example written using [`~modular_pipelines.ConditionalPipelineBlocks`] directly:
+
+```py
+from diffusers.modular_pipelines import ConditionalPipelineBlocks
+
+class AutoImageBlocks(ConditionalPipelineBlocks):
+    block_classes = [InpaintBlock, ImageToImageBlock, TextToImageBlock]
+    block_names = ["inpaint", "img2img", "text2img"]
+    block_trigger_inputs = ["mask", "image"]
+    default_block_name = "text2img"
+
+    @property
+    def description(self):
+        return (
+            "Pipeline generates images given different types of conditions!\n"
+            + "This is an auto pipeline block that works for text2img, img2img and inpainting tasks.\n"
+            + " - inpaint workflow is run when `mask` is provided.\n"
+            + " - img2img workflow is run when `image` is provided (but only when `mask` is not provided).\n"
+            + " - text2img workflow is run when neither `image` nor `mask` is provided.\n"
+        )
+
+    def select_block(self, mask=None, image=None) -> str | None:
+        if mask is not None:
+            return "inpaint"
+        if image is not None:
+            return "img2img"
+        return None  # falls back to default_block_name ("text2img")
+```
+
+The inputs listed in `block_trigger_inputs` are passed as keyword arguments to `select_block()`. When `select_block` returns `None`, it falls back to `default_block_name`. If `default_block_name` is also `None`, the entire conditional block is skipped — this is useful for optional processing steps that should only run when specific inputs are provided.
+
+## Workflows
+
+Pipelines that contain conditional blocks ([`~modular_pipelines.AutoPipelineBlocks`] or [`~modular_pipelines.ConditionalPipelineBlocks]`) can support multiple workflows — for example, our SDXL modular pipeline supports a dozen workflows all in one pipeline. But this also means it can be confusing for users to know what workflows are supported and how to run them. For pipeline builders, it's useful to be able to extract only the blocks relevant to a specific workflow.
+
+We recommend defining a `_workflow_map` to give each workflow a name and explicitly list the inputs it requires.
+
+```py
+from diffusers.modular_pipelines import SequentialPipelineBlocks
+
+class MyPipelineBlocks(SequentialPipelineBlocks):
+    block_classes = [TextEncoderBlock, AutoImageBlocks, DecodeBlock]
+    block_names = ["text_encoder", "auto_image", "decode"]
+
+    _workflow_map = {
+        "text2image": {"prompt": True},
+        "image2image": {"image": True, "prompt": True},
+        "inpaint": {"mask": True, "image": True, "prompt": True},
+    }
+```
+
+All of our built-in modular pipelines come with pre-defined workflows. The `available_workflows` property lists all supported workflows:
+
+```py
+pipeline_blocks = MyPipelineBlocks()
+pipeline_blocks.available_workflows
+# ['text2image', 'image2image', 'inpaint']
+```
+
+Retrieve a specific workflow with `get_workflow` to inspect and debug a specific block that executes the workflow.
+
+```py
+pipeline_blocks.get_workflow("inpaint")
 ```

docs/source/en/training/distributed_inference.md

@@ -111,7 +111,7 @@ if __name__ == "__main__":
 Call `torchrun` to run the inference script and use the `--nproc_per_node` argument to set the number of GPUs to use.
 
 ```bash
-torchrun run_distributed.py --nproc_per_node=2
+torchrun --nproc_per_node=2 run_distributed.py
 ```
 
 ## device_map

examples/custom_diffusion/test_custom_diffusion.py

@@ -17,6 +17,9 @@ import logging
 import os
 import sys
 import tempfile
+import unittest
+
+from diffusers.utils import is_transformers_version
 
 
 sys.path.append("..")
@@ -30,6 +33,7 @@ stream_handler = logging.StreamHandler(sys.stdout)
 logger.addHandler(stream_handler)
 
 
+@unittest.skipIf(is_transformers_version(">=", "4.57.5"), "Size mismatch")
 class CustomDiffusion(ExamplesTestsAccelerate):
     def test_custom_diffusion(self):
         with tempfile.TemporaryDirectory() as tmpdir:

scripts/convert_cosmos_to_diffusers.py

@@ -94,9 +94,15 @@ python scripts/convert_cosmos_to_diffusers.py \
     --transformer_type Cosmos-2.5-Transfer-General-2B \
     --transformer_ckpt_path $transformer_ckpt_path \
     --vae_type wan2.1 \
-    --output_path converted/transfer/2b/general/depth \
+    --output_path converted/transfer/2b/general/depth/pipeline \
     --save_pipeline
 
+python scripts/convert_cosmos_to_diffusers.py \
+    --transformer_type Cosmos-2.5-Transfer-General-2B \
+    --transformer_ckpt_path $transformer_ckpt_path \
+    --vae_type wan2.1 \
+    --output_path converted/transfer/2b/general/depth/models
+
 # edge
 transformer_ckpt_path=~/.cache/huggingface/hub/models--nvidia--Cosmos-Transfer2.5-2B/snapshots/eb5325b77d358944da58a690157dd2b8071bbf85/general/edge/61f5694b-0ad5-4ecd-8ad7-c8545627d125_ema_bf16.pt
 
@@ -120,9 +126,15 @@ python scripts/convert_cosmos_to_diffusers.py \
     --transformer_type Cosmos-2.5-Transfer-General-2B \
     --transformer_ckpt_path $transformer_ckpt_path \
     --vae_type wan2.1 \
-    --output_path converted/transfer/2b/general/blur \
+    --output_path converted/transfer/2b/general/blur/pipeline \
     --save_pipeline
 
+python scripts/convert_cosmos_to_diffusers.py \
+    --transformer_type Cosmos-2.5-Transfer-General-2B \
+    --transformer_ckpt_path $transformer_ckpt_path \
+    --vae_type wan2.1 \
+    --output_path converted/transfer/2b/general/blur/models
+
 # seg
 transformer_ckpt_path=~/.cache/huggingface/hub/models--nvidia--Cosmos-Transfer2.5-2B/snapshots/eb5325b77d358944da58a690157dd2b8071bbf85/general/seg/5136ef49-6d8d-42e8-8abf-7dac722a304a_ema_bf16.pt
 
@@ -130,8 +142,14 @@ python scripts/convert_cosmos_to_diffusers.py \
     --transformer_type Cosmos-2.5-Transfer-General-2B \
     --transformer_ckpt_path $transformer_ckpt_path \
     --vae_type wan2.1 \
-    --output_path converted/transfer/2b/general/seg \
+    --output_path converted/transfer/2b/general/seg/pipeline \
     --save_pipeline
+
+python scripts/convert_cosmos_to_diffusers.py \
+    --transformer_type Cosmos-2.5-Transfer-General-2B \
+    --transformer_ckpt_path $transformer_ckpt_path \
+    --vae_type wan2.1 \
+    --output_path converted/transfer/2b/general/seg/models
 ```
 """
 

src/diffusers/configuration_utils.py

@@ -648,6 +648,28 @@ class ConfigMixin:
             )
         return config_file
 
+    @classmethod
+    def _get_dataclass_from_config(cls, config_dict: dict[str, Any]):
+        sig = inspect.signature(cls.__init__)
+        fields = []
+        for name, param in sig.parameters.items():
+            if name == "self" or name == "kwargs" or name in cls.ignore_for_config:
+                continue
+            annotation = param.annotation if param.annotation is not inspect.Parameter.empty else Any
+            if param.default is not inspect.Parameter.empty:
+                fields.append((name, annotation, dataclasses.field(default=param.default)))
+            else:
+                fields.append((name, annotation))
+
+        dc_cls = dataclasses.make_dataclass(
+            f"{cls.__name__}Config",
+            fields,
+            frozen=True,
+        )
+        valid_fields = {f.name for f in dataclasses.fields(dc_cls)}
+        init_kwargs = {k: v for k, v in config_dict.items() if k in valid_fields}
+        return dc_cls(**init_kwargs)
+
 
 def register_to_config(init):
     r"""

src/diffusers/hooks/_common.py

@@ -48,6 +48,7 @@ _GO_LC_SUPPORTED_PYTORCH_LAYERS = (
     torch.nn.ConvTranspose2d,
     torch.nn.ConvTranspose3d,
     torch.nn.Linear,
+    torch.nn.Embedding,
     # TODO(aryan): look into torch.nn.LayerNorm, torch.nn.GroupNorm later, seems to be causing some issues with CogVideoX
     # because of double invocation of the same norm layer in CogVideoXLayerNorm
 )

src/diffusers/loaders/lora_conversion_utils.py

@@ -856,7 +856,7 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
                 )
             state_dict = {k: v for k, v in state_dict.items() if not k.startswith("text_encoders.t5xxl.transformer.")}
 
-        has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_")) for k in state_dict)
+        has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_", "lora_te1_")) for k in state_dict)
         if has_diffb:
             zero_status_diff_b = state_dict_all_zero(state_dict, ".diff_b")
             if zero_status_diff_b:
@@ -895,7 +895,7 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
         state_dict = {
             _custom_replace(k, limit_substrings): v
             for k, v in state_dict.items()
-            if k.startswith(("lora_unet_", "lora_te_"))
+            if k.startswith(("lora_unet_", "lora_te_", "lora_te1_"))
         }
 
         if any("text_projection" in k for k in state_dict):

src/diffusers/loaders/textual_inversion.py

@@ -22,7 +22,12 @@ from tokenizers import Tokenizer as TokenizerFast
 from torch import nn
 
 from ..models.modeling_utils import load_state_dict
-from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
+from ..utils import (
+    _get_model_file,
+    is_accelerate_available,
+    is_transformers_available,
+    logging,
+)
 
 
 if is_transformers_available():

src/diffusers/models/attention_dispatch.py

@@ -62,6 +62,8 @@ _REQUIRED_FLEX_VERSION = "2.5.0"
 _REQUIRED_XLA_VERSION = "2.2"
 _REQUIRED_XFORMERS_VERSION = "0.0.29"
 
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
 _CAN_USE_FLASH_ATTN = is_flash_attn_available() and is_flash_attn_version(">=", _REQUIRED_FLASH_VERSION)
 _CAN_USE_FLASH_ATTN_3 = is_flash_attn_3_available()
 _CAN_USE_AITER_ATTN = is_aiter_available() and is_aiter_version(">=", _REQUIRED_AITER_VERSION)
@@ -73,8 +75,18 @@ _CAN_USE_XFORMERS_ATTN = is_xformers_available() and is_xformers_version(">=", _
 
 
 if _CAN_USE_FLASH_ATTN:
-    from flash_attn import flash_attn_func, flash_attn_varlen_func
-    from flash_attn.flash_attn_interface import _wrapped_flash_attn_backward, _wrapped_flash_attn_forward
+    try:
+        from flash_attn import flash_attn_func, flash_attn_varlen_func
+        from flash_attn.flash_attn_interface import _wrapped_flash_attn_backward, _wrapped_flash_attn_forward
+    except (ImportError, OSError, RuntimeError) as e:
+        # Handle ABI mismatch or other import failures gracefully.
+        # This can happen when flash_attn was compiled against a different PyTorch version.
+        logger.warning(f"flash_attn is installed but failed to import: {e}. Falling back to native PyTorch attention.")
+        _CAN_USE_FLASH_ATTN = False
+        flash_attn_func = None
+        flash_attn_varlen_func = None
+        _wrapped_flash_attn_backward = None
+        _wrapped_flash_attn_forward = None
 else:
     flash_attn_func = None
     flash_attn_varlen_func = None
@@ -83,26 +95,47 @@ else:
 
 
 if _CAN_USE_FLASH_ATTN_3:
-    from flash_attn_interface import flash_attn_func as flash_attn_3_func
-    from flash_attn_interface import flash_attn_varlen_func as flash_attn_3_varlen_func
+    try:
+        from flash_attn_interface import flash_attn_func as flash_attn_3_func
+        from flash_attn_interface import flash_attn_varlen_func as flash_attn_3_varlen_func
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"flash_attn_3 failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_FLASH_ATTN_3 = False
+        flash_attn_3_func = None
+        flash_attn_3_varlen_func = None
 else:
     flash_attn_3_func = None
     flash_attn_3_varlen_func = None
 
 if _CAN_USE_AITER_ATTN:
-    from aiter import flash_attn_func as aiter_flash_attn_func
+    try:
+        from aiter import flash_attn_func as aiter_flash_attn_func
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"aiter failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_AITER_ATTN = False
+        aiter_flash_attn_func = None
 else:
     aiter_flash_attn_func = None
 
 if _CAN_USE_SAGE_ATTN:
-    from sageattention import (
-        sageattn,
-        sageattn_qk_int8_pv_fp8_cuda,
-        sageattn_qk_int8_pv_fp8_cuda_sm90,
-        sageattn_qk_int8_pv_fp16_cuda,
-        sageattn_qk_int8_pv_fp16_triton,
-        sageattn_varlen,
-    )
+    try:
+        from sageattention import (
+            sageattn,
+            sageattn_qk_int8_pv_fp8_cuda,
+            sageattn_qk_int8_pv_fp8_cuda_sm90,
+            sageattn_qk_int8_pv_fp16_cuda,
+            sageattn_qk_int8_pv_fp16_triton,
+            sageattn_varlen,
+        )
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"sageattention failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_SAGE_ATTN = False
+        sageattn = None
+        sageattn_qk_int8_pv_fp8_cuda = None
+        sageattn_qk_int8_pv_fp8_cuda_sm90 = None
+        sageattn_qk_int8_pv_fp16_cuda = None
+        sageattn_qk_int8_pv_fp16_triton = None
+        sageattn_varlen = None
 else:
     sageattn = None
     sageattn_qk_int8_pv_fp16_cuda = None
@@ -113,26 +146,48 @@ else:
 
 
 if _CAN_USE_FLEX_ATTN:
-    # We cannot import the flex_attention function from the package directly because it is expected (from the
-    # pytorch documentation) that the user may compile it. If we import directly, we will not have access to the
-    # compiled function.
-    import torch.nn.attention.flex_attention as flex_attention
+    try:
+        # We cannot import the flex_attention function from the package directly because it is expected (from the
+        # pytorch documentation) that the user may compile it. If we import directly, we will not have access to the
+        # compiled function.
+        import torch.nn.attention.flex_attention as flex_attention
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"flex_attention failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_FLEX_ATTN = False
+        flex_attention = None
+else:
+    flex_attention = None
 
 
 if _CAN_USE_NPU_ATTN:
-    from torch_npu import npu_fusion_attention
+    try:
+        from torch_npu import npu_fusion_attention
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"torch_npu failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_NPU_ATTN = False
+        npu_fusion_attention = None
 else:
     npu_fusion_attention = None
 
 
 if _CAN_USE_XLA_ATTN:
-    from torch_xla.experimental.custom_kernel import flash_attention as xla_flash_attention
+    try:
+        from torch_xla.experimental.custom_kernel import flash_attention as xla_flash_attention
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"torch_xla failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_XLA_ATTN = False
+        xla_flash_attention = None
 else:
     xla_flash_attention = None
 
 
 if _CAN_USE_XFORMERS_ATTN:
-    import xformers.ops as xops
+    try:
+        import xformers.ops as xops
+    except (ImportError, OSError, RuntimeError) as e:
+        logger.warning(f"xformers failed to import: {e}. Falling back to native attention.")
+        _CAN_USE_XFORMERS_ATTN = False
+        xops = None
 else:
     xops = None
 
@@ -158,8 +213,6 @@ else:
     _register_fake = register_fake_no_op
 
 
-logger = get_logger(__name__)  # pylint: disable=invalid-name
-
 # TODO(aryan): Add support for the following:
 # - Sage Attention++
 # - block sparse, radial and other attention methods
@@ -276,7 +329,11 @@ class _HubKernelConfig:
 _HUB_KERNELS_REGISTRY: dict["AttentionBackendName", _HubKernelConfig] = {
     # TODO: temporary revision for now. Remove when merged upstream into `main`.
     AttentionBackendName._FLASH_3_HUB: _HubKernelConfig(
-        repo_id="kernels-community/flash-attn3", function_attr="flash_attn_func", revision="fake-ops-return-probs"
+        repo_id="kernels-community/flash-attn3",
+        function_attr="flash_attn_func",
+        revision="fake-ops-return-probs",
+        wrapped_forward_attr="flash_attn_interface._flash_attn_forward",
+        wrapped_backward_attr="flash_attn_interface._flash_attn_backward",
     ),
     AttentionBackendName._FLASH_3_VARLEN_HUB: _HubKernelConfig(
         repo_id="kernels-community/flash-attn3",
@@ -676,7 +733,7 @@ def _wrapped_flash_attn_3(
 ) -> tuple[torch.Tensor, torch.Tensor]:
     # Hardcoded for now because pytorch does not support tuple/int type hints
     window_size = (-1, -1)
-    out, lse, *_ = flash_attn_3_func(
+    result = flash_attn_3_func(
         q=q,
         k=k,
         v=v,
@@ -693,7 +750,9 @@ def _wrapped_flash_attn_3(
         pack_gqa=pack_gqa,
         deterministic=deterministic,
         sm_margin=sm_margin,
+        return_attn_probs=True,
     )
+    out, lse, *_ = result
     lse = lse.permute(0, 2, 1)
     return out, lse
 
@@ -1237,36 +1296,62 @@ def _flash_attention_3_hub_forward_op(
     if enable_gqa:
         raise ValueError("`enable_gqa` is not yet supported for flash-attn 3 hub kernels.")
 
-    func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
-    out = func(
-        q=query,
-        k=key,
-        v=value,
-        softmax_scale=scale,
+    config = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB]
+    wrapped_forward_fn = config.wrapped_forward_fn
+    if wrapped_forward_fn is None:
+        raise RuntimeError(
+            "Flash attention 3 hub kernels must expose `flash_attn_interface._flash_attn_forward` "
+            "for context parallel execution."
+        )
+
+    if scale is None:
+        scale = query.shape[-1] ** (-0.5)
+
+    out, softmax_lse, *_ = wrapped_forward_fn(
+        query,
+        key,
+        value,
+        None,
+        None,  # k_new, v_new
+        None,  # qv
+        None,  # out
+        None,
+        None,
+        None,  # cu_seqlens_q/k/k_new
+        None,
+        None,  # seqused_q/k
+        None,
+        None,  # max_seqlen_q/k
+        None,
+        None,
+        None,  # page_table, kv_batch_idx, leftpad_k
+        None,
+        None,
+        None,  # rotary_cos/sin, seqlens_rotary
+        None,
+        None,
+        None,  # q_descale, k_descale, v_descale
+        scale,
         causal=is_causal,
-        qv=None,
-        q_descale=None,
-        k_descale=None,
-        v_descale=None,
-        window_size=window_size,
+        window_size_left=window_size[0],
+        window_size_right=window_size[1],
+        attention_chunk=0,
         softcap=softcap,
         num_splits=num_splits,
         pack_gqa=pack_gqa,
-        deterministic=deterministic,
         sm_margin=sm_margin,
-        return_attn_probs=return_lse,
     )
 
-    lse = None
-    if return_lse:
-        out, lse = out
-        lse = lse.permute(0, 2, 1).contiguous()
+    lse = softmax_lse.permute(0, 2, 1).contiguous() if return_lse else None
 
     if _save_ctx:
-        ctx.save_for_backward(query, key, value)
+        ctx.save_for_backward(query, key, value, out, softmax_lse)
         ctx.scale = scale
         ctx.is_causal = is_causal
-        ctx._hub_kernel = func
+        ctx.window_size = window_size
+        ctx.softcap = softcap
+        ctx.deterministic = deterministic
+        ctx.sm_margin = sm_margin
 
     return (out, lse) if return_lse else out
 
@@ -1275,54 +1360,49 @@ def _flash_attention_3_hub_backward_op(
     ctx: torch.autograd.function.FunctionCtx,
     grad_out: torch.Tensor,
     *args,
-    window_size: tuple[int, int] = (-1, -1),
-    softcap: float = 0.0,
-    num_splits: int = 1,
-    pack_gqa: bool | None = None,
-    deterministic: bool = False,
-    sm_margin: int = 0,
+    **kwargs,
 ):
-    query, key, value = ctx.saved_tensors
-    kernel_fn = ctx._hub_kernel
-    # NOTE: Unlike the FA2 hub kernel, the FA3 hub kernel does not expose separate wrapped forward/backward
-    # primitives (no `wrapped_forward_attr`/`wrapped_backward_attr` in its `_HubKernelConfig`). We
-    # therefore rerun the forward pass under `torch.enable_grad()` and differentiate through it with
-    # `torch.autograd.grad()`. This is a second forward pass during backward; it can be avoided once
-    # the FA3 hub exposes a dedicated fused backward kernel (analogous to `_wrapped_flash_attn_backward`
-    # in the FA2 hub), at which point this can be refactored to match `_flash_attention_hub_backward_op`.
-    with torch.enable_grad():
-        query_r = query.detach().requires_grad_(True)
-        key_r = key.detach().requires_grad_(True)
-        value_r = value.detach().requires_grad_(True)
-
-        out = kernel_fn(
-            q=query_r,
-            k=key_r,
-            v=value_r,
-            softmax_scale=ctx.scale,
-            causal=ctx.is_causal,
-            qv=None,
-            q_descale=None,
-            k_descale=None,
-            v_descale=None,
-            window_size=window_size,
-            softcap=softcap,
-            num_splits=num_splits,
-            pack_gqa=pack_gqa,
-            deterministic=deterministic,
-            sm_margin=sm_margin,
-            return_attn_probs=False,
+    config = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB]
+    wrapped_backward_fn = config.wrapped_backward_fn
+    if wrapped_backward_fn is None:
+        raise RuntimeError(
+            "Flash attention 3 hub kernels must expose `flash_attn_interface._flash_attn_backward` "
+            "for context parallel execution."
         )
-        if isinstance(out, tuple):
-            out = out[0]
 
-        grad_query, grad_key, grad_value = torch.autograd.grad(
-            out,
-            (query_r, key_r, value_r),
-            grad_out,
-            retain_graph=False,
-            allow_unused=False,
-        )
+    query, key, value, out, softmax_lse = ctx.saved_tensors
+    grad_query = torch.empty_like(query)
+    grad_key = torch.empty_like(key)
+    grad_value = torch.empty_like(value)
+
+    wrapped_backward_fn(
+        grad_out,
+        query,
+        key,
+        value,
+        out,
+        softmax_lse,
+        None,
+        None,  # cu_seqlens_q, cu_seqlens_k
+        None,
+        None,  # seqused_q, seqused_k
+        None,
+        None,  # max_seqlen_q, max_seqlen_k
+        grad_query,
+        grad_key,
+        grad_value,
+        ctx.scale,
+        ctx.is_causal,
+        ctx.window_size[0],
+        ctx.window_size[1],
+        ctx.softcap,
+        ctx.deterministic,
+        ctx.sm_margin,
+    )
+
+    grad_query = grad_query[..., : grad_out.shape[-1]]
+    grad_key = grad_key[..., : grad_out.shape[-1]]
+    grad_value = grad_value[..., : grad_out.shape[-1]]
 
     return grad_query, grad_key, grad_value
 
@@ -2623,7 +2703,7 @@ def _flash_varlen_attention_3(
     key_packed = torch.cat(key_valid, dim=0)
     value_packed = torch.cat(value_valid, dim=0)
 
-    out, lse, *_ = flash_attn_3_varlen_func(
+    result = flash_attn_3_varlen_func(
         q=query_packed,
         k=key_packed,
         v=value_packed,
@@ -2633,7 +2713,13 @@ def _flash_varlen_attention_3(
         max_seqlen_k=max_seqlen_k,
         softmax_scale=scale,
         causal=is_causal,
+        return_attn_probs=return_lse,
     )
+    if isinstance(result, tuple):
+        out, lse, *_ = result
+    else:
+        out = result
+        lse = None
     out = out.unflatten(0, (batch_size, -1))
 
     return (out, lse) if return_lse else out

src/diffusers/models/auto_model.py

@@ -30,10 +30,126 @@ class AutoModel(ConfigMixin):
     def __init__(self, *args, **kwargs):
         raise EnvironmentError(
             f"{self.__class__.__name__} is designed to be instantiated "
-            f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
+            f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)`, "
+            f"`{self.__class__.__name__}.from_config(config)`, or "
             f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
         )
 
+    @classmethod
+    def from_config(cls, pretrained_model_name_or_path_or_dict: str | os.PathLike | dict | None = None, **kwargs):
+        r"""
+        Instantiate a model from a config dictionary or a pretrained model configuration file with random weights (no
+        pretrained weights are loaded).
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str`, `os.PathLike`, or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model
+                      configuration hosted on the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing a model configuration
+                      file.
+                    - A config dictionary.
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model configuration, overriding the cached version if
+                it exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint.
+            local_files_only(`bool`, *optional*, defaults to `False`):
+                Whether to only load local model configuration files or not.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether to trust remote code.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+
+        Returns:
+            A model object instantiated from the config with random weights.
+
+        Example:
+
+        ```py
+        from diffusers import AutoModel
+
+        model = AutoModel.from_config("stable-diffusion-v1-5/stable-diffusion-v1-5", subfolder="unet")
+        ```
+        """
+        subfolder = kwargs.pop("subfolder", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+
+        hub_kwargs_names = [
+            "cache_dir",
+            "force_download",
+            "local_files_only",
+            "proxies",
+            "revision",
+            "token",
+        ]
+        hub_kwargs = {name: kwargs.pop(name, None) for name in hub_kwargs_names}
+
+        if pretrained_model_name_or_path_or_dict is None:
+            raise ValueError(
+                "Please provide a `pretrained_model_name_or_path_or_dict` as the first positional argument."
+            )
+
+        if isinstance(pretrained_model_name_or_path_or_dict, (str, os.PathLike)):
+            pretrained_model_name_or_path = pretrained_model_name_or_path_or_dict
+            config = cls.load_config(pretrained_model_name_or_path, subfolder=subfolder, **hub_kwargs)
+        else:
+            config = pretrained_model_name_or_path_or_dict
+            pretrained_model_name_or_path = config.get("_name_or_path", None)
+
+        has_remote_code = "auto_map" in config and cls.__name__ in config["auto_map"]
+        trust_remote_code = resolve_trust_remote_code(
+            trust_remote_code, pretrained_model_name_or_path, has_remote_code
+        )
+
+        if has_remote_code and trust_remote_code:
+            class_ref = config["auto_map"][cls.__name__]
+            module_file, class_name = class_ref.split(".")
+            module_file = module_file + ".py"
+            model_cls = get_class_from_dynamic_module(
+                pretrained_model_name_or_path,
+                subfolder=subfolder,
+                module_file=module_file,
+                class_name=class_name,
+                **hub_kwargs,
+            )
+        else:
+            if "_class_name" in config:
+                class_name = config["_class_name"]
+                library = "diffusers"
+            elif "model_type" in config:
+                class_name = "AutoModel"
+                library = "transformers"
+            else:
+                raise ValueError(
+                    f"Couldn't find a model class associated with the config: {config}. Make sure the config "
+                    "contains a `_class_name` or `model_type` key."
+                )
+
+            from ..pipelines.pipeline_loading_utils import ALL_IMPORTABLE_CLASSES, get_class_obj_and_candidates
+
+            model_cls, _ = get_class_obj_and_candidates(
+                library_name=library,
+                class_name=class_name,
+                importable_classes=ALL_IMPORTABLE_CLASSES,
+                pipelines=None,
+                is_pipeline_module=False,
+            )
+
+        if model_cls is None:
+            raise ValueError(f"AutoModel can't find a model linked to {class_name}.")
+
+        return model_cls.from_config(config, **kwargs)
+
     @classmethod
     @validate_hf_hub_args
     def from_pretrained(cls, pretrained_model_or_path: str | os.PathLike | None = None, **kwargs):

src/diffusers/models/controlnets/controlnet_cosmos.py

@@ -191,7 +191,12 @@ class CosmosControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
                 dim=1,
             )
 
-        control_hidden_states = torch.cat([control_hidden_states, torch.zeros_like(controls_latents[:, :1])], dim=1)
+        if condition_mask is not None:
+            control_hidden_states = torch.cat([control_hidden_states, condition_mask], dim=1)
+        else:
+            control_hidden_states = torch.cat(
+                [control_hidden_states, torch.zeros_like(controls_latents[:, :1])], dim=1
+            )
 
         padding_mask_resized = transforms.functional.resize(
             padding_mask, list(control_hidden_states.shape[-2:]), interpolation=transforms.InterpolationMode.NEAREST

src/diffusers/modular_pipelines/modular_pipeline.py

@@ -1633,7 +1633,14 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):
                 blocks_class_name = self.default_blocks_name
             if blocks_class_name is not None:
                 diffusers_module = importlib.import_module("diffusers")
-                blocks_class = getattr(diffusers_module, blocks_class_name)
+                blocks_class = getattr(diffusers_module, blocks_class_name, None)
+                # If the blocks_class is not found or is a base class (e.g. SequentialPipelineBlocks saved by from_blocks_dict) with empty block_classes
+                # fall back to default_blocks_name
+                if blocks_class is None or not blocks_class.block_classes:
+                    blocks_class_name = self.default_blocks_name
+                    blocks_class = getattr(diffusers_module, blocks_class_name)
+
+            if blocks_class is not None:
                 blocks = blocks_class()
             else:
                 logger.warning(f"`blocks` is `None`, no default blocks class found for {self.__class__.__name__}")

src/diffusers/pipelines/audioldm2/pipeline_audioldm2.py

@@ -502,6 +502,10 @@ class AudioLDM2Pipeline(DiffusionPipeline):
                         text_input_ids,
                         attention_mask=attention_mask,
                     )
+                    # Extract the pooler output if it's a BaseModelOutputWithPooling (Transformers v5+)
+                    # otherwise use it directly (Transformers v4)
+                    if hasattr(prompt_embeds, "pooler_output"):
+                        prompt_embeds = prompt_embeds.pooler_output
                     # append the seq-len dim: (bs, hidden_size) -> (bs, seq_len, hidden_size)
                     prompt_embeds = prompt_embeds[:, None, :]
                     # make sure that we attend to this single hidden-state
@@ -610,6 +614,10 @@ class AudioLDM2Pipeline(DiffusionPipeline):
                         uncond_input_ids,
                         attention_mask=negative_attention_mask,
                     )
+                    # Extract the pooler output if it's a BaseModelOutputWithPooling (Transformers v5+)
+                    # otherwise use it directly (Transformers v4)
+                    if hasattr(negative_prompt_embeds, "pooler_output"):
+                        negative_prompt_embeds = negative_prompt_embeds.pooler_output
                     # append the seq-len dim: (bs, hidden_size) -> (bs, seq_len, hidden_size)
                     negative_prompt_embeds = negative_prompt_embeds[:, None, :]
                     # make sure that we attend to this single hidden-state

src/diffusers/pipelines/cosmos/pipeline_cosmos2_5_predict.py

@@ -287,6 +287,9 @@ class Cosmos2_5_PredictBasePipeline(DiffusionPipeline):
                 truncation=True,
                 padding="max_length",
             )
+            input_ids = (
+                input_ids["input_ids"] if not isinstance(input_ids, list) and "input_ids" in input_ids else input_ids
+            )
             input_ids = torch.LongTensor(input_ids)
             input_ids_batch.append(input_ids)
 

src/diffusers/pipelines/cosmos/pipeline_cosmos2_5_transfer.py

@@ -17,9 +17,6 @@ from typing import Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
-import torchvision
-import torchvision.transforms
-import torchvision.transforms.functional
 from transformers import AutoTokenizer, Qwen2_5_VLForConditionalGeneration
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
@@ -54,11 +51,13 @@ else:
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-def _maybe_pad_video(video: torch.Tensor, num_frames: int):
+def _maybe_pad_or_trim_video(video: torch.Tensor, num_frames: int):
     n_pad_frames = num_frames - video.shape[2]
     if n_pad_frames > 0:
         last_frame = video[:, :, -1:, :, :]
         video = torch.cat((video, last_frame.repeat(1, 1, n_pad_frames, 1, 1)), dim=2)
+    elif num_frames < video.shape[2]:
+        video = video[:, :, :num_frames, :, :]
     return video
 
 
@@ -134,8 +133,8 @@ EXAMPLE_DOC_STRING = """
         >>> controls = [Image.fromarray(x.numpy()) for x in controls.permute(1, 2, 3, 0)]
         >>> export_to_video(controls, "edge_controlled_video_edge.mp4", fps=30)
 
+        >>> # Transfer inference with controls.
         >>> video = pipe(
-        ...     video=input_video[:num_frames],
         ...     controls=controls,
         ...     controls_conditioning_scale=1.0,
         ...     prompt=prompt,
@@ -149,7 +148,7 @@ EXAMPLE_DOC_STRING = """
 
 class Cosmos2_5_TransferPipeline(DiffusionPipeline):
     r"""
-    Pipeline for Cosmos Transfer2.5 base model.
+    Pipeline for Cosmos Transfer2.5, supporting auto-regressive inference.
 
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
@@ -166,12 +165,14 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
         vae ([`AutoencoderKLWan`]):
             Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        controlnet ([`CosmosControlNetModel`]):
+            ControlNet used to condition generation on control inputs.
     """
 
     model_cpu_offload_seq = "text_encoder->transformer->controlnet->vae"
     _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
     # We mark safety_checker as optional here to get around some test failures, but it is not really optional
-    _optional_components = ["safety_checker", "controlnet"]
+    _optional_components = ["safety_checker"]
     _exclude_from_cpu_offload = ["safety_checker"]
 
     def __init__(
@@ -181,8 +182,8 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         transformer: CosmosTransformer3DModel,
         vae: AutoencoderKLWan,
         scheduler: UniPCMultistepScheduler,
-        controlnet: Optional[CosmosControlNetModel],
-        safety_checker: CosmosSafetyChecker = None,
+        controlnet: CosmosControlNetModel,
+        safety_checker: Optional[CosmosSafetyChecker] = None,
     ):
         super().__init__()
 
@@ -262,6 +263,9 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
                 truncation=True,
                 padding="max_length",
             )
+            input_ids = (
+                input_ids["input_ids"] if not isinstance(input_ids, list) and "input_ids" in input_ids else input_ids
+            )
             input_ids = torch.LongTensor(input_ids)
             input_ids_batch.append(input_ids)
 
@@ -381,10 +385,11 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         num_frames_in: int = 93,
         num_frames_out: int = 93,
         do_classifier_free_guidance: bool = True,
-        dtype: torch.dtype | None = None,
-        device: torch.device | None = None,
-        generator: torch.Generator | list[torch.Generator] | None = None,
-        latents: torch.Tensor | None = None,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        num_cond_latent_frames: int = 0,
     ) -> torch.Tensor:
         if isinstance(generator, list) and len(generator) != batch_size:
             raise ValueError(
@@ -399,10 +404,14 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         W = width // self.vae_scale_factor_spatial
         shape = (B, C, T, H, W)
 
-        if num_frames_in == 0:
-            if latents is None:
-                latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        if latents is not None:
+            if latents.shape[1:] != shape[1:]:
+                raise ValueError(f"Unexpected `latents` shape, got {latents.shape}, expected {shape}.")
+            latents = latents.to(device=device, dtype=dtype)
+        else:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
 
+        if num_frames_in == 0:
             cond_mask = torch.zeros((B, 1, T, H, W), dtype=latents.dtype, device=latents.device)
             cond_indicator = torch.zeros((B, 1, T, 1, 1), dtype=latents.dtype, device=latents.device)
 
@@ -432,16 +441,12 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             latents_std = self.latents_std.to(device=device, dtype=dtype)
             cond_latents = (cond_latents - latents_mean) / latents_std
 
-            if latents is None:
-                latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
-            else:
-                latents = latents.to(device=device, dtype=dtype)
-
             padding_shape = (B, 1, T, H, W)
             ones_padding = latents.new_ones(padding_shape)
             zeros_padding = latents.new_zeros(padding_shape)
 
-            cond_indicator = latents.new_zeros(1, 1, latents.size(2), 1, 1)
+            cond_indicator = latents.new_zeros(B, 1, latents.size(2), 1, 1)
+            cond_indicator[:, :, 0:num_cond_latent_frames, :, :] = 1.0
             cond_mask = cond_indicator * ones_padding + (1 - cond_indicator) * zeros_padding
 
             return (
@@ -451,34 +456,7 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
                 cond_indicator,
             )
 
-    def _encode_controls(
-        self,
-        controls: Optional[torch.Tensor],
-        height: int,
-        width: int,
-        num_frames: int,
-        dtype: torch.dtype,
-        device: torch.device,
-        generator: torch.Generator | list[torch.Generator] | None,
-    ) -> Optional[torch.Tensor]:
-        if controls is None:
-            return None
-
-        control_video = self.video_processor.preprocess_video(controls, height, width)
-        control_video = _maybe_pad_video(control_video, num_frames)
-
-        control_video = control_video.to(device=device, dtype=self.vae.dtype)
-        control_latents = [
-            retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator=generator) for vid in control_video
-        ]
-        control_latents = torch.cat(control_latents, dim=0).to(dtype)
-
-        latents_mean = self.latents_mean.to(device=device, dtype=dtype)
-        latents_std = self.latents_std.to(device=device, dtype=dtype)
-        control_latents = (control_latents - latents_mean) / latents_std
-        return control_latents
-
-    # Copied from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.check_inputs
+    # Modified from diffusers.pipelines.cosmos.pipeline_cosmos_text2world.CosmosTextToWorldPipeline.check_inputs
     def check_inputs(
         self,
         prompt,
@@ -486,9 +464,25 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         width,
         prompt_embeds=None,
         callback_on_step_end_tensor_inputs=None,
+        num_ar_conditional_frames=None,
+        num_ar_latent_conditional_frames=None,
+        num_frames_per_chunk=None,
+        num_frames=None,
+        conditional_frame_timestep=0.1,
     ):
-        if height % 16 != 0 or width % 16 != 0:
-            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+        if width <= 0 or height <= 0 or height % 16 != 0 or width % 16 != 0:
+            raise ValueError(
+                f"`height` and `width` have to be divisible by 16 (& positive) but are {height} and {width}."
+            )
+
+        if num_frames is not None and num_frames <= 0:
+            raise ValueError(f"`num_frames` has to be a positive integer when provided but is {num_frames}.")
+
+        if conditional_frame_timestep < 0 or conditional_frame_timestep > 1:
+            raise ValueError(
+                "`conditional_frame_timestep` has to be a float in the [0, 1] interval but is "
+                f"{conditional_frame_timestep}."
+            )
 
         if callback_on_step_end_tensor_inputs is not None and not all(
             k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
@@ -509,6 +503,46 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
             raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
 
+        if num_ar_latent_conditional_frames is not None and num_ar_conditional_frames is not None:
+            raise ValueError(
+                "Provide only one of `num_ar_conditional_frames` or `num_ar_latent_conditional_frames`, not both."
+            )
+        if num_ar_latent_conditional_frames is None and num_ar_conditional_frames is None:
+            raise ValueError("Provide either `num_ar_conditional_frames` or `num_ar_latent_conditional_frames`.")
+        if num_ar_latent_conditional_frames is not None and num_ar_latent_conditional_frames < 0:
+            raise ValueError("`num_ar_latent_conditional_frames` must be >= 0.")
+        if num_ar_conditional_frames is not None and num_ar_conditional_frames < 0:
+            raise ValueError("`num_ar_conditional_frames` must be >= 0.")
+
+        if num_ar_latent_conditional_frames is not None:
+            num_ar_conditional_frames = max(
+                0, (num_ar_latent_conditional_frames - 1) * self.vae_scale_factor_temporal + 1
+            )
+
+        min_chunk_len = self.vae_scale_factor_temporal + 1
+        if num_frames_per_chunk < min_chunk_len:
+            logger.warning(f"{num_frames_per_chunk=} must be larger than {min_chunk_len=}, setting to min_chunk_len")
+            num_frames_per_chunk = min_chunk_len
+
+        max_frames_by_rope = None
+        if getattr(self.transformer.config, "max_size", None) is not None:
+            max_frames_by_rope = max(
+                size // patch
+                for size, patch in zip(self.transformer.config.max_size, self.transformer.config.patch_size)
+            )
+            if num_frames_per_chunk > max_frames_by_rope:
+                raise ValueError(
+                    f"{num_frames_per_chunk=} is too large for RoPE setting ({max_frames_by_rope=}). "
+                    "Please reduce `num_frames_per_chunk`."
+                )
+
+        if num_ar_conditional_frames >= num_frames_per_chunk:
+            raise ValueError(
+                f"{num_ar_conditional_frames=} must be smaller than {num_frames_per_chunk=} for chunked generation."
+            )
+
+        return num_frames_per_chunk
+
     @property
     def guidance_scale(self):
         return self._guidance_scale
@@ -533,23 +567,22 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
         self,
-        image: PipelineImageInput | None = None,
-        video: List[PipelineImageInput] | None = None,
+        controls: PipelineImageInput | List[PipelineImageInput],
+        controls_conditioning_scale: Union[float, List[float]] = 1.0,
         prompt: Union[str, List[str]] | None = None,
         negative_prompt: Union[str, List[str]] = DEFAULT_NEGATIVE_PROMPT,
         height: int = 704,
-        width: int | None = None,
-        num_frames: int = 93,
+        width: Optional[int] = None,
+        num_frames: Optional[int] = None,
+        num_frames_per_chunk: int = 93,
         num_inference_steps: int = 36,
         guidance_scale: float = 3.0,
-        num_videos_per_prompt: Optional[int] = 1,
-        generator: torch.Generator | list[torch.Generator] | None = None,
-        latents: torch.Tensor | None = None,
-        controls: Optional[PipelineImageInput | List[PipelineImageInput]] = None,
-        controls_conditioning_scale: float | list[float] = 1.0,
-        prompt_embeds: torch.Tensor | None = None,
-        negative_prompt_embeds: torch.Tensor | None = None,
-        output_type: str = "pil",
+        num_videos_per_prompt: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback_on_step_end: Optional[
             Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
@@ -557,24 +590,26 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
         max_sequence_length: int = 512,
         conditional_frame_timestep: float = 0.1,
+        num_ar_conditional_frames: Optional[int] = 1,
+        num_ar_latent_conditional_frames: Optional[int] = None,
     ):
         r"""
-        The call function to the pipeline for generation. Supports three modes:
+        `controls` drive the conditioning through ControlNet. Controls are assumed to be pre-processed, e.g. edge maps
+        are pre-computed.
 
-        - **Text2World**: `image=None`, `video=None`, `prompt` provided. Generates a world clip.
-        - **Image2World**: `image` provided, `video=None`, `prompt` provided. Conditions on a single frame.
-        - **Video2World**: `video` provided, `image=None`, `prompt` provided. Conditions on an input clip.
+        Setting `num_frames` will restrict the total number of frames output, if not provided or assigned to None
+        (default) then the number of output frames will match the input `controls`.
 
-        Set `num_frames=93` (default) to produce a world video, or `num_frames=1` to produce a single image frame (the
-        above in "*2Image mode").
-
-        Outputs follow `output_type` (e.g., `"pil"` returns a list of `num_frames` PIL images per prompt).
+        Auto-regressive inference is supported and thus a sliding window of `num_frames_per_chunk` frames are used per
+        denoising loop. In addition, when auto-regressive inference is performed, the previous
+        `num_ar_latent_conditional_frames` or `num_ar_conditional_frames` are used to condition the following denoising
+        inference loops.
 
         Args:
-            image (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, *optional*):
-                Optional single image for Image2World conditioning. Must be `None` when `video` is provided.
-            video (`List[PIL.Image.Image]`, `np.ndarray`, `torch.Tensor`, *optional*):
-                Optional input video for Video2World conditioning. Must be `None` when `image` is provided.
+            controls (`PipelineImageInput`, `List[PipelineImageInput]`):
+                Control image or video input used by the ControlNet.
+            controls_conditioning_scale (`float` or `List[float]`, *optional*, defaults to `1.0`):
+                The scale factor(s) for the ControlNet outputs. A single float is broadcast to all control blocks.
             prompt (`str` or `List[str]`, *optional*):
                 The prompt or prompts to guide generation. Required unless `prompt_embeds` is supplied.
             height (`int`, defaults to `704`):
@@ -582,9 +617,10 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             width (`int`, *optional*):
                 The width in pixels of the generated image. If not provided, this will be determined based on the
                 aspect ratio of the input and the provided height.
-            num_frames (`int`, defaults to `93`):
-                Number of output frames. Use `93` for world (video) generation; set to `1` to return a single frame.
-            num_inference_steps (`int`, defaults to `35`):
+            num_frames (`int`, *optional*):
+                Number of output frames. Defaults to `None` to output the same number of frames as the input
+                `controls`.
+            num_inference_steps (`int`, defaults to `36`):
                 The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                 expense of slower inference.
             guidance_scale (`float`, defaults to `3.0`):
@@ -598,13 +634,9 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
                 A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                 generation deterministic.
             latents (`torch.Tensor`, *optional*):
-                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
-                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
-                tensor is generated by sampling using the supplied random `generator`.
-            controls (`PipelineImageInput`, `List[PipelineImageInput]`, *optional*):
-                Control image or video input used by the ControlNet. If `None`, ControlNet is skipped.
-            controls_conditioning_scale (`float` or `List[float]`, *optional*, defaults to `1.0`):
-                The scale factor(s) for the ControlNet outputs. A single float is broadcast to all control blocks.
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs. Can be used to
+                tweak the same generation with different prompts. If not provided, a latents tensor is generated by
+                sampling using the supplied random `generator`.
             prompt_embeds (`torch.Tensor`, *optional*):
                 Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                 provided, text embeddings will be generated from `prompt` input argument.
@@ -627,7 +659,18 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             max_sequence_length (`int`, defaults to `512`):
                 The maximum number of tokens in the prompt. If the prompt exceeds this length, it will be truncated. If
                 the prompt is shorter than this length, it will be padded.
+            num_ar_conditional_frames (`int`, *optional*, defaults to `1`):
+                Number of frames to condition on subsequent inference loops in auto-regressive inference, i.e. for the
+                second chunk and onwards. Only used if `num_ar_latent_conditional_frames` is `None`.
 
+                This is only used when auto-regressive inference is performed, i.e. when the number of frames in
+                controls is > num_frames_per_chunk
+            num_ar_latent_conditional_frames (`int`, *optional*):
+                Number of latent frames to condition on subsequent inference loops in auto-regressive inference, i.e.
+                for the second chunk and onwards. Only used if `num_ar_conditional_frames` is `None`.
+
+                This is only used when auto-regressive inference is performed, i.e. when the number of frames in
+                controls is > num_frames_per_chunk
         Examples:
 
         Returns:
@@ -647,21 +690,40 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
 
         if width is None:
-            frame = image or video[0] if image or video else None
-            if frame is None and controls is not None:
-                frame = controls[0] if isinstance(controls, list) else controls
-                if isinstance(frame, (torch.Tensor, np.ndarray)) and len(frame.shape) == 4:
-                    frame = controls[0]
+            frame = controls[0] if isinstance(controls, list) else controls
+            if isinstance(frame, list):
+                frame = frame[0]
+            if isinstance(frame, (torch.Tensor, np.ndarray)):
+                if frame.ndim == 5:
+                    frame = frame[0, 0]
+                elif frame.ndim == 4:
+                    frame = frame[0]
 
-            if frame is None:
-                width = int((height + 16) * (1280 / 720))
-            elif isinstance(frame, PIL.Image.Image):
+            if isinstance(frame, PIL.Image.Image):
                 width = int((height + 16) * (frame.width / frame.height))
             else:
+                if frame.ndim != 3:
+                    raise ValueError("`controls` must contain 3D frames in CHW format.")
                 width = int((height + 16) * (frame.shape[2] / frame.shape[1]))  # NOTE: assuming C H W
 
-        # Check inputs. Raise error if not correct
-        self.check_inputs(prompt, height, width, prompt_embeds, callback_on_step_end_tensor_inputs)
+        num_frames_per_chunk = self.check_inputs(
+            prompt,
+            height,
+            width,
+            prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            num_ar_conditional_frames,
+            num_ar_latent_conditional_frames,
+            num_frames_per_chunk,
+            num_frames,
+            conditional_frame_timestep,
+        )
+
+        if num_ar_latent_conditional_frames is not None:
+            num_cond_latent_frames = num_ar_latent_conditional_frames
+            num_ar_conditional_frames = max(0, (num_cond_latent_frames - 1) * self.vae_scale_factor_temporal + 1)
+        else:
+            num_cond_latent_frames = max(0, (num_ar_conditional_frames - 1) // self.vae_scale_factor_temporal + 1)
 
         self._guidance_scale = guidance_scale
         self._current_timestep = None
@@ -706,102 +768,137 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
         vae_dtype = self.vae.dtype
         transformer_dtype = self.transformer.dtype
 
-        img_context = torch.zeros(
-            batch_size,
-            self.transformer.config.img_context_num_tokens,
-            self.transformer.config.img_context_dim_in,
-            device=prompt_embeds.device,
-            dtype=transformer_dtype,
-        )
-        encoder_hidden_states = (prompt_embeds, img_context)
-        neg_encoder_hidden_states = (negative_prompt_embeds, img_context)
-
-        num_frames_in = None
-        if image is not None:
-            if batch_size != 1:
-                raise ValueError(f"batch_size must be 1 for image input (given {batch_size})")
-
-            image = torchvision.transforms.functional.to_tensor(image).unsqueeze(0)
-            video = torch.cat([image, torch.zeros_like(image).repeat(num_frames - 1, 1, 1, 1)], dim=0)
-            video = video.unsqueeze(0)
-            num_frames_in = 1
-        elif video is None:
-            video = torch.zeros(batch_size, num_frames, 3, height, width, dtype=torch.uint8)
-            num_frames_in = 0
-        else:
-            num_frames_in = len(video)
-
-            if batch_size != 1:
-                raise ValueError(f"batch_size must be 1 for video input (given {batch_size})")
-
-        assert video is not None
-        video = self.video_processor.preprocess_video(video, height, width)
-
-        # pad with last frame (for video2world)
-        num_frames_out = num_frames
-        video = _maybe_pad_video(video, num_frames_out)
-        assert num_frames_in <= num_frames_out, f"expected ({num_frames_in=}) <= ({num_frames_out=})"
-
-        video = video.to(device=device, dtype=vae_dtype)
-
-        num_channels_latents = self.transformer.config.in_channels - 1
-        latents, cond_latent, cond_mask, cond_indicator = self.prepare_latents(
-            video=video,
-            batch_size=batch_size * num_videos_per_prompt,
-            num_channels_latents=num_channels_latents,
-            height=height,
-            width=width,
-            num_frames_in=num_frames_in,
-            num_frames_out=num_frames,
-            do_classifier_free_guidance=self.do_classifier_free_guidance,
-            dtype=torch.float32,
-            device=device,
-            generator=generator,
-            latents=latents,
-        )
-        cond_timestep = torch.ones_like(cond_indicator) * conditional_frame_timestep
-        cond_mask = cond_mask.to(transformer_dtype)
-
-        controls_latents = None
-        if controls is not None:
-            controls_latents = self._encode_controls(
-                controls,
-                height=height,
-                width=width,
-                num_frames=num_frames,
+        if getattr(self.transformer.config, "img_context_dim_in", None):
+            img_context = torch.zeros(
+                batch_size,
+                self.transformer.config.img_context_num_tokens,
+                self.transformer.config.img_context_dim_in,
+                device=prompt_embeds.device,
                 dtype=transformer_dtype,
-                device=device,
-                generator=generator,
             )
 
-        padding_mask = latents.new_zeros(1, 1, height, width, dtype=transformer_dtype)
+            if num_videos_per_prompt > 1:
+                img_context = img_context.repeat_interleave(num_videos_per_prompt, dim=0)
 
-        # Denoising loop
-        self.scheduler.set_timesteps(num_inference_steps, device=device)
-        timesteps = self.scheduler.timesteps
-        self._num_timesteps = len(timesteps)
-        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+            encoder_hidden_states = (prompt_embeds, img_context)
+            neg_encoder_hidden_states = (negative_prompt_embeds, img_context)
+        else:
+            encoder_hidden_states = prompt_embeds
+            neg_encoder_hidden_states = negative_prompt_embeds
 
-        gt_velocity = (latents - cond_latent) * cond_mask
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
-            for i, t in enumerate(timesteps):
-                if self.interrupt:
-                    continue
-
-                self._current_timestep = t.cpu().item()
-
-                # NOTE: assumes sigma(t) \in [0, 1]
-                sigma_t = (
-                    torch.tensor(self.scheduler.sigmas[i].item())
-                    .unsqueeze(0)
-                    .to(device=device, dtype=transformer_dtype)
+        control_video = self.video_processor.preprocess_video(controls, height, width)
+        if control_video.shape[0] != batch_size:
+            if control_video.shape[0] == 1:
+                control_video = control_video.repeat(batch_size, 1, 1, 1, 1)
+            else:
+                raise ValueError(
+                    f"Expected controls batch size {batch_size} to match prompt batch size, but got {control_video.shape[0]}."
                 )
 
-                in_latents = cond_mask * cond_latent + (1 - cond_mask) * latents
-                in_latents = in_latents.to(transformer_dtype)
-                in_timestep = cond_indicator * cond_timestep + (1 - cond_indicator) * sigma_t
-                control_blocks = None
-                if controls_latents is not None and self.controlnet is not None:
+        num_frames_out = control_video.shape[2]
+        if num_frames is not None:
+            num_frames_out = min(num_frames_out, num_frames)
+
+        control_video = _maybe_pad_or_trim_video(control_video, num_frames_out)
+
+        # chunk information
+        num_latent_frames_per_chunk = (num_frames_per_chunk - 1) // self.vae_scale_factor_temporal + 1
+        chunk_stride = num_frames_per_chunk - num_ar_conditional_frames
+        chunk_idxs = [
+            (start_idx, min(start_idx + num_frames_per_chunk, num_frames_out))
+            for start_idx in range(0, num_frames_out - num_ar_conditional_frames, chunk_stride)
+        ]
+
+        video_chunks = []
+        latents_mean = self.latents_mean.to(dtype=vae_dtype, device=device)
+        latents_std = self.latents_std.to(dtype=vae_dtype, device=device)
+
+        def decode_latents(latents):
+            latents = latents * latents_std + latents_mean
+            video = self.vae.decode(latents.to(dtype=self.vae.dtype, device=device), return_dict=False)[0]
+            return video
+
+        latents_arg = latents
+        initial_num_cond_latent_frames = 0
+        latent_chunks = []
+        num_chunks = len(chunk_idxs)
+        total_steps = num_inference_steps * num_chunks
+        with self.progress_bar(total=total_steps) as progress_bar:
+            for chunk_idx, (start_idx, end_idx) in enumerate(chunk_idxs):
+                if chunk_idx == 0:
+                    prev_output = torch.zeros((batch_size, num_frames_per_chunk, 3, height, width), dtype=vae_dtype)
+                    prev_output = self.video_processor.preprocess_video(prev_output, height, width)
+                else:
+                    prev_output = video_chunks[-1].clone()
+                    if num_ar_conditional_frames > 0:
+                        prev_output[:, :, :num_ar_conditional_frames] = prev_output[:, :, -num_ar_conditional_frames:]
+                        prev_output[:, :, num_ar_conditional_frames:] = -1  # -1 == 0 in processed video space
+                    else:
+                        prev_output.fill_(-1)
+
+                chunk_video = prev_output.to(device=device, dtype=vae_dtype)
+                chunk_video = _maybe_pad_or_trim_video(chunk_video, num_frames_per_chunk)
+                latents, cond_latent, cond_mask, cond_indicator = self.prepare_latents(
+                    video=chunk_video,
+                    batch_size=batch_size * num_videos_per_prompt,
+                    num_channels_latents=self.transformer.config.in_channels - 1,
+                    height=height,
+                    width=width,
+                    num_frames_in=chunk_video.shape[2],
+                    num_frames_out=num_frames_per_chunk,
+                    do_classifier_free_guidance=self.do_classifier_free_guidance,
+                    dtype=torch.float32,
+                    device=device,
+                    generator=generator,
+                    num_cond_latent_frames=initial_num_cond_latent_frames
+                    if chunk_idx == 0
+                    else num_cond_latent_frames,
+                    latents=latents_arg,
+                )
+                cond_mask = cond_mask.to(transformer_dtype)
+                cond_timestep = torch.ones_like(cond_indicator) * conditional_frame_timestep
+                padding_mask = latents.new_zeros(1, 1, height, width, dtype=transformer_dtype)
+
+                chunk_control_video = control_video[:, :, start_idx:end_idx, ...].to(
+                    device=device, dtype=self.vae.dtype
+                )
+                chunk_control_video = _maybe_pad_or_trim_video(chunk_control_video, num_frames_per_chunk)
+                if isinstance(generator, list):
+                    controls_latents = [
+                        retrieve_latents(self.vae.encode(chunk_control_video[i].unsqueeze(0)), generator=generator[i])
+                        for i in range(chunk_control_video.shape[0])
+                    ]
+                else:
+                    controls_latents = [
+                        retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator=generator)
+                        for vid in chunk_control_video
+                    ]
+                controls_latents = torch.cat(controls_latents, dim=0).to(transformer_dtype)
+
+                controls_latents = (controls_latents - latents_mean) / latents_std
+
+                # Denoising loop
+                self.scheduler.set_timesteps(num_inference_steps, device=device)
+                timesteps = self.scheduler.timesteps
+                self._num_timesteps = len(timesteps)
+
+                gt_velocity = (latents - cond_latent) * cond_mask
+                for i, t in enumerate(timesteps):
+                    if self.interrupt:
+                        continue
+
+                    self._current_timestep = t.cpu().item()
+
+                    # NOTE: assumes sigma(t) \in [0, 1]
+                    sigma_t = (
+                        torch.tensor(self.scheduler.sigmas[i].item())
+                        .unsqueeze(0)
+                        .to(device=device, dtype=transformer_dtype)
+                    )
+
+                    in_latents = cond_mask * cond_latent + (1 - cond_mask) * latents
+                    in_latents = in_latents.to(transformer_dtype)
+                    in_timestep = cond_indicator * cond_timestep + (1 - cond_indicator) * sigma_t
                     control_output = self.controlnet(
                         controls_latents=controls_latents,
                         latents=in_latents,
@@ -814,20 +911,18 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
                     )
                     control_blocks = control_output[0]
 
-                noise_pred = self.transformer(
-                    hidden_states=in_latents,
-                    timestep=in_timestep,
-                    encoder_hidden_states=encoder_hidden_states,
-                    block_controlnet_hidden_states=control_blocks,
-                    condition_mask=cond_mask,
-                    padding_mask=padding_mask,
-                    return_dict=False,
-                )[0]
-                noise_pred = gt_velocity + noise_pred * (1 - cond_mask)
+                    noise_pred = self.transformer(
+                        hidden_states=in_latents,
+                        timestep=in_timestep,
+                        encoder_hidden_states=encoder_hidden_states,
+                        block_controlnet_hidden_states=control_blocks,
+                        condition_mask=cond_mask,
+                        padding_mask=padding_mask,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = gt_velocity + noise_pred * (1 - cond_mask)
 
-                if self.do_classifier_free_guidance:
-                    control_blocks = None
-                    if controls_latents is not None and self.controlnet is not None:
+                    if self.do_classifier_free_guidance:
                         control_output = self.controlnet(
                             controls_latents=controls_latents,
                             latents=in_latents,
@@ -840,46 +935,50 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
                         )
                         control_blocks = control_output[0]
 
-                    noise_pred_neg = self.transformer(
-                        hidden_states=in_latents,
-                        timestep=in_timestep,
-                        encoder_hidden_states=neg_encoder_hidden_states,  # NOTE: negative prompt
-                        block_controlnet_hidden_states=control_blocks,
-                        condition_mask=cond_mask,
-                        padding_mask=padding_mask,
-                        return_dict=False,
-                    )[0]
-                    # NOTE: replace velocity (noise_pred_neg) with gt_velocity for conditioning inputs only
-                    noise_pred_neg = gt_velocity + noise_pred_neg * (1 - cond_mask)
-                    noise_pred = noise_pred + self.guidance_scale * (noise_pred - noise_pred_neg)
+                        noise_pred_neg = self.transformer(
+                            hidden_states=in_latents,
+                            timestep=in_timestep,
+                            encoder_hidden_states=neg_encoder_hidden_states,  # NOTE: negative prompt
+                            block_controlnet_hidden_states=control_blocks,
+                            condition_mask=cond_mask,
+                            padding_mask=padding_mask,
+                            return_dict=False,
+                        )[0]
+                        # NOTE: replace velocity (noise_pred_neg) with gt_velocity for conditioning inputs only
+                        noise_pred_neg = gt_velocity + noise_pred_neg * (1 - cond_mask)
+                        noise_pred = noise_pred + self.guidance_scale * (noise_pred - noise_pred_neg)
 
-                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+                    latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
 
-                if callback_on_step_end is not None:
-                    callback_kwargs = {}
-                    for k in callback_on_step_end_tensor_inputs:
-                        callback_kwargs[k] = locals()[k]
-                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+                    # call the callback, if provided
+                    if callback_on_step_end is not None:
+                        callback_kwargs = {}
+                        for k in callback_on_step_end_tensor_inputs:
+                            callback_kwargs[k] = locals()[k]
+                        callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
 
-                    latents = callback_outputs.pop("latents", latents)
-                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                        latents = callback_outputs.pop("latents", latents)
+                        prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                        negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
 
-                # call the callback, if provided
-                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
-                    progress_bar.update()
+                    if i == total_steps - 1 or ((i + 1) % self.scheduler.order == 0):
+                        progress_bar.update()
 
-                if XLA_AVAILABLE:
-                    xm.mark_step()
+                    if XLA_AVAILABLE:
+                        xm.mark_step()
+
+                video_chunks.append(decode_latents(latents).detach().cpu())
+                latent_chunks.append(latents.detach().cpu())
 
         self._current_timestep = None
 
         if not output_type == "latent":
-            latents_mean = self.latents_mean.to(latents.device, latents.dtype)
-            latents_std = self.latents_std.to(latents.device, latents.dtype)
-            latents = latents * latents_std + latents_mean
-            video = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
-            video = self._match_num_frames(video, num_frames)
+            video_chunks = [
+                chunk[:, :, num_ar_conditional_frames:, ...] if chunk_idx != 0 else chunk
+                for chunk_idx, chunk in enumerate(video_chunks)
+            ]
+            video = torch.cat(video_chunks, dim=2)
+            video = video[:, :, :num_frames_out, ...]
 
             assert self.safety_checker is not None
             self.safety_checker.to(device)
@@ -896,7 +995,13 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             video = torch.from_numpy(video).permute(0, 4, 1, 2, 3)
             video = self.video_processor.postprocess_video(video, output_type=output_type)
         else:
-            video = latents
+            latent_T = (num_frames_out - 1) // self.vae_scale_factor_temporal + 1
+            latent_chunks = [
+                chunk[:, :, num_cond_latent_frames:, ...] if chunk_idx != 0 else chunk
+                for chunk_idx, chunk in enumerate(latent_chunks)
+            ]
+            video = torch.cat(latent_chunks, dim=2)
+            video = video[:, :, :latent_T, ...]
 
         # Offload all models
         self.maybe_free_model_hooks()
@@ -905,19 +1010,3 @@ class Cosmos2_5_TransferPipeline(DiffusionPipeline):
             return (video,)
 
         return CosmosPipelineOutput(frames=video)
-
-    def _match_num_frames(self, video: torch.Tensor, target_num_frames: int) -> torch.Tensor:
-        if target_num_frames <= 0 or video.shape[2] == target_num_frames:
-            return video
-
-        frames_per_latent = max(self.vae_scale_factor_temporal, 1)
-        video = torch.repeat_interleave(video, repeats=frames_per_latent, dim=2)
-
-        current_frames = video.shape[2]
-        if current_frames < target_num_frames:
-            pad = video[:, :, -1:, :, :].repeat(1, 1, target_num_frames - current_frames, 1, 1)
-            video = torch.cat([video, pad], dim=2)
-        elif current_frames > target_num_frames:
-            video = video[:, :, :target_num_frames]
-
-        return video

src/diffusers/pipelines/kandinsky/text_encoder.py

@@ -20,6 +20,8 @@ class MultilingualCLIP(PreTrainedModel):
         self.LinearTransformation = torch.nn.Linear(
             in_features=config.transformerDimensions, out_features=config.numDims
         )
+        if hasattr(self, "post_init"):
+            self.post_init()
 
     def forward(self, input_ids, attention_mask):
         embs = self.transformer(input_ids=input_ids, attention_mask=attention_mask)[0]

src/diffusers/pipelines/kolors/text_encoder.py

@@ -781,6 +781,9 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
             self.prefix_encoder = PrefixEncoder(config)
             self.dropout = torch.nn.Dropout(0.1)
 
+        if hasattr(self, "post_init"):
+            self.post_init()
+
     def get_input_embeddings(self):
         return self.embedding.word_embeddings
 
@@ -810,7 +813,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        use_cache = use_cache if use_cache is not None else getattr(self.config, "use_cache", None)
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         batch_size, seq_length = input_ids.shape

src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py

@@ -699,9 +699,13 @@ class LTX2ImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraL
         mask_shape = (batch_size, 1, num_frames, height, width)
 
         if latents is not None:
-            conditioning_mask = latents.new_zeros(mask_shape)
-            conditioning_mask[:, :, 0] = 1.0
             if latents.ndim == 5:
+                # conditioning_mask needs to the same shape as latents in two stages generation.
+                batch_size, _, num_frames, height, width = latents.shape
+                mask_shape = (batch_size, 1, num_frames, height, width)
+                conditioning_mask = latents.new_zeros(mask_shape)
+                conditioning_mask[:, :, 0] = 1.0
+
                 latents = self._normalize_latents(
                     latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
                 )
@@ -710,6 +714,9 @@ class LTX2ImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraL
                 latents = self._pack_latents(
                     latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
                 )
+            else:
+                conditioning_mask = latents.new_zeros(mask_shape)
+                conditioning_mask[:, :, 0] = 1.0
             conditioning_mask = self._pack_latents(
                 conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
             ).squeeze(-1)

src/diffusers/pipelines/pipeline_utils.py

@@ -341,6 +341,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             save_method_accept_safe = "safe_serialization" in save_method_signature.parameters
             save_method_accept_variant = "variant" in save_method_signature.parameters
             save_method_accept_max_shard_size = "max_shard_size" in save_method_signature.parameters
+            save_method_accept_peft_format = "save_peft_format" in save_method_signature.parameters
 
             save_kwargs = {}
             if save_method_accept_safe:
@@ -350,6 +351,11 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             if save_method_accept_max_shard_size and max_shard_size is not None:
                 # max_shard_size is expected to not be None in ModelMixin
                 save_kwargs["max_shard_size"] = max_shard_size
+            if save_method_accept_peft_format:
+                # Set save_peft_format=False for transformers>=5.0.0 compatibility
+                # In transformers 5.0.0+, the default save_peft_format=True adds "base_model.model" prefix
+                # to adapter keys, but from_pretrained expects keys without this prefix
+                save_kwargs["save_peft_format"] = False
 
             save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs)
 

src/diffusers/pipelines/prx/__init__.py

@@ -24,14 +24,25 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_prx"] = ["PRXPipeline"]
 
-# Import T5GemmaEncoder for pipeline loading compatibility
+# Wrap T5GemmaEncoder to pass config.encoder (T5GemmaModuleConfig) instead of the
+# composite T5GemmaConfig, which lacks flat attributes expected by T5GemmaEncoder.__init__.
 try:
     if is_transformers_available():
         import transformers
-        from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder
+        from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder as _T5GemmaEncoder
+
+        class T5GemmaEncoder(_T5GemmaEncoder):
+            @classmethod
+            def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
+                if "config" not in kwargs:
+                    from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig
+
+                    config = T5GemmaConfig.from_pretrained(pretrained_model_name_or_path)
+                    if hasattr(config, "encoder"):
+                        kwargs["config"] = config.encoder
+                return super().from_pretrained(pretrained_model_name_or_path, *args, **kwargs)
 
         _additional_imports["T5GemmaEncoder"] = T5GemmaEncoder
-        # Patch transformers module directly for serialization
         if not hasattr(transformers, "T5GemmaEncoder"):
             transformers.T5GemmaEncoder = T5GemmaEncoder
 except ImportError:

src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2.py

@@ -17,7 +17,7 @@ from typing import Any, Callable
 
 import regex as re
 import torch
-from transformers import AutoTokenizer, UMT5EncoderModel
+from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...loaders import SkyReelsV2LoraLoaderMixin
@@ -132,7 +132,7 @@ class SkyReelsV2Pipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin):
     def __init__(
         self,
         tokenizer: AutoTokenizer,
-        text_encoder: UMT5EncoderModel,
+        text_encoder: T5EncoderModel | UMT5EncoderModel,
         transformer: SkyReelsV2Transformer3DModel,
         vae: AutoencoderKLWan,
         scheduler: UniPCMultistepScheduler,

src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing.py

@@ -19,7 +19,7 @@ from copy import deepcopy
 from typing import Any, Callable
 
 import torch
-from transformers import AutoTokenizer, UMT5EncoderModel
+from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...loaders import SkyReelsV2LoraLoaderMixin
@@ -153,7 +153,7 @@ class SkyReelsV2DiffusionForcingPipeline(DiffusionPipeline, SkyReelsV2LoraLoader
     def __init__(
         self,
         tokenizer: AutoTokenizer,
-        text_encoder: UMT5EncoderModel,
+        text_encoder: T5EncoderModel | UMT5EncoderModel,
         transformer: SkyReelsV2Transformer3DModel,
         vae: AutoencoderKLWan,
         scheduler: UniPCMultistepScheduler,

src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_i2v.py

@@ -20,7 +20,7 @@ from typing import Any, Callable
 
 import PIL
 import torch
-from transformers import AutoTokenizer, UMT5EncoderModel
+from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel
 
 from diffusers.image_processor import PipelineImageInput
 from diffusers.utils.torch_utils import randn_tensor
@@ -158,7 +158,7 @@ class SkyReelsV2DiffusionForcingImageToVideoPipeline(DiffusionPipeline, SkyReels
     def __init__(
         self,
         tokenizer: AutoTokenizer,
-        text_encoder: UMT5EncoderModel,
+        text_encoder: T5EncoderModel | UMT5EncoderModel,
         transformer: SkyReelsV2Transformer3DModel,
         vae: AutoencoderKLWan,
         scheduler: UniPCMultistepScheduler,

src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_v2v.py

@@ -21,7 +21,7 @@ from typing import Any, Callable
 
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, UMT5EncoderModel
+from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...loaders import SkyReelsV2LoraLoaderMixin
@@ -214,7 +214,7 @@ class SkyReelsV2DiffusionForcingVideoToVideoPipeline(DiffusionPipeline, SkyReels
     def __init__(
         self,
         tokenizer: AutoTokenizer,
-        text_encoder: UMT5EncoderModel,
+        text_encoder: T5EncoderModel | UMT5EncoderModel,
         transformer: SkyReelsV2Transformer3DModel,
         vae: AutoencoderKLWan,
         scheduler: UniPCMultistepScheduler,

src/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_i2v.py

@@ -18,7 +18,7 @@ from typing import Any, Callable
 import PIL
 import regex as re
 import torch
-from transformers import AutoTokenizer, CLIPProcessor, CLIPVisionModelWithProjection, UMT5EncoderModel
+from transformers import AutoTokenizer, CLIPProcessor, CLIPVisionModelWithProjection, T5EncoderModel, UMT5EncoderModel
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...image_processor import PipelineImageInput
@@ -157,7 +157,7 @@ class SkyReelsV2ImageToVideoPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixi
     def __init__(
         self,
         tokenizer: AutoTokenizer,
-        text_encoder: UMT5EncoderModel,
+        text_encoder: T5EncoderModel | UMT5EncoderModel,
         image_encoder: CLIPVisionModelWithProjection,
         image_processor: CLIPProcessor,
         transformer: SkyReelsV2Transformer3DModel,

src/diffusers/pipelines/transformers_loading_utils.py

@@ -112,6 +112,8 @@ def _load_transformers_model_from_dduf(
                 tensors = safetensors.torch.load(mmap)
                 # Update the state dictionary with tensors
                 state_dict.update(tensors)
+            # `from_pretrained` sets the model to eval mode by default, which is the
+            # correct behavior for inference. Do not call `model.train()` here.
             return cls.from_pretrained(
                 pretrained_model_name_or_path=None,
                 config=config,

src/diffusers/pipelines/z_image/pipeline_z_image.py

@@ -276,7 +276,7 @@ class ZImagePipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMix
 
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1
+        return self._guidance_scale > 0
 
     @property
     def joint_attention_kwargs(self):

tests/models/controlnets/test_models_controlnet_cosmos.py

@@ -131,6 +131,26 @@ class CosmosControlNetModelTests(ModelTesterMixin, unittest.TestCase):
         self.assertIsInstance(output[0], list)
         self.assertEqual(len(output[0]), init_dict["n_controlnet_blocks"])
 
+    def test_condition_mask_changes_output(self):
+        """Test that condition mask affects control outputs."""
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        inputs_no_mask = dict(inputs_dict)
+        inputs_no_mask["condition_mask"] = torch.zeros_like(inputs_dict["condition_mask"])
+
+        with torch.no_grad():
+            output_no_mask = model(**inputs_no_mask)
+            output_with_mask = model(**inputs_dict)
+
+        self.assertEqual(len(output_no_mask.control_block_samples), len(output_with_mask.control_block_samples))
+        for no_mask_tensor, with_mask_tensor in zip(
+            output_no_mask.control_block_samples, output_with_mask.control_block_samples
+        ):
+            self.assertFalse(torch.allclose(no_mask_tensor, with_mask_tensor))
+
     def test_conditioning_scale_single(self):
         """Test that a single conditioning scale is broadcast to all blocks."""
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

tests/models/test_models_auto.py

@@ -1,5 +1,5 @@
 import unittest
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 from transformers import CLIPTextModel, LongformerModel
 
@@ -20,7 +20,9 @@ class TestAutoModel(unittest.TestCase):
         side_effect=[EnvironmentError("File not found"), {"model_type": "clip_text_model"}],
     )
     def test_load_from_config_transformers_with_subfolder(self, mock_load_config):
-        model = AutoModel.from_pretrained("hf-internal-testing/tiny-stable-diffusion-torch", subfolder="text_encoder")
+        model = AutoModel.from_pretrained(
+            "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="text_encoder", use_safetensors=False
+        )
         assert isinstance(model, CLIPTextModel)
 
     def test_load_from_config_without_subfolder(self):
@@ -28,5 +30,73 @@ class TestAutoModel(unittest.TestCase):
         assert isinstance(model, LongformerModel)
 
     def test_load_from_model_index(self):
-        model = AutoModel.from_pretrained("hf-internal-testing/tiny-stable-diffusion-torch", subfolder="text_encoder")
+        model = AutoModel.from_pretrained(
+            "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="text_encoder", use_safetensors=False
+        )
         assert isinstance(model, CLIPTextModel)
+
+
+class TestAutoModelFromConfig(unittest.TestCase):
+    @patch(
+        "diffusers.pipelines.pipeline_loading_utils.get_class_obj_and_candidates",
+        return_value=(MagicMock(), None),
+    )
+    def test_from_config_with_dict_diffusers_class(self, mock_get_class):
+        config = {"_class_name": "UNet2DConditionModel", "sample_size": 64}
+        mock_model = MagicMock()
+        mock_get_class.return_value[0].from_config.return_value = mock_model
+
+        result = AutoModel.from_config(config)
+
+        mock_get_class.assert_called_once_with(
+            library_name="diffusers",
+            class_name="UNet2DConditionModel",
+            importable_classes=unittest.mock.ANY,
+            pipelines=None,
+            is_pipeline_module=False,
+        )
+        mock_get_class.return_value[0].from_config.assert_called_once_with(config)
+        assert result is mock_model
+
+    @patch(
+        "diffusers.pipelines.pipeline_loading_utils.get_class_obj_and_candidates",
+        return_value=(MagicMock(), None),
+    )
+    @patch("diffusers.models.AutoModel.load_config", return_value={"_class_name": "UNet2DConditionModel"})
+    def test_from_config_with_string_path(self, mock_load_config, mock_get_class):
+        mock_model = MagicMock()
+        mock_get_class.return_value[0].from_config.return_value = mock_model
+
+        result = AutoModel.from_config("hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet")
+
+        mock_load_config.assert_called_once()
+        assert result is mock_model
+
+    def test_from_config_raises_on_missing_class_info(self):
+        config = {"some_key": "some_value"}
+        with self.assertRaises(ValueError, msg="Couldn't find a model class"):
+            AutoModel.from_config(config)
+
+    @patch(
+        "diffusers.pipelines.pipeline_loading_utils.get_class_obj_and_candidates",
+        return_value=(MagicMock(), None),
+    )
+    def test_from_config_with_model_type_routes_to_transformers(self, mock_get_class):
+        config = {"model_type": "clip_text_model"}
+        mock_model = MagicMock()
+        mock_get_class.return_value[0].from_config.return_value = mock_model
+
+        result = AutoModel.from_config(config)
+
+        mock_get_class.assert_called_once_with(
+            library_name="transformers",
+            class_name="AutoModel",
+            importable_classes=unittest.mock.ANY,
+            pipelines=None,
+            is_pipeline_module=False,
+        )
+        assert result is mock_model
+
+    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)

tests/modular_pipelines/test_modular_pipelines_common.py

@@ -1,4 +1,6 @@
 import gc
+import json
+import os
 import tempfile
 from typing import Callable
 
@@ -349,6 +351,33 @@ class ModularPipelineTesterMixin:
 
         assert torch.abs(image_slices[0] - image_slices[1]).max() < 1e-3
 
+    def test_modular_index_consistency(self):
+        pipe = self.get_pipeline()
+        components_spec = pipe._component_specs
+        components = sorted(components_spec.keys())
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir)
+            index_file = os.path.join(tmpdir, "modular_model_index.json")
+            assert os.path.exists(index_file)
+
+            with open(index_file) as f:
+                index_contents = json.load(f)
+
+            compulsory_keys = {"_blocks_class_name", "_class_name", "_diffusers_version"}
+            for k in compulsory_keys:
+                assert k in index_contents
+
+            to_check_attrs = {"pretrained_model_name_or_path", "revision", "subfolder"}
+            for component in components:
+                spec = components_spec[component]
+                for attr in to_check_attrs:
+                    if getattr(spec, "pretrained_model_name_or_path", None) is not None:
+                        for attr in to_check_attrs:
+                            assert component in index_contents, f"{component} should be present in index but isn't."
+                            attr_value_from_index = index_contents[component][2][attr]
+                            assert getattr(spec, attr) == attr_value_from_index
+
     def test_workflow_map(self):
         blocks = self.pipeline_blocks_class()
         if blocks._workflow_map is None:
@@ -699,3 +728,27 @@ class TestLoadComponentsSkipBehavior:
 
         # Verify test_component was not loaded
         assert not hasattr(pipe, "test_component") or pipe.test_component is None
+
+
+class TestModularPipelineInitFallback:
+    """Test that ModularPipeline.__init__ falls back to default_blocks_name when
+    _blocks_class_name is a base class (e.g. SequentialPipelineBlocks saved by from_blocks_dict)."""
+
+    def test_init_fallback_when_blocks_class_name_is_base_class(self, tmp_path):
+        # 1. Load pipeline and get a workflow (returns a base SequentialPipelineBlocks)
+        pipe = ModularPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
+        t2i_blocks = pipe.blocks.get_workflow("text2image")
+        assert t2i_blocks.__class__.__name__ == "SequentialPipelineBlocks"
+
+        # 2. Use init_pipeline to create a new pipeline from the workflow blocks
+        t2i_pipe = t2i_blocks.init_pipeline("hf-internal-testing/tiny-stable-diffusion-xl-pipe")
+
+        # 3. Save and reload — the saved config will have _blocks_class_name="SequentialPipelineBlocks"
+        save_dir = str(tmp_path / "pipeline")
+        t2i_pipe.save_pretrained(save_dir)
+        loaded_pipe = ModularPipeline.from_pretrained(save_dir)
+
+        # 4. Verify it fell back to default_blocks_name and has correct blocks
+        assert loaded_pipe.__class__.__name__ == pipe.__class__.__name__
+        assert loaded_pipe._blocks.__class__.__name__ == pipe._blocks.__class__.__name__
+        assert len(loaded_pipe._blocks.sub_blocks) == len(pipe._blocks.sub_blocks)

tests/others/test_config.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import dataclasses
 import json
 import tempfile
 import unittest
@@ -305,3 +306,96 @@ class ConfigTester(unittest.TestCase):
         result = json.loads(json_string)
         assert result["test_file_1"] == config.config.test_file_1.as_posix()
         assert result["test_file_2"] == config.config.test_file_2.as_posix()
+
+
+class SampleObjectTyped(ConfigMixin):
+    config_name = "config.json"
+
+    @register_to_config
+    def __init__(
+        self,
+        a: int = 2,
+        b: int = 5,
+        c: str = "hello",
+    ):
+        pass
+
+
+class SampleObjectWithIgnore(ConfigMixin):
+    config_name = "config.json"
+    ignore_for_config = ["secret"]
+
+    @register_to_config
+    def __init__(
+        self,
+        a: int = 2,
+        secret: str = "hidden",
+    ):
+        pass
+
+
+class DataclassFromConfigTester(unittest.TestCase):
+    def test_get_dataclass_from_config_returns_frozen_dataclass(self):
+        obj = SampleObject()
+        tc = SampleObject._get_dataclass_from_config(dict(obj.config))
+        assert dataclasses.is_dataclass(tc)
+        with self.assertRaises(dataclasses.FrozenInstanceError):
+            tc.a = 99
+
+    def test_get_dataclass_from_config_class_name(self):
+        obj = SampleObject()
+        tc = SampleObject._get_dataclass_from_config(dict(obj.config))
+        assert type(tc).__name__ == "SampleObjectConfig"
+
+    def test_get_dataclass_from_config_values_match_config(self):
+        obj = SampleObject(a=10, b=20)
+        tc = SampleObject._get_dataclass_from_config(dict(obj.config))
+        assert tc.a == 10
+        assert tc.b == 20
+        assert tc.c == (2, 5)
+        assert tc.d == "for diffusion"
+        assert tc.e == [1, 3]
+
+    def test_get_dataclass_from_config_from_raw_dict(self):
+        tc = SampleObjectTyped._get_dataclass_from_config({"a": 7, "b": 3, "c": "world"})
+        assert tc.a == 7
+        assert tc.b == 3
+        assert tc.c == "world"
+
+    def test_get_dataclass_from_config_annotations(self):
+        tc = SampleObjectTyped._get_dataclass_from_config({"a": 1, "b": 2, "c": "hi"})
+        fields = {f.name: f.type for f in dataclasses.fields(tc)}
+        assert fields["a"] is int
+        assert fields["b"] is int
+        assert fields["c"] is str
+
+    def test_get_dataclass_from_config_asdict_roundtrip(self):
+        tc = SampleObjectTyped._get_dataclass_from_config({"a": 7, "b": 3, "c": "world"})
+        d = dataclasses.asdict(tc)
+        assert d == {"a": 7, "b": 3, "c": "world"}
+
+    def test_get_dataclass_from_config_ignores_extra_keys(self):
+        tc = SampleObjectTyped._get_dataclass_from_config(
+            {"a": 1, "b": 2, "c": "hi", "_class_name": "Foo", "extra": 99}
+        )
+        assert tc.a == 1
+        assert not hasattr(tc, "_class_name")
+        assert not hasattr(tc, "extra")
+
+    def test_get_dataclass_from_config_respects_ignore_for_config(self):
+        tc = SampleObjectWithIgnore._get_dataclass_from_config({"a": 5})
+        assert not hasattr(tc, "secret")
+        assert tc.a == 5
+
+    def test_get_dataclass_from_config_works_for_scheduler(self):
+        scheduler = DDIMScheduler()
+        tc = DDIMScheduler._get_dataclass_from_config(dict(scheduler.config))
+        assert dataclasses.is_dataclass(tc)
+        assert type(tc).__name__ == "DDIMSchedulerConfig"
+        assert tc.num_train_timesteps == scheduler.config.num_train_timesteps
+
+    def test_get_dataclass_from_config_different_values(self):
+        tc1 = SampleObjectTyped._get_dataclass_from_config({"a": 1, "b": 2, "c": "x"})
+        tc2 = SampleObjectTyped._get_dataclass_from_config({"a": 9, "b": 8, "c": "y"})
+        assert tc1.a == 1
+        assert tc2.a == 9

tests/pipelines/audioldm2/test_audioldm2.py

@@ -282,6 +282,8 @@ class AudioLDM2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         text_inputs = text_inputs["input_ids"].to(torch_device)
 
         clap_prompt_embeds = audioldm_pipe.text_encoder.get_text_features(text_inputs)
+        if hasattr(clap_prompt_embeds, "pooler_output"):
+            clap_prompt_embeds = clap_prompt_embeds.pooler_output
         clap_prompt_embeds = clap_prompt_embeds[:, None, :]
 
         text_inputs = audioldm_pipe.tokenizer_2(
@@ -341,6 +343,8 @@ class AudioLDM2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_inputs = text_inputs["input_ids"].to(torch_device)
 
             clap_prompt_embeds = audioldm_pipe.text_encoder.get_text_features(text_inputs)
+            if hasattr(clap_prompt_embeds, "pooler_output"):
+                clap_prompt_embeds = clap_prompt_embeds.pooler_output
             clap_prompt_embeds = clap_prompt_embeds[:, None, :]
 
             text_inputs = audioldm_pipe.tokenizer_2(

tests/pipelines/bria/test_pipeline_bria.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 import torch
 from huggingface_hub import hf_hub_download
-from transformers import T5EncoderModel, T5TokenizerFast
+from transformers import AutoConfig, T5EncoderModel, T5TokenizerFast
 
 from diffusers import (
     AutoencoderKL,
@@ -89,7 +89,8 @@ class BriaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         scheduler = FlowMatchEulerDiscreteScheduler()
 
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = T5TokenizerFast.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/chroma/test_pipeline_chroma.py

@@ -2,7 +2,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, ChromaPipeline, ChromaTransformer2DModel, FlowMatchEulerDiscreteScheduler
 
@@ -41,7 +41,8 @@ class ChromaPipelineFastTests(
         )
 
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/chroma/test_pipeline_chroma_img2img.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, ChromaImg2ImgPipeline, ChromaTransformer2DModel, FlowMatchEulerDiscreteScheduler
 
@@ -42,7 +42,8 @@ class ChromaImg2ImgPipelineFastTests(
         )
 
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/chronoedit/test_chronoedit.py

@@ -17,6 +17,7 @@ import unittest
 import torch
 from PIL import Image
 from transformers import (
+    AutoConfig,
     AutoTokenizer,
     CLIPImageProcessor,
     CLIPVisionConfig,
@@ -71,7 +72,8 @@ class ChronoEditPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         torch.manual_seed(0)
         # TODO: impl FlowDPMSolverMultistepScheduler
         scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         torch.manual_seed(0)

tests/pipelines/cogvideo/test_cogvideox.py

@@ -18,7 +18,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCogVideoX, CogVideoXPipeline, CogVideoXTransformer3DModel, DDIMScheduler
 
@@ -117,7 +117,8 @@ class CogVideoXPipelineFastTests(
 
         torch.manual_seed(0)
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -235,6 +236,9 @@ class CogVideoXPipelineFastTests(
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/cogvideo/test_cogvideox_fun_control.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCogVideoX, CogVideoXFunControlPipeline, CogVideoXTransformer3DModel, DDIMScheduler
 
@@ -104,7 +104,8 @@ class CogVideoXFunControlPipelineFastTests(PipelineTesterMixin, unittest.TestCas
 
         torch.manual_seed(0)
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -228,6 +229,9 @@ class CogVideoXFunControlPipelineFastTests(PipelineTesterMixin, unittest.TestCas
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/cogvideo/test_cogvideox_image2video.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCogVideoX, CogVideoXImageToVideoPipeline, CogVideoXTransformer3DModel, DDIMScheduler
 from diffusers.utils import load_image
@@ -113,7 +113,8 @@ class CogVideoXImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestC
 
         torch.manual_seed(0)
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -237,6 +238,9 @@ class CogVideoXImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestC
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/cogvideo/test_cogvideox_video2video.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel, CogVideoXVideoToVideoPipeline, DDIMScheduler
 
@@ -99,7 +99,8 @@ class CogVideoXVideoToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestC
 
         torch.manual_seed(0)
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/cogview3/test_cogview3plus.py

@@ -18,7 +18,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, CogVideoXDDIMScheduler, CogView3PlusPipeline, CogView3PlusTransformer2DModel
 
@@ -89,7 +89,8 @@ class CogView3PlusPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         torch.manual_seed(0)
         scheduler = CogVideoXDDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/cogview4/test_cogview4.py

@@ -108,7 +108,7 @@ class CogView4PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             generator = torch.Generator(device=device).manual_seed(seed)
         inputs = {
             "prompt": "dance monkey",
-            "negative_prompt": "",
+            "negative_prompt": "bad",
             "generator": generator,
             "num_inference_steps": 2,
             "guidance_scale": 6.0,

tests/pipelines/consisid/test_consisid.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCogVideoX, ConsisIDPipeline, ConsisIDTransformer3DModel, DDIMScheduler
 from diffusers.utils import load_image
@@ -122,7 +122,8 @@ class ConsisIDPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         torch.manual_seed(0)
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -248,6 +249,9 @@ class ConsisIDPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/controlnet_flux/test_controlnet_flux.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 import torch
 from huggingface_hub import hf_hub_download
-from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
+from transformers import AutoConfig, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
 
 from diffusers import (
     AutoencoderKL,
@@ -97,7 +97,8 @@ class FluxControlNetPipelineFastTests(unittest.TestCase, PipelineTesterMixin, Fl
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = T5TokenizerFast.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/controlnet_flux/test_controlnet_flux_img2img.py

@@ -2,7 +2,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -13,9 +13,7 @@ from diffusers import (
 )
 from diffusers.utils.torch_utils import randn_tensor
 
-from ...testing_utils import (
-    torch_device,
-)
+from ...testing_utils import torch_device
 from ..test_pipelines_common import PipelineTesterMixin, check_qkv_fused_layers_exist
 
 
@@ -70,7 +68,8 @@ class FluxControlNetImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMi
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/controlnet_flux/test_controlnet_flux_inpaint.py

@@ -3,15 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-
-# torch_device,  # {{ edit_1 }} Removed unused import
-from transformers import (
-    AutoTokenizer,
-    CLIPTextConfig,
-    CLIPTextModel,
-    CLIPTokenizer,
-    T5EncoderModel,
-)
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -22,11 +14,7 @@ from diffusers import (
 )
 from diffusers.utils.torch_utils import randn_tensor
 
-from ...testing_utils import (
-    enable_full_determinism,
-    floats_tensor,
-    torch_device,
-)
+from ...testing_utils import enable_full_determinism, floats_tensor, torch_device
 from ..test_pipelines_common import PipelineTesterMixin
 
 
@@ -85,7 +73,8 @@ class FluxControlNetInpaintPipelineTests(unittest.TestCase, PipelineTesterMixin)
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/controlnet_hunyuandit/test_controlnet_hunyuandit.py

@@ -18,7 +18,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, BertModel, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, BertModel, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -96,7 +96,10 @@ class HunyuanDiTControlNetPipelineFastTests(unittest.TestCase, PipelineTesterMix
         scheduler = DDPMScheduler()
         text_encoder = BertModel.from_pretrained("hf-internal-testing/tiny-random-BertModel")
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/controlnet_sd3/test_controlnet_inpaint_sd3.py

@@ -17,7 +17,14 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    CLIPTextConfig,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKL,
@@ -28,10 +35,7 @@ from diffusers import (
 from diffusers.models import SD3ControlNetModel
 from diffusers.utils.torch_utils import randn_tensor
 
-from ...testing_utils import (
-    enable_full_determinism,
-    torch_device,
-)
+from ...testing_utils import enable_full_determinism, torch_device
 from ..test_pipelines_common import PipelineTesterMixin
 
 
@@ -103,7 +107,8 @@ class StableDiffusion3ControlInpaintNetPipelineFastTests(unittest.TestCase, Pipe
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_3 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

tests/pipelines/controlnet_sd3/test_controlnet_sd3.py

@@ -18,7 +18,14 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    CLIPTextConfig,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKL,
@@ -117,7 +124,8 @@ class StableDiffusion3ControlNetPipelineFastTests(unittest.TestCase, PipelineTes
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_3 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

tests/pipelines/cosmos/test_cosmos.py

@@ -20,7 +20,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCosmos, CosmosTextToWorldPipeline, CosmosTransformer3DModel, EDMEulerScheduler
 
@@ -107,7 +107,8 @@ class CosmosTextToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
             rho=7.0,
             final_sigmas_type="sigma_min",
         )
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -232,6 +233,9 @@ class CosmosTextToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/cosmos/test_cosmos2_5_transfer.py

@@ -55,7 +55,7 @@ class Cosmos2_5_TransferWrapper(Cosmos2_5_TransferPipeline):
 class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = Cosmos2_5_TransferWrapper
     params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
-    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS.union({"controls"})
     image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
     image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
     required_optional_params = frozenset(
@@ -176,15 +176,19 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
         else:
             generator = torch.Generator(device=device).manual_seed(seed)
 
+        controls_generator = torch.Generator(device="cpu").manual_seed(seed)
+
         inputs = {
             "prompt": "dance monkey",
             "negative_prompt": "bad quality",
+            "controls": [torch.randn(3, 32, 32, generator=controls_generator) for _ in range(5)],
             "generator": generator,
             "num_inference_steps": 2,
             "guidance_scale": 3.0,
             "height": 32,
             "width": 32,
             "num_frames": 3,
+            "num_frames_per_chunk": 16,
             "max_sequence_length": 16,
             "output_type": "pt",
         }
@@ -212,6 +216,56 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
         self.assertEqual(generated_video.shape, (3, 3, 32, 32))
         self.assertTrue(torch.isfinite(generated_video).all())
 
+    def test_inference_autoregressive_multi_chunk(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["num_frames"] = 5
+        inputs["num_frames_per_chunk"] = 3
+        inputs["num_ar_conditional_frames"] = 1
+
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+        self.assertEqual(generated_video.shape, (5, 3, 32, 32))
+        self.assertTrue(torch.isfinite(generated_video).all())
+
+    def test_inference_autoregressive_multi_chunk_no_condition_frames(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["num_frames"] = 5
+        inputs["num_frames_per_chunk"] = 3
+        inputs["num_ar_conditional_frames"] = 0
+
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+        self.assertEqual(generated_video.shape, (5, 3, 32, 32))
+        self.assertTrue(torch.isfinite(generated_video).all())
+
+    def test_num_frames_per_chunk_above_rope_raises(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["num_frames_per_chunk"] = 17
+
+        with self.assertRaisesRegex(ValueError, "too large for RoPE setting"):
+            pipe(**inputs)
+
     def test_inference_with_controls(self):
         """Test inference with control inputs (ControlNet)."""
         device = "cpu"
@@ -222,13 +276,13 @@ class Cosmos2_5_TransferPipelineFastTests(PipelineTesterMixin, unittest.TestCase
         pipe.set_progress_bar_config(disable=None)
 
         inputs = self.get_dummy_inputs(device)
-        # Add control video input - should be a video tensor
-        inputs["controls"] = [torch.randn(3, 3, 32, 32)]  # num_frames, channels, height, width
+        inputs["controls"] = [torch.randn(3, 32, 32) for _ in range(5)]  # list of 5 frames (C, H, W)
         inputs["controls_conditioning_scale"] = 1.0
+        inputs["num_frames"] = None
 
         video = pipe(**inputs).frames
         generated_video = video[0]
-        self.assertEqual(generated_video.shape, (3, 3, 32, 32))
+        self.assertEqual(generated_video.shape, (5, 3, 32, 32))
         self.assertTrue(torch.isfinite(generated_video).all())
 
     def test_callback_inputs(self):

tests/pipelines/cosmos/test_cosmos2_text2image.py

@@ -20,7 +20,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKLWan,
@@ -95,7 +95,8 @@ class Cosmos2TextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler(use_karras_sigmas=True)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/cosmos/test_cosmos2_video2world.py

@@ -21,7 +21,7 @@ import unittest
 import numpy as np
 import PIL.Image
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKLWan,
@@ -96,7 +96,8 @@ class Cosmos2VideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCas
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler(use_karras_sigmas=True)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/cosmos/test_cosmos_video2world.py

@@ -21,7 +21,7 @@ import unittest
 import numpy as np
 import PIL.Image
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLCosmos, CosmosTransformer3DModel, CosmosVideoToWorldPipeline, EDMEulerScheduler
 
@@ -108,7 +108,8 @@ class CosmosVideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase
             rho=7.0,
             final_sigmas_type="sigma_min",
         )
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -245,6 +246,9 @@ class CosmosVideoToWorldPipelineFastTests(PipelineTesterMixin, unittest.TestCase
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/deepfloyd_if/__init__.py

@@ -2,7 +2,7 @@ import tempfile
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import DDPMScheduler, UNet2DConditionModel
 from diffusers.models.attention_processor import AttnAddedKVProcessor
@@ -18,7 +18,8 @@ from ..test_pipelines_common import to_np
 class IFPipelineTesterMixin:
     def _get_dummy_components(self):
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         torch.manual_seed(0)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
@@ -75,7 +76,8 @@ class IFPipelineTesterMixin:
 
     def _get_superresolution_dummy_components(self):
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         torch.manual_seed(0)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
@@ -250,6 +252,9 @@ class IFPipelineTesterMixin:
     # This should be handled in the base test and then this method can be removed.
     def _test_save_load_local(self):
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)

tests/pipelines/deepfloyd_if/test_if.py

@@ -18,9 +18,7 @@ import unittest
 
 import torch
 
-from diffusers import (
-    IFPipeline,
-)
+from diffusers import IFPipeline
 from diffusers.models.attention_processor import AttnAddedKVProcessor
 from diffusers.utils.import_utils import is_xformers_available
 

tests/pipelines/flux/test_pipeline_flux.py

@@ -4,7 +4,7 @@ import unittest
 import numpy as np
 import torch
 from huggingface_hub import hf_hub_download
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -93,7 +93,8 @@ class FluxPipelineFastTests(
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_control.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxControlPipeline, FluxTransformer2DModel
 
@@ -53,7 +53,8 @@ class FluxControlPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_control_img2img.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -57,7 +57,8 @@ class FluxControlImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_control_inpaint.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -58,7 +58,8 @@ class FluxControlInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_fill.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxFillPipeline, FluxTransformer2DModel
 
@@ -58,7 +58,8 @@ class FluxFillPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_img2img.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxImg2ImgPipeline, FluxTransformer2DModel
 
@@ -55,7 +55,8 @@ class FluxImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin, FluxI
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_inpaint.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxInpaintPipeline, FluxTransformer2DModel
 
@@ -55,7 +55,8 @@ class FluxInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin, FluxI
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_kontext.py

@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 import PIL.Image
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -79,7 +79,8 @@ class FluxKontextPipelineFastTests(
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/flux/test_pipeline_flux_kontext_inpaint.py

@@ -3,7 +3,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -79,7 +79,8 @@ class FluxKontextInpaintPipelineFastTests(
         text_encoder = CLIPTextModel(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

tests/pipelines/glm_image/test_glm_image.py

@@ -16,7 +16,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, GlmImagePipeline, GlmImageTransformer2DModel
 from diffusers.utils import is_transformers_version
@@ -57,7 +57,8 @@ class GlmImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
     def get_dummy_components(self):
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         glm_config = GlmImageConfig(

tests/pipelines/hidream_image/test_pipeline_hidream.py

@@ -18,6 +18,7 @@ import unittest
 import numpy as np
 import torch
 from transformers import (
+    AutoConfig,
     AutoTokenizer,
     CLIPTextConfig,
     CLIPTextModelWithProjection,
@@ -94,7 +95,8 @@ class HiDreamImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
         torch.manual_seed(0)
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_3 = T5EncoderModel(config)
 
         torch.manual_seed(0)
         text_encoder_4 = LlamaForCausalLM.from_pretrained("hf-internal-testing/tiny-random-LlamaForCausalLM")
@@ -149,12 +151,12 @@ class HiDreamImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         self.assertEqual(generated_image.shape, (128, 128, 3))
 
         # fmt: off
-        expected_slice = np.array([0.4507, 0.5256, 0.4205, 0.5791, 0.4848, 0.4831, 0.4443, 0.5107, 0.6586, 0.3163, 0.7318, 0.5933, 0.6252, 0.5512, 0.5357, 0.5983])
+        expected_slice = np.array([0.4501, 0.5256, 0.4207, 0.5783, 0.4842, 0.4833, 0.4441, 0.5112, 0.6587, 0.3169, 0.7308, 0.5927, 0.6251, 0.5509, 0.5355, 0.5969])
         # fmt: on
 
         generated_slice = generated_image.flatten()
         generated_slice = np.concatenate([generated_slice[:8], generated_slice[-8:]])
-        self.assertTrue(np.allclose(generated_slice, expected_slice, atol=1e-3))
+        self.assertTrue(np.allclose(generated_slice, expected_slice, atol=5e-3))
 
     def test_inference_batch_single_identical(self):
         super().test_inference_batch_single_identical(expected_max_diff=3e-4)

tests/pipelines/hunyuan_image_21/test_hunyuanimage.py

@@ -223,7 +223,7 @@ class HunyuanImagePipelineFastTests(
         self.assertEqual(generated_image.shape, (3, 16, 16))
 
         expected_slice_np = np.array(
-            [0.61494756, 0.49616697, 0.60327923, 0.6115793, 0.49047345, 0.56977504, 0.53066164, 0.58880305, 0.5570612]
+            [0.6068114, 0.48716035, 0.5984431, 0.60241306, 0.48849544, 0.5624479, 0.53696984, 0.58964247, 0.54248774]
         )
         output_slice = generated_image[0, -3:, -3:].flatten().cpu().numpy()
 

tests/pipelines/hunyuan_video/test_hunyuan_image2video.py

@@ -233,7 +233,7 @@ class HunyuanVideoImageToVideoPipelineFastTests(
         self.assertEqual(generated_video.shape, (5, 3, 16, 16))
 
         # fmt: off
-        expected_slice = torch.tensor([0.444, 0.479, 0.4485, 0.5752, 0.3539, 0.1548, 0.2706, 0.3593, 0.5323, 0.6635, 0.6795, 0.5255, 0.5091, 0.345, 0.4276, 0.4128])
+        expected_slice = torch.tensor([0.4441, 0.4790, 0.4485, 0.5748, 0.3539, 0.1553, 0.2707, 0.3594, 0.5331, 0.6645, 0.6799, 0.5257, 0.5092, 0.3450, 0.4276, 0.4127])
         # fmt: on
 
         generated_slice = generated_video.flatten()

tests/pipelines/hunyuan_video1_5/test_hunyuan_1_5.py

@@ -15,7 +15,14 @@
 import unittest
 
 import torch
-from transformers import ByT5Tokenizer, Qwen2_5_VLTextConfig, Qwen2_5_VLTextModel, Qwen2Tokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    ByT5Tokenizer,
+    Qwen2_5_VLTextConfig,
+    Qwen2_5_VLTextModel,
+    Qwen2Tokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKLHunyuanVideo15,
@@ -114,7 +121,8 @@ class HunyuanVideo15PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         tokenizer = Qwen2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration")
 
         torch.manual_seed(0)
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
         tokenizer_2 = ByT5Tokenizer()
 
         guider = ClassifierFreeGuidance(guidance_scale=1.0)

tests/pipelines/hunyuandit/test_hunyuan_dit.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, BertModel, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, BertModel, T5EncoderModel
 
 from diffusers import AutoencoderKL, DDPMScheduler, HunyuanDiT2DModel, HunyuanDiTPipeline
 
@@ -74,7 +74,9 @@ class HunyuanDiTPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         scheduler = DDPMScheduler()
         text_encoder = BertModel.from_pretrained("hf-internal-testing/tiny-random-BertModel")
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/kandinsky3/test_kandinsky3.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoPipelineForImage2Image,
@@ -108,7 +108,8 @@ class Kandinsky3PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         torch.manual_seed(0)
         movq = self.dummy_movq
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config).eval()
 
         torch.manual_seed(0)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
@@ -155,9 +156,9 @@ class Kandinsky3PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         assert image.shape == (1, 16, 16, 3)
 
-        expected_slice = np.array([0.3768, 0.4373, 0.4865, 0.4890, 0.4299, 0.5122, 0.4921, 0.4924, 0.5599])
+        expected_slice = np.array([0.3944, 0.3680, 0.4842, 0.5333, 0.4412, 0.4812, 0.5089, 0.5381, 0.5578])
 
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-1, (
             f" expected_slice {expected_slice}, but got {image_slice.flatten()}"
         )
 

tests/pipelines/kandinsky3/test_kandinsky3_img2img.py

@@ -20,7 +20,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoPipelineForImage2Image,
@@ -119,7 +119,8 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
         torch.manual_seed(0)
         movq = self.dummy_movq
         torch.manual_seed(0)
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config).eval()
 
         torch.manual_seed(0)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
@@ -155,10 +156,7 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
         return inputs
 
     def test_dict_tuple_outputs_equivalent(self):
-        expected_slice = None
-        if torch_device == "cpu":
-            expected_slice = np.array([0.5762, 0.6112, 0.4150, 0.6018, 0.6167, 0.4626, 0.5426, 0.5641, 0.6536])
-        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+        super().test_dict_tuple_outputs_equivalent()
 
     def test_kandinsky3_img2img(self):
         device = "cpu"
@@ -177,11 +175,9 @@ class Kandinsky3Img2ImgPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
 
         assert image.shape == (1, 64, 64, 3)
 
-        expected_slice = np.array(
-            [0.576259, 0.6132097, 0.41703486, 0.603196, 0.62062526, 0.4655338, 0.5434324, 0.5660727, 0.65433365]
-        )
+        expected_slice = np.array([0.5725, 0.6248, 0.4355, 0.5732, 0.6105, 0.5267, 0.5470, 0.5512, 0.6618])
 
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-1, (
             f" expected_slice {expected_slice}, but got {image_slice.flatten()}"
         )
 

tests/pipelines/latte/test_latte.py

@@ -20,7 +20,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -109,7 +109,8 @@ class LattePipelineFastTests(
         vae = AutoencoderKL()
 
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/ltx/test_ltx.py

@@ -17,7 +17,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLLTXVideo, FlowMatchEulerDiscreteScheduler, LTXPipeline, LTXVideoTransformer3DModel
 
@@ -88,7 +88,8 @@ class LTXPipelineFastTests(PipelineTesterMixin, FirstBlockCacheTesterMixin, unit
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/ltx/test_ltx_condition.py

@@ -17,7 +17,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKLLTXVideo,
@@ -92,7 +92,8 @@ class LTXConditionPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/ltx/test_ltx_image2video.py

@@ -17,7 +17,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKLLTXVideo,
@@ -91,7 +91,8 @@ class LTXImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/ltx2/test_ltx2_image2video.py

@@ -24,7 +24,8 @@ from diffusers import (
     LTX2ImageToVideoPipeline,
     LTX2VideoTransformer3DModel,
 )
-from diffusers.pipelines.ltx2 import LTX2TextConnectors
+from diffusers.pipelines.ltx2 import LTX2LatentUpsamplePipeline, LTX2TextConnectors
+from diffusers.pipelines.ltx2.latent_upsampler import LTX2LatentUpsamplerModel
 from diffusers.pipelines.ltx2.vocoder import LTX2Vocoder
 
 from ...testing_utils import enable_full_determinism
@@ -174,6 +175,15 @@ class LTX2ImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         return components
 
+    def get_dummy_upsample_component(self, in_channels=4, mid_channels=32, num_blocks_per_stage=1):
+        upsampler = LTX2LatentUpsamplerModel(
+            in_channels=in_channels,
+            mid_channels=mid_channels,
+            num_blocks_per_stage=num_blocks_per_stage,
+        )
+
+        return upsampler
+
     def get_dummy_inputs(self, device, seed=0):
         if str(device).startswith("mps"):
             generator = torch.manual_seed(seed)
@@ -287,5 +297,60 @@ class LTX2ImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         assert torch.allclose(expected_video_slice, generated_video_slice, atol=1e-4, rtol=1e-4)
         assert torch.allclose(expected_audio_slice, generated_audio_slice, atol=1e-4, rtol=1e-4)
 
+    def test_two_stages_inference_with_upsampler(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["output_type"] = "latent"
+        first_stage_output = pipe(**inputs)
+        video_latent = first_stage_output.frames
+        audio_latent = first_stage_output.audio
+
+        self.assertEqual(video_latent.shape, (1, 4, 3, 16, 16))
+        self.assertEqual(audio_latent.shape, (1, 2, 5, 2))
+        self.assertEqual(audio_latent.shape[1], components["vocoder"].config.out_channels)
+
+        upsampler = self.get_dummy_upsample_component(in_channels=video_latent.shape[1])
+        upsample_pipe = LTX2LatentUpsamplePipeline(vae=pipe.vae, latent_upsampler=upsampler)
+        upscaled_video_latent = upsample_pipe(latents=video_latent, output_type="latent", return_dict=False)[0]
+        self.assertEqual(upscaled_video_latent.shape, (1, 4, 3, 32, 32))
+
+        inputs["latents"] = upscaled_video_latent
+        inputs["audio_latents"] = audio_latent
+        inputs["output_type"] = "pt"
+        second_stage_output = pipe(**inputs)
+        video = second_stage_output.frames
+        audio = second_stage_output.audio
+
+        self.assertEqual(video.shape, (1, 5, 3, 64, 64))
+        self.assertEqual(audio.shape[0], 1)
+        self.assertEqual(audio.shape[1], components["vocoder"].config.out_channels)
+
+        # fmt: off
+        expected_video_slice = torch.tensor(
+            [
+                0.4497, 0.6757, 0.4219, 0.7686, 0.4525, 0.6483, 0.3969, 0.7404, 0.3541, 0.3039, 0.4592, 0.3521, 0.3665, 0.2785, 0.3336, 0.3079
+            ]
+        )
+        expected_audio_slice = torch.tensor(
+            [
+                0.0271, 0.0492, 0.1249, 0.1126, 0.1661, 0.1060, 0.1717, 0.0944, 0.0672, -0.0069, 0.0688, 0.0097, 0.0808, 0.1231, 0.0986, 0.0739
+            ]
+        )
+        # fmt: on
+
+        video = video.flatten()
+        audio = audio.flatten()
+        generated_video_slice = torch.cat([video[:8], video[-8:]])
+        generated_audio_slice = torch.cat([audio[:8], audio[-8:]])
+
+        assert torch.allclose(expected_video_slice, generated_video_slice, atol=1e-4, rtol=1e-4)
+        assert torch.allclose(expected_audio_slice, generated_audio_slice, atol=1e-4, rtol=1e-4)
+
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(batch_size=2, expected_max_diff=2e-2)

tests/pipelines/mochi/test_mochi.py

@@ -18,7 +18,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKLMochi, FlowMatchEulerDiscreteScheduler, MochiPipeline, MochiTransformer3DModel
 
@@ -89,7 +89,8 @@ class MochiPipelineFastTests(
 
         torch.manual_seed(0)
         scheduler = FlowMatchEulerDiscreteScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {
@@ -207,6 +208,9 @@ class MochiPipelineFastTests(
             return
 
         components = self.get_dummy_components()
+        for key in components:
+            if "text_encoder" in key and hasattr(components[key], "eval"):
+                components[key].eval()
         pipe = self.pipeline_class(**components)
         for component in pipe.components.values():
             if hasattr(component, "set_default_attn_processor"):

tests/pipelines/pag/test_pag_hunyuan_dit.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, BertModel, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, BertModel, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -67,7 +67,9 @@ class HunyuanDiTPAGPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         scheduler = DDPMScheduler()
         text_encoder = BertModel.from_pretrained("hf-internal-testing/tiny-random-BertModel")
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
-        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_2 = T5EncoderModel(config)
         tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 
         components = {

tests/pipelines/pag/test_pag_pixart_sigma.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 import diffusers
 from diffusers import (
@@ -80,7 +80,8 @@ class PixArtSigmaPAGPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         vae = AutoencoderKL()
 
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/pag/test_pag_sd3.py

@@ -3,7 +3,14 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    CLIPTextConfig,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKL,
@@ -73,7 +80,9 @@ class StableDiffusion3PAGPipelineFastTests(unittest.TestCase, PipelineTesterMixi
         torch.manual_seed(0)
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_3 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

tests/pipelines/pag/test_pag_sd3_img2img.py

@@ -5,7 +5,14 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    CLIPTextConfig,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    T5EncoderModel,
+)
 
 from diffusers import (
     AutoencoderKL,
@@ -84,7 +91,9 @@ class StableDiffusion3PAGImg2ImgPipelineFastTests(unittest.TestCase, PipelineTes
         torch.manual_seed(0)
         text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)
 
-        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder_3 = T5EncoderModel(config)
 
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
         tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

tests/pipelines/pixart_alpha/test_pixart.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -77,7 +77,10 @@ class PixArtAlphaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         vae = AutoencoderKL()
 
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/pixart_sigma/test_pixart.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import AutoTokenizer, T5EncoderModel
+from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKL,
@@ -83,7 +83,10 @@ class PixArtSigmaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         vae = AutoencoderKL()
 
         scheduler = DDIMScheduler()
-        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        config = AutoConfig.from_pretrained("hf-internal-testing/tiny-random-t5")
+        text_encoder = T5EncoderModel(config)
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
 

tests/pipelines/prx/test_pipeline_prx.py

@@ -92,7 +92,7 @@ class PRXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         }
         encoder_config = T5GemmaModuleConfig(**encoder_params)
         text_encoder_config = T5GemmaConfig(encoder=encoder_config, is_encoder_decoder=False, **encoder_params)
-        text_encoder = T5GemmaEncoder(text_encoder_config)
+        text_encoder = T5GemmaEncoder(text_encoder_config.encoder)
 
         return {
             "transformer": transformer,
@@ -256,3 +256,27 @@ class PRXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         expected_image = torch.zeros(3, 32, 32)
         max_diff = np.abs(generated_image - expected_image).max()
         self.assertLessEqual(max_diff, 1e10)
+
+    @unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
+    def test_save_load_dduf(self):
+        pass
+
+    @unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
+    def test_loading_with_variants(self):
+        pass
+
+    @unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
+    def test_pipeline_with_accelerator_device_map(self):
+        pass
+
+    @unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
+    def test_save_load_local(self):
+        pass
+
+    @unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
+    def test_save_load_optional_components(self):
+        pass
+
+    @unittest.skip("Custom T5GemmaEncoder not compatible with transformers v5.")
+    def test_torch_dtype_dict(self):
+        pass

tests/pipelines/qwenimage/test_qwenimage.py

@@ -113,7 +113,7 @@ class QwenImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             vision_start_token_id=151652,
             vision_token_id=151654,
         )
-        text_encoder = Qwen2_5_VLForConditionalGeneration(config)
+        text_encoder = Qwen2_5_VLForConditionalGeneration(config).eval()
         tokenizer = Qwen2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration")
 
         components = {
@@ -160,12 +160,12 @@ class QwenImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         self.assertEqual(generated_image.shape, (3, 32, 32))
 
         # fmt: off
-        expected_slice = torch.tensor([0.56331, 0.63677, 0.6015, 0.56369, 0.58166, 0.55277, 0.57176, 0.63261, 0.41466, 0.35561, 0.56229, 0.48334, 0.49714, 0.52622, 0.40872, 0.50208])
+        expected_slice = torch.tensor([0.5633, 0.6368, 0.6015, 0.5637, 0.5817, 0.5528, 0.5718, 0.6326, 0.4147, 0.3556, 0.5623, 0.4833, 0.4971, 0.5262, 0.4087, 0.5021])
         # fmt: on
 
         generated_slice = generated_image.flatten()
         generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
-        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=5e-3))
 
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-1)

tests/pipelines/qwenimage/test_qwenimage_controlnet.py

@@ -211,7 +211,7 @@ class QwenControlNetPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         generated_slice = generated_image.flatten()
         generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
-        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=5e-3))
 
     def test_qwen_controlnet_multicondition(self):
         device = "cpu"
@@ -255,7 +255,7 @@ class QwenControlNetPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         generated_slice = generated_image.flatten()
         generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
-        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=5e-3))
 
     def test_attention_slicing_forward_pass(
         self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3

tests/pipelines/qwenimage/test_qwenimage_edit.py

@@ -115,7 +115,7 @@ class QwenImageEditPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             vision_start_token_id=151652,
             vision_token_id=151654,
         )
-        text_encoder = Qwen2_5_VLForConditionalGeneration(config)
+        text_encoder = Qwen2_5_VLForConditionalGeneration(config).eval()
         tokenizer = Qwen2Tokenizer.from_pretrained(tiny_ckpt_id)
 
         components = {
@@ -163,12 +163,12 @@ class QwenImageEditPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         self.assertEqual(generated_image.shape, (3, 32, 32))
 
         # fmt: off
-        expected_slice = torch.tensor([[0.5637, 0.6341, 0.6001, 0.5620, 0.5794, 0.5498, 0.5757, 0.6389, 0.4174, 0.3597, 0.5649, 0.4894, 0.4969, 0.5255, 0.4083, 0.4986]])
+        expected_slice = torch.tensor([0.5637, 0.6341, 0.6001, 0.5620, 0.5794, 0.5498, 0.5757, 0.6389, 0.4174, 0.3597, 0.5649, 0.4894, 0.4969, 0.5255, 0.4083, 0.4986])
         # fmt: on
 
         generated_slice = generated_image.flatten()
         generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
-        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=5e-3))
 
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-1)

tests/pipelines/qwenimage/test_qwenimage_edit_plus.py

@@ -164,7 +164,7 @@ class QwenImageEditPlusPipelineFastTests(PipelineTesterMixin, unittest.TestCase)
         self.assertEqual(generated_image.shape, (3, 32, 32))
 
         # fmt: off
-        expected_slice = torch.tensor([[0.5637, 0.6341, 0.6001, 0.5620, 0.5794, 0.5498, 0.5757, 0.6389, 0.4174, 0.3597, 0.5649, 0.4894, 0.4969, 0.5255, 0.4083, 0.4986]])
+        expected_slice = torch.tensor([0.5640, 0.6339, 0.5997, 0.5607, 0.5799, 0.5496, 0.5760, 0.6393, 0.4172, 0.3595, 0.5655, 0.4896, 0.4971, 0.5255, 0.4088, 0.4987])
         # fmt: on
 
         generated_slice = generated_image.flatten()

tests/pipelines/skyreels_v2/test_skyreels_v2.py

@@ -18,20 +18,11 @@ import numpy as np
 import torch
 from transformers import AutoTokenizer, T5EncoderModel
 
-from diffusers import (
-    AutoencoderKLWan,
-    SkyReelsV2Pipeline,
-    SkyReelsV2Transformer3DModel,
-    UniPCMultistepScheduler,
-)
+from diffusers import AutoencoderKLWan, SkyReelsV2Pipeline, SkyReelsV2Transformer3DModel, UniPCMultistepScheduler
 
-from ...testing_utils import (
-    enable_full_determinism,
-)
+from ...testing_utils import enable_full_determinism
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
-from ..test_pipelines_common import (
-    PipelineTesterMixin,
-)
+from ..test_pipelines_common import PipelineTesterMixin
 
 
 enable_full_determinism()

tests/pipelines/skyreels_v2/test_skyreels_v2_df.py

@@ -25,13 +25,9 @@ from diffusers import (
     UniPCMultistepScheduler,
 )
 
-from ...testing_utils import (
-    enable_full_determinism,
-)
+from ...testing_utils import enable_full_determinism
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
-from ..test_pipelines_common import (
-    PipelineTesterMixin,
-)
+from ..test_pipelines_common import PipelineTesterMixin
 
 
 enable_full_determinism()

tests/pipelines/skyreels_v2/test_skyreels_v2_df_image_to_video.py

@@ -17,10 +17,7 @@ import unittest
 import numpy as np
 import torch
 from PIL import Image
-from transformers import (
-    AutoTokenizer,
-    T5EncoderModel,
-)
+from transformers import AutoTokenizer, T5EncoderModel
 
 from diffusers import (
     AutoencoderKLWan,

tests/pipelines/skyreels_v2/test_skyreels_v2_df_video_to_video.py

@@ -27,14 +27,9 @@ from diffusers import (
     UniPCMultistepScheduler,
 )
 
-from ...testing_utils import (
-    enable_full_determinism,
-    torch_device,
-)
+from ...testing_utils import enable_full_determinism, torch_device
 from ..pipeline_params import TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
-from ..test_pipelines_common import (
-    PipelineTesterMixin,
-)
+from ..test_pipelines_common import PipelineTesterMixin
 
 
 enable_full_determinism()

tests/pipelines/stable_audio/test_stable_audio.py

@@ -19,10 +19,7 @@ import unittest
 
 import numpy as np
 import torch
-from transformers import (
-    T5EncoderModel,
-    T5Tokenizer,
-)
+from transformers import AutoConfig, T5EncoderModel, T5Tokenizer
 
 from diffusers import (
     AutoencoderOobleck,
@@ -111,7 +108,8 @@ class StableAudioPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         )
         torch.manual_seed(0)
         t5_repo_id = "hf-internal-testing/tiny-random-T5ForConditionalGeneration"
-        text_encoder = T5EncoderModel.from_pretrained(t5_repo_id)
+        config = AutoConfig.from_pretrained(t5_repo_id)
+        text_encoder = T5EncoderModel(config)
         tokenizer = T5Tokenizer.from_pretrained(t5_repo_id, truncation=True, model_max_length=25)
 
         torch.manual_seed(0)