update

* fix: graceful fallback when attention backends fail to import

## Problem

External attention backends (flash_attn, xformers, sageattention, etc.) may be
installed but fail to import at runtime due to ABI mismatches. For example,
when `flash_attn` is compiled against PyTorch 2.4 but used with PyTorch 2.8,
the import fails with:

```
OSError: .../flash_attn_2_cuda.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda9SetDeviceEab
```

The current code uses `importlib.util.find_spec()` to check if packages exist,
but this only verifies the package is installed—not that it can actually be
imported. When the import fails, diffusers crashes instead of falling back to
native PyTorch attention.

## Solution

Wrap all external attention backend imports in try-except blocks that catch
`ImportError` and `OSError`. On failure:
1. Log a warning message explaining the issue
2. Set the corresponding `_CAN_USE_*` flag to `False`
3. Set the imported functions to `None`

This allows diffusers to gracefully degrade to PyTorch's native SDPA
(scaled_dot_product_attention) instead of crashing.

## Affected backends

- flash_attn (Flash Attention)
- flash_attn_3 (Flash Attention 3)
- aiter (AMD Instinct)
- sageattention (SageAttention)
- flex_attention (PyTorch Flex Attention)
- torch_npu (Huawei NPU)
- torch_xla (TPU/XLA)
- xformers (Meta xFormers)

## Testing

Tested with PyTorch 2.8.0 and flash_attn 2.7.4.post1 (compiled for PyTorch 2.4).
Before: crashes on import. After: logs warning and uses native attention.

* address review: use single logger and catch RuntimeError

- Move logger to module level instead of creating per-backend loggers
- Add RuntimeError to exception list alongside ImportError and OSError

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Apply style fixes

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

.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: |
@@ -199,16 +198,10 @@ jobs:
 
     - name: Install dependencies
       run: |
-        # Install pkgs which depend on setuptools<81 for pkg_resources first with no build isolation
-        uv pip install pip==25.2 setuptools==80.10.2
-        uv pip install --no-build-isolation k-diffusion==0.0.12
-        uv pip install --upgrade pip setuptools
-        # Install the rest as normal
         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: |
@@ -269,8 +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

@@ -76,9 +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
-
+          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
@@ -127,16 +125,10 @@ jobs:
 
     - name: Install dependencies
       run: |
-        # Install pkgs which depend on setuptools<81 for pkg_resources first with no build isolation
-        uv pip install pip==25.2 setuptools==80.10.2
-        uv pip install --no-build-isolation k-diffusion==0.0.12
-        uv pip install --upgrade pip setuptools
-        # Install the rest as normal
         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
+        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
     - name: Environment
       run: |
@@ -188,8 +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
+        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

.github/workflows/push_tests_mps.yml

@@ -41,7 +41,7 @@ jobs:
       shell: arch -arch arm64 bash {0}
       run: |
         ${CONDA_RUN} python -m pip install --upgrade pip uv
-        ${CONDA_RUN} python -m uv pip install -e ".[quality,test]"
+        ${CONDA_RUN} python -m uv pip install -e ".[quality]"
         ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio
         ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
         ${CONDA_RUN} python -m uv pip install transformers --upgrade

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

@@ -29,7 +29,7 @@ Qwen-Image comes in the following variants:
 | Qwen-Image-Edit Plus | [Qwen/Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) |
 
 > [!TIP]
-> [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
+> See the [Caching](../../optimization/cache) guide to speed up inference by storing and reusing intermediate outputs.
 
 ## LoRA for faster inference
 
@@ -190,6 +190,12 @@ For detailed benchmark scripts and results, see [this gist](https://gist.github.
   - all
   - __call__
 
+## QwenImageLayeredPipeline
+
+[[autodoc]] QwenImageLayeredPipeline
+  - all
+  - __call__
+
 ## QwenImagePipelineOutput
 
 [[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

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:

setup.py

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

src/diffusers/dependency_versions_table.py

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

src/diffusers/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_pipeline.py

@@ -5472,6 +5472,10 @@ class Flux2LoraLoaderMixin(LoraBaseMixin):
             logger.warning(warn_msg)
             state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
 
+        is_peft_format = any(k.startswith("base_model.model.") for k in state_dict)
+        if is_peft_format:
+            state_dict = {k.replace("base_model.model.", "diffusion_model."): v for k, v in state_dict.items()}
+
         is_ai_toolkit = any(k.startswith("diffusion_model.") for k in state_dict)
         if is_ai_toolkit:
             state_dict = _convert_non_diffusers_flux2_lora_to_diffusers(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
@@ -266,6 +319,10 @@ class _HubKernelConfig:
     function_attr: str
     revision: str | None = None
     kernel_fn: Callable | None = None
+    wrapped_forward_attr: str | None = None
+    wrapped_backward_attr: str | None = None
+    wrapped_forward_fn: Callable | None = None
+    wrapped_backward_fn: Callable | None = None
 
 
 # Registry for hub-based attention kernels
@@ -280,7 +337,11 @@ _HUB_KERNELS_REGISTRY: dict["AttentionBackendName", _HubKernelConfig] = {
         # revision="fake-ops-return-probs",
     ),
     AttentionBackendName.FLASH_HUB: _HubKernelConfig(
-        repo_id="kernels-community/flash-attn2", function_attr="flash_attn_func", revision=None
+        repo_id="kernels-community/flash-attn2",
+        function_attr="flash_attn_func",
+        revision=None,
+        wrapped_forward_attr="flash_attn_interface._wrapped_flash_attn_forward",
+        wrapped_backward_attr="flash_attn_interface._wrapped_flash_attn_backward",
     ),
     AttentionBackendName.FLASH_VARLEN_HUB: _HubKernelConfig(
         repo_id="kernels-community/flash-attn2", function_attr="flash_attn_varlen_func", revision=None
@@ -605,22 +666,39 @@ def _flex_attention_causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
 
 
 # ===== Helpers for downloading kernels =====
+def _resolve_kernel_attr(module, attr_path: str):
+    target = module
+    for attr in attr_path.split("."):
+        if not hasattr(target, attr):
+            raise AttributeError(f"Kernel module '{module.__name__}' does not define attribute path '{attr_path}'.")
+        target = getattr(target, attr)
+    return target
+
+
 def _maybe_download_kernel_for_backend(backend: AttentionBackendName) -> None:
     if backend not in _HUB_KERNELS_REGISTRY:
         return
     config = _HUB_KERNELS_REGISTRY[backend]
 
-    if config.kernel_fn is not None:
+    needs_kernel = config.kernel_fn is None
+    needs_wrapped_forward = config.wrapped_forward_attr is not None and config.wrapped_forward_fn is None
+    needs_wrapped_backward = config.wrapped_backward_attr is not None and config.wrapped_backward_fn is None
+
+    if not (needs_kernel or needs_wrapped_forward or needs_wrapped_backward):
         return
 
     try:
         from kernels import get_kernel
 
         kernel_module = get_kernel(config.repo_id, revision=config.revision)
-        kernel_func = getattr(kernel_module, config.function_attr)
+        if needs_kernel:
+            config.kernel_fn = _resolve_kernel_attr(kernel_module, config.function_attr)
 
-        # Cache the downloaded kernel function in the config object
-        config.kernel_fn = kernel_func
+        if needs_wrapped_forward:
+            config.wrapped_forward_fn = _resolve_kernel_attr(kernel_module, config.wrapped_forward_attr)
+
+        if needs_wrapped_backward:
+            config.wrapped_backward_fn = _resolve_kernel_attr(kernel_module, config.wrapped_backward_attr)
 
     except Exception as e:
         logger.error(f"An error occurred while fetching kernel '{config.repo_id}' from the Hub: {e}")
@@ -1071,6 +1149,237 @@ def _flash_attention_backward_op(
     return grad_query, grad_key, grad_value
 
 
+def _flash_attention_hub_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: torch.Tensor | None = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: float | None = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: "ParallelConfig" | None = None,
+):
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not yet supported for flash-attn hub kernels.")
+    if enable_gqa:
+        raise ValueError("`enable_gqa` is not yet supported for flash-attn hub kernels.")
+
+    config = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB]
+    wrapped_forward_fn = config.wrapped_forward_fn
+    wrapped_backward_fn = config.wrapped_backward_fn
+    if wrapped_forward_fn is None or wrapped_backward_fn is None:
+        raise RuntimeError(
+            "Flash attention hub kernels must expose `_wrapped_flash_attn_forward` and `_wrapped_flash_attn_backward` "
+            "for context parallel execution."
+        )
+
+    if scale is None:
+        scale = query.shape[-1] ** (-0.5)
+
+    window_size = (-1, -1)
+    softcap = 0.0
+    alibi_slopes = None
+    deterministic = False
+    grad_enabled = any(x.requires_grad for x in (query, key, value))
+
+    if grad_enabled or (_parallel_config is not None and _parallel_config.context_parallel_config._world_size > 1):
+        dropout_p = dropout_p if dropout_p > 0 else 1e-30
+
+    with torch.set_grad_enabled(grad_enabled):
+        out, lse, S_dmask, rng_state = wrapped_forward_fn(
+            query,
+            key,
+            value,
+            dropout_p,
+            scale,
+            is_causal,
+            window_size[0],
+            window_size[1],
+            softcap,
+            alibi_slopes,
+            return_lse,
+        )
+        lse = lse.permute(0, 2, 1).contiguous()
+
+    if _save_ctx:
+        ctx.save_for_backward(query, key, value, out, lse, rng_state)
+        ctx.dropout_p = dropout_p
+        ctx.scale = scale
+        ctx.is_causal = is_causal
+        ctx.window_size = window_size
+        ctx.softcap = softcap
+        ctx.alibi_slopes = alibi_slopes
+        ctx.deterministic = deterministic
+
+    return (out, lse) if return_lse else out
+
+
+def _flash_attention_hub_backward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    grad_out: torch.Tensor,
+    *args,
+    **kwargs,
+):
+    config = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB]
+    wrapped_backward_fn = config.wrapped_backward_fn
+    if wrapped_backward_fn is None:
+        raise RuntimeError(
+            "Flash attention hub kernels must expose `_wrapped_flash_attn_backward` for context parallel execution."
+        )
+
+    query, key, value, out, lse, rng_state = ctx.saved_tensors
+    grad_query, grad_key, grad_value = torch.empty_like(query), torch.empty_like(key), torch.empty_like(value)
+
+    _ = wrapped_backward_fn(
+        grad_out,
+        query,
+        key,
+        value,
+        out,
+        lse,
+        grad_query,
+        grad_key,
+        grad_value,
+        ctx.dropout_p,
+        ctx.scale,
+        ctx.is_causal,
+        ctx.window_size[0],
+        ctx.window_size[1],
+        ctx.softcap,
+        ctx.alibi_slopes,
+        ctx.deterministic,
+        rng_state,
+    )
+
+    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
+
+
+def _flash_attention_3_hub_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: torch.Tensor | None = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: float | None = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: "ParallelConfig" | None = None,
+    *,
+    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,
+):
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not yet supported for flash-attn 3 hub kernels.")
+    if dropout_p != 0.0:
+        raise ValueError("`dropout_p` is not yet supported for flash-attn 3 hub kernels.")
+    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,
+        causal=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=return_lse,
+    )
+
+    lse = None
+    if return_lse:
+        out, lse = out
+        lse = lse.permute(0, 2, 1).contiguous()
+
+    if _save_ctx:
+        ctx.save_for_backward(query, key, value)
+        ctx.scale = scale
+        ctx.is_causal = is_causal
+        ctx._hub_kernel = func
+
+    return (out, lse) if return_lse else out
+
+
+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,
+):
+    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,
+        )
+        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,
+        )
+
+    return grad_query, grad_key, grad_value
+
+
 def _sage_attention_forward_op(
     ctx: torch.autograd.function.FunctionCtx,
     query: torch.Tensor,
@@ -1109,6 +1418,46 @@ def _sage_attention_forward_op(
     return (out, lse) if return_lse else out
 
 
+def _sage_attention_hub_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: torch.Tensor | None = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: float | None = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: "ParallelConfig" | None = None,
+):
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not yet supported for Sage attention.")
+    if dropout_p > 0.0:
+        raise ValueError("`dropout_p` is not yet supported for Sage attention.")
+    if enable_gqa:
+        raise ValueError("`enable_gqa` is not yet supported for Sage attention.")
+
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName.SAGE_HUB].kernel_fn
+    out = func(
+        q=query,
+        k=key,
+        v=value,
+        tensor_layout="NHD",
+        is_causal=is_causal,
+        sm_scale=scale,
+        return_lse=return_lse,
+    )
+
+    lse = None
+    if return_lse:
+        out, lse, *_ = out
+        lse = lse.permute(0, 2, 1).contiguous()
+
+    return (out, lse) if return_lse else out
+
+
 def _sage_attention_backward_op(
     ctx: torch.autograd.function.FunctionCtx,
     grad_out: torch.Tensor,
@@ -1117,6 +1466,26 @@ def _sage_attention_backward_op(
     raise NotImplementedError("Backward pass is not implemented for Sage attention.")
 
 
+def _maybe_modify_attn_mask_npu(query: torch.Tensor, key: torch.Tensor, attn_mask: torch.Tensor | None = None):
+    # Skip Attention Mask if all values are 1, `None` mask can speedup the computation
+    if attn_mask is not None and torch.all(attn_mask != 0):
+        attn_mask = None
+
+    # Reshape Attention Mask: [batch_size, seq_len_k] -> [batch_size, 1, sqe_len_q, seq_len_k]
+    # https://www.hiascend.com/document/detail/zh/Pytorch/730/apiref/torchnpuCustomsapi/docs/context/torch_npu-npu_fusion_attention.md
+    if (
+        attn_mask is not None
+        and attn_mask.ndim == 2
+        and attn_mask.shape[0] == query.shape[0]
+        and attn_mask.shape[1] == key.shape[1]
+    ):
+        B, Sq, Skv = attn_mask.shape[0], query.shape[1], key.shape[1]
+        attn_mask = ~attn_mask.to(torch.bool)
+        attn_mask = attn_mask.unsqueeze(1).expand(B, Sq, Skv).unsqueeze(1).contiguous()
+
+    return attn_mask
+
+
 def _npu_attention_forward_op(
     ctx: torch.autograd.function.FunctionCtx,
     query: torch.Tensor,
@@ -1134,11 +1503,14 @@ def _npu_attention_forward_op(
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
 
+    attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
+
     out = npu_fusion_attention(
         query,
         key,
         value,
         query.size(2),  # num_heads
+        atten_mask=attn_mask,
         input_layout="BSND",
         pse=None,
         scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -1942,7 +2314,7 @@ def _flash_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.FLASH_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=False,
+    supports_context_parallel=True,
 )
 def _flash_attention_hub(
     query: torch.Tensor,
@@ -1960,17 +2332,35 @@ def _flash_attention_hub(
         raise ValueError("`attn_mask` is not supported for flash-attn 2.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_HUB].kernel_fn
-    out = func(
-        q=query,
-        k=key,
-        v=value,
-        dropout_p=dropout_p,
-        softmax_scale=scale,
-        causal=is_causal,
-        return_attn_probs=return_lse,
-    )
-    if return_lse:
-        out, lse, *_ = out
+    if _parallel_config is None:
+        out = func(
+            q=query,
+            k=key,
+            v=value,
+            dropout_p=dropout_p,
+            softmax_scale=scale,
+            causal=is_causal,
+            return_attn_probs=return_lse,
+        )
+        if return_lse:
+            out, lse, *_ = out
+    else:
+        out = _templated_context_parallel_attention(
+            query,
+            key,
+            value,
+            None,
+            dropout_p,
+            is_causal,
+            scale,
+            False,
+            return_lse,
+            forward_op=_flash_attention_hub_forward_op,
+            backward_op=_flash_attention_hub_backward_op,
+            _parallel_config=_parallel_config,
+        )
+        if return_lse:
+            out, lse = out
 
     return (out, lse) if return_lse else out
 
@@ -2117,7 +2507,7 @@ def _flash_attention_3(
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_3_HUB,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=False,
+    supports_context_parallel=True,
 )
 def _flash_attention_3_hub(
     query: torch.Tensor,
@@ -2132,33 +2522,68 @@ def _flash_attention_3_hub(
     return_attn_probs: bool = False,
     _parallel_config: "ParallelConfig" | None = None,
 ) -> torch.Tensor:
-    if _parallel_config:
-        raise NotImplementedError(f"{AttentionBackendName._FLASH_3_HUB.value} is not implemented for parallelism yet.")
     if attn_mask is not None:
         raise ValueError("`attn_mask` is not supported for flash-attn 3.")
 
     func = _HUB_KERNELS_REGISTRY[AttentionBackendName._FLASH_3_HUB].kernel_fn
-    out = func(
-        q=query,
-        k=key,
-        v=value,
-        softmax_scale=scale,
-        causal=is_causal,
-        qv=None,
-        q_descale=None,
-        k_descale=None,
-        v_descale=None,
+    if _parallel_config is None:
+        out = func(
+            q=query,
+            k=key,
+            v=value,
+            softmax_scale=scale,
+            causal=is_causal,
+            qv=None,
+            q_descale=None,
+            k_descale=None,
+            v_descale=None,
+            window_size=window_size,
+            softcap=softcap,
+            num_splits=1,
+            pack_gqa=None,
+            deterministic=deterministic,
+            sm_margin=0,
+            return_attn_probs=return_attn_probs,
+        )
+        return (out[0], out[1]) if return_attn_probs else out
+
+    forward_op = functools.partial(
+        _flash_attention_3_hub_forward_op,
         window_size=window_size,
         softcap=softcap,
         num_splits=1,
         pack_gqa=None,
         deterministic=deterministic,
         sm_margin=0,
-        return_attn_probs=return_attn_probs,
     )
-    # When `return_attn_probs` is True, the above returns a tuple of
-    # actual outputs and lse.
-    return (out[0], out[1]) if return_attn_probs else out
+    backward_op = functools.partial(
+        _flash_attention_3_hub_backward_op,
+        window_size=window_size,
+        softcap=softcap,
+        num_splits=1,
+        pack_gqa=None,
+        deterministic=deterministic,
+        sm_margin=0,
+    )
+    out = _templated_context_parallel_attention(
+        query,
+        key,
+        value,
+        None,
+        0.0,
+        is_causal,
+        scale,
+        False,
+        return_attn_probs,
+        forward_op=forward_op,
+        backward_op=backward_op,
+        _parallel_config=_parallel_config,
+    )
+    if return_attn_probs:
+        out, lse = out
+        return out, lse
+
+    return out
 
 
 @_AttentionBackendRegistry.register(
@@ -2668,16 +3093,17 @@ def _native_npu_attention(
     return_lse: bool = False,
     _parallel_config: "ParallelConfig" | None = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for NPU attention")
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
     if _parallel_config is None:
+        attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
+
         out = npu_fusion_attention(
             query,
             key,
             value,
             query.size(2),  # num_heads
+            atten_mask=attn_mask,
             input_layout="BSND",
             pse=None,
             scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -2692,7 +3118,7 @@ def _native_npu_attention(
             query,
             key,
             value,
-            None,
+            attn_mask,
             dropout_p,
             None,
             scale,
@@ -2789,7 +3215,7 @@ def _sage_attention(
 @_AttentionBackendRegistry.register(
     AttentionBackendName.SAGE_HUB,
     constraints=[_check_device_cuda, _check_qkv_dtype_bf16_or_fp16, _check_shape],
-    supports_context_parallel=False,
+    supports_context_parallel=True,
 )
 def _sage_attention_hub(
     query: torch.Tensor,
@@ -2817,6 +3243,23 @@ def _sage_attention_hub(
         )
         if return_lse:
             out, lse, *_ = out
+    else:
+        out = _templated_context_parallel_attention(
+            query,
+            key,
+            value,
+            None,
+            0.0,
+            is_causal,
+            scale,
+            False,
+            return_lse,
+            forward_op=_sage_attention_hub_forward_op,
+            backward_op=_sage_attention_backward_op,
+            _parallel_config=_parallel_config,
+        )
+        if return_lse:
+            out, lse = out
 
     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/transformers/transformer_flux2.py

@@ -424,7 +424,7 @@ class Flux2SingleTransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor | None,
-        temb_mod_params: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
+        temb_mod: torch.Tensor,
         image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
         joint_attention_kwargs: dict[str, Any] | None = None,
         split_hidden_states: bool = False,
@@ -436,7 +436,7 @@ class Flux2SingleTransformerBlock(nn.Module):
             text_seq_len = encoder_hidden_states.shape[1]
             hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
-        mod_shift, mod_scale, mod_gate = temb_mod_params
+        mod_shift, mod_scale, mod_gate = Flux2Modulation.split(temb_mod, 1)[0]
 
         norm_hidden_states = self.norm(hidden_states)
         norm_hidden_states = (1 + mod_scale) * norm_hidden_states + mod_shift
@@ -498,16 +498,18 @@ class Flux2TransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
-        temb_mod_params_img: tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
-        temb_mod_params_txt: tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
+        temb_mod_img: torch.Tensor,
+        temb_mod_txt: torch.Tensor,
         image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
         joint_attention_kwargs: dict[str, Any] | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         joint_attention_kwargs = joint_attention_kwargs or {}
 
         # Modulation parameters shape: [1, 1, self.dim]
-        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = temb_mod_params_img
-        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = temb_mod_params_txt
+        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = Flux2Modulation.split(temb_mod_img, 2)
+        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = Flux2Modulation.split(
+            temb_mod_txt, 2
+        )
 
         # Img stream
         norm_hidden_states = self.norm1(hidden_states)
@@ -627,15 +629,19 @@ class Flux2Modulation(nn.Module):
         self.linear = nn.Linear(dim, dim * 3 * self.mod_param_sets, bias=bias)
         self.act_fn = nn.SiLU()
 
-    def forward(self, temb: torch.Tensor) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
+    def forward(self, temb: torch.Tensor) -> torch.Tensor:
         mod = self.act_fn(temb)
         mod = self.linear(mod)
+        return mod
 
+    @staticmethod
+    # split inside the transformer blocks, to avoid passing tuples into checkpoints https://github.com/huggingface/diffusers/issues/12776
+    def split(mod: torch.Tensor, mod_param_sets: int) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
         if mod.ndim == 2:
             mod = mod.unsqueeze(1)
-        mod_params = torch.chunk(mod, 3 * self.mod_param_sets, dim=-1)
+        mod_params = torch.chunk(mod, 3 * mod_param_sets, dim=-1)
         # Return tuple of 3-tuples of modulation params shift/scale/gate
-        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(self.mod_param_sets))
+        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(mod_param_sets))
 
 
 class Flux2Transformer2DModel(
@@ -824,7 +830,7 @@ class Flux2Transformer2DModel(
 
         double_stream_mod_img = self.double_stream_modulation_img(temb)
         double_stream_mod_txt = self.double_stream_modulation_txt(temb)
-        single_stream_mod = self.single_stream_modulation(temb)[0]
+        single_stream_mod = self.single_stream_modulation(temb)
 
         # 2. Input projection for image (hidden_states) and conditioning text (encoder_hidden_states)
         hidden_states = self.x_embedder(hidden_states)
@@ -861,8 +867,8 @@ class Flux2Transformer2DModel(
                 encoder_hidden_states, hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
-                    temb_mod_params_img=double_stream_mod_img,
-                    temb_mod_params_txt=double_stream_mod_txt,
+                    temb_mod_img=double_stream_mod_img,
+                    temb_mod_txt=double_stream_mod_txt,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )
@@ -884,7 +890,7 @@ class Flux2Transformer2DModel(
                 hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=None,
-                    temb_mod_params=single_stream_mod,
+                    temb_mod=single_stream_mod,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )

src/diffusers/models/transformers/transformer_qwenimage.py

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

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

@@ -262,6 +262,9 @@ 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)
 

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/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/prx/pipeline_prx.py

@@ -18,7 +18,6 @@ import re
 import urllib.parse as ul
 from typing import Callable
 
-import ftfy
 import torch
 from transformers import (
     AutoTokenizer,
@@ -34,13 +33,13 @@ from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.prx.pipeline_output import PRXPipelineOutput
 from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
-from diffusers.utils import (
-    logging,
-    replace_example_docstring,
-)
+from diffusers.utils import is_ftfy_available, logging, replace_example_docstring
 from diffusers.utils.torch_utils import randn_tensor
 
 
+if is_ftfy_available():
+    import ftfy
+
 DEFAULT_RESOLUTION = 512
 
 ASPECT_RATIO_256_BIN = {

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/quantizers/gguf/utils.py

@@ -516,6 +516,9 @@ def dequantize_gguf_tensor(tensor):
 
     block_size, type_size = GGML_QUANT_SIZES[quant_type]
 
+    # Conver to plain tensor to avoid unnecessary __torch_function__ overhead.
+    tensor = tensor.as_tensor()
+
     tensor = tensor.view(torch.uint8)
     shape = _quant_shape_from_byte_shape(tensor.shape, type_size, block_size)
 
@@ -525,7 +528,7 @@ def dequantize_gguf_tensor(tensor):
     dequant = dequant_fn(blocks, block_size, type_size)
     dequant = dequant.reshape(shape)
 
-    return dequant.as_tensor()
+    return dequant
 
 
 class GGUFParameter(torch.nn.Parameter):

src/diffusers/quantizers/quantization_config.py

@@ -36,7 +36,7 @@ from typing import Any, Callable
 
 from packaging import version
 
-from ..utils import is_torch_available, is_torchao_available, is_torchao_version, logging
+from ..utils import deprecate, is_torch_available, is_torchao_available, is_torchao_version, logging
 
 
 if is_torch_available():
@@ -844,6 +844,8 @@ class QuantoConfig(QuantizationConfigMixin):
         modules_to_not_convert: list[str] | None = None,
         **kwargs,
     ):
+        deprecation_message = "`QuantoConfig` is deprecated and will be removed in version 1.0.0."
+        deprecate("QuantoConfig", "1.0.0", deprecation_message)
         self.quant_method = QuantizationMethod.QUANTO
         self.weights_dtype = weights_dtype
         self.modules_to_not_convert = modules_to_not_convert

src/diffusers/quantizers/quanto/quanto_quantizer.py

@@ -3,6 +3,7 @@ from typing import TYPE_CHECKING, Any
 from diffusers.utils.import_utils import is_optimum_quanto_version
 
 from ...utils import (
+    deprecate,
     get_module_from_name,
     is_accelerate_available,
     is_accelerate_version,
@@ -42,6 +43,9 @@ class QuantoQuantizer(DiffusersQuantizer):
         super().__init__(quantization_config, **kwargs)
 
     def validate_environment(self, *args, **kwargs):
+        deprecation_message = "The Quanto quantizer is deprecated and will be removed in version 1.0.0."
+        deprecate("QuantoQuantizer", "1.0.0", deprecation_message)
+
         if not is_optimum_quanto_available():
             raise ImportError(
                 "Loading an optimum-quanto quantized model requires optimum-quanto library (`pip install optimum-quanto`)"

src/diffusers/schedulers/scheduling_flow_match_lcm.py

@@ -14,6 +14,7 @@
 
 import math
 from dataclasses import dataclass
+from typing import Literal
 
 import numpy as np
 import torch
@@ -41,7 +42,7 @@ class FlowMatchLCMSchedulerOutput(BaseOutput):
             denoising loop.
     """
 
-    prev_sample: torch.FloatTensor
+    prev_sample: torch.Tensor
 
 
 class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
@@ -79,11 +80,11 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         use_beta_sigmas (`bool`, defaults to False):
             Whether to use beta sigmas for step sizes in the noise schedule during sampling.
         time_shift_type (`str`, defaults to "exponential"):
-            The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear".
-        scale_factors ('list', defaults to None)
+            The type of dynamic resolution-dependent timestep shifting to apply.
+        scale_factors (`list[float]`, *optional*, defaults to `None`):
             It defines how to scale the latents at which predictions are made.
-        upscale_mode ('str', defaults to 'bicubic')
-            Upscaling method, applied if scale-wise generation is considered
+        upscale_mode (`str`, *optional*, defaults to "bicubic"):
+            Upscaling method, applied if scale-wise generation is considered.
     """
 
     _compatibles = []
@@ -101,16 +102,33 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         max_image_seq_len: int = 4096,
         invert_sigmas: bool = False,
         shift_terminal: float | None = None,
-        use_karras_sigmas: bool = False,
-        use_exponential_sigmas: bool = False,
-        use_beta_sigmas: bool = False,
-        time_shift_type: str = "exponential",
+        use_karras_sigmas: bool | None = False,
+        use_exponential_sigmas: bool | None = False,
+        use_beta_sigmas: bool | None = False,
+        time_shift_type: Literal["exponential", "linear"] = "exponential",
         scale_factors: list[float] | None = None,
-        upscale_mode: str = "bicubic",
+        upscale_mode: Literal[
+            "nearest",
+            "linear",
+            "bilinear",
+            "bicubic",
+            "trilinear",
+            "area",
+            "nearest-exact",
+        ] = "bicubic",
     ):
         if self.config.use_beta_sigmas and not is_scipy_available():
             raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
-        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
+        if (
+            sum(
+                [
+                    self.config.use_beta_sigmas,
+                    self.config.use_exponential_sigmas,
+                    self.config.use_karras_sigmas,
+                ]
+            )
+            > 1
+        ):
             raise ValueError(
                 "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
             )
@@ -162,7 +180,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         return self._begin_index
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
-    def set_begin_index(self, begin_index: int = 0):
+    def set_begin_index(self, begin_index: int = 0) -> None:
         """
         Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
 
@@ -172,18 +190,18 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         """
         self._begin_index = begin_index
 
-    def set_shift(self, shift: float):
+    def set_shift(self, shift: float) -> None:
         self._shift = shift
 
-    def set_scale_factors(self, scale_factors: list, upscale_mode):
+    def set_scale_factors(self, scale_factors: list[float], upscale_mode: str) -> None:
         """
         Sets scale factors for a scale-wise generation regime.
 
         Args:
-            scale_factors (`list`):
-                The scale factors for each step
+            scale_factors (`list[float]`):
+                The scale factors for each step.
             upscale_mode (`str`):
-                Upscaling method
+                Upscaling method.
         """
         self._scale_factors = scale_factors
         self._upscale_mode = upscale_mode
@@ -238,16 +256,18 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
 
         return sample
 
-    def _sigma_to_t(self, sigma):
+    def _sigma_to_t(self, sigma: float | torch.FloatTensor) -> float | torch.FloatTensor:
         return sigma * self.config.num_train_timesteps
 
-    def time_shift(self, mu: float, sigma: float, t: torch.Tensor):
+    def time_shift(
+        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
+    ) -> float | np.ndarray | torch.Tensor:
         if self.config.time_shift_type == "exponential":
             return self._time_shift_exponential(mu, sigma, t)
         elif self.config.time_shift_type == "linear":
             return self._time_shift_linear(mu, sigma, t)
 
-    def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor:
+    def stretch_shift_to_terminal(self, t: np.ndarray | torch.Tensor) -> np.ndarray | torch.Tensor:
         r"""
         Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config
         value.
@@ -256,12 +276,13 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51
 
         Args:
-            t (`torch.Tensor`):
-                A tensor of timesteps to be stretched and shifted.
+            t (`torch.Tensor` or `np.ndarray`):
+                A tensor or numpy array of timesteps to be stretched and shifted.
 
         Returns:
-            `torch.Tensor`:
-                A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`.
+            `torch.Tensor` or `np.ndarray`:
+                A tensor or numpy array of adjusted timesteps such that the final value equals
+                `self.config.shift_terminal`.
         """
         one_minus_z = 1 - t
         scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal)
@@ -270,12 +291,12 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
 
     def set_timesteps(
         self,
-        num_inference_steps: int = None,
-        device: str | torch.device = None,
+        num_inference_steps: int | None = None,
+        device: str | torch.device | None = None,
         sigmas: list[float] | None = None,
-        mu: float = None,
+        mu: float | None = None,
         timesteps: list[float] | None = None,
-    ):
+    ) -> None:
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
@@ -317,43 +338,45 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         is_timesteps_provided = timesteps is not None
 
         if is_timesteps_provided:
-            timesteps = np.array(timesteps).astype(np.float32)
+            timesteps = np.array(timesteps).astype(np.float32)  # type: ignore
 
         if sigmas is None:
             if timesteps is None:
-                timesteps = np.linspace(
-                    self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps
+                timesteps = np.linspace(  # type: ignore
+                    self._sigma_to_t(self.sigma_max),
+                    self._sigma_to_t(self.sigma_min),
+                    num_inference_steps,
                 )
-            sigmas = timesteps / self.config.num_train_timesteps
+            sigmas = timesteps / self.config.num_train_timesteps  # type: ignore
         else:
-            sigmas = np.array(sigmas).astype(np.float32)
+            sigmas = np.array(sigmas).astype(np.float32)  # type: ignore
             num_inference_steps = len(sigmas)
 
         # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of
         #    "exponential" or "linear" type is applied
         if self.config.use_dynamic_shifting:
-            sigmas = self.time_shift(mu, 1.0, sigmas)
+            sigmas = self.time_shift(mu, 1.0, sigmas)  # type: ignore
         else:
-            sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)
+            sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)  # type: ignore
 
         # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value
         if self.config.shift_terminal:
-            sigmas = self.stretch_shift_to_terminal(sigmas)
+            sigmas = self.stretch_shift_to_terminal(sigmas)  # type: ignore
 
         # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules
         if self.config.use_karras_sigmas:
-            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
         elif self.config.use_exponential_sigmas:
-            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
         elif self.config.use_beta_sigmas:
-            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
 
         # 5. Convert sigmas and timesteps to tensors and move to specified device
-        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
+        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)  # type: ignore
         if not is_timesteps_provided:
-            timesteps = sigmas * self.config.num_train_timesteps
+            timesteps = sigmas * self.config.num_train_timesteps  # type: ignore
         else:
-            timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)
+            timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)  # type: ignore
 
         # 6. Append the terminal sigma value.
         #    If a model requires inverted sigma schedule for denoising but timesteps without inversion, the
@@ -370,7 +393,11 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         self._step_index = None
         self._begin_index = None
 
-    def index_for_timestep(self, timestep, schedule_timesteps=None):
+    def index_for_timestep(
+        self,
+        timestep: float | torch.Tensor,
+        schedule_timesteps: torch.Tensor | None = None,
+    ) -> int:
         if schedule_timesteps is None:
             schedule_timesteps = self.timesteps
 
@@ -382,9 +409,9 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         # case we start in the middle of the denoising schedule (e.g. for image-to-image)
         pos = 1 if len(indices) > 1 else 0
 
-        return indices[pos].item()
+        return int(indices[pos].item())
 
-    def _init_step_index(self, timestep):
+    def _init_step_index(self, timestep: float | torch.Tensor) -> None:
         if self.begin_index is None:
             if isinstance(timestep, torch.Tensor):
                 timestep = timestep.to(self.timesteps.device)
@@ -459,7 +486,12 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
                 size = [round(self._scale_factors[self._step_index] * size) for size in self._init_size]
                 x0_pred = torch.nn.functional.interpolate(x0_pred, size=size, mode=self._upscale_mode)
 
-        noise = randn_tensor(x0_pred.shape, generator=generator, device=x0_pred.device, dtype=x0_pred.dtype)
+        noise = randn_tensor(
+            x0_pred.shape,
+            generator=generator,
+            device=x0_pred.device,
+            dtype=x0_pred.dtype,
+        )
         prev_sample = (1 - sigma_next) * x0_pred + sigma_next * noise
 
         # upon completion increase step index by one
@@ -473,7 +505,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         return FlowMatchLCMSchedulerOutput(prev_sample=prev_sample)
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
-    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
+    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
         """
         Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative
         Models](https://huggingface.co/papers/2206.00364).
@@ -594,11 +626,15 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         )
         return sigmas
 
-    def _time_shift_exponential(self, mu, sigma, t):
+    def _time_shift_exponential(
+        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
+    ) -> float | np.ndarray | torch.Tensor:
         return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
 
-    def _time_shift_linear(self, mu, sigma, t):
+    def _time_shift_linear(
+        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
+    ) -> float | np.ndarray | torch.Tensor:
         return mu / (mu + (1 / t - 1) ** sigma)
 
-    def __len__(self):
+    def __len__(self) -> int:
         return self.config.num_train_timesteps

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/models/testing_utils/lora.py

@@ -375,7 +375,7 @@ class LoraHotSwappingForModelTesterMixin:
             # additionally check if dynamic compilation works.
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output0_after = model(**inputs_dict)["sample"]
@@ -390,7 +390,7 @@ class LoraHotSwappingForModelTesterMixin:
         with torch.inference_mode():
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
+                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output1_after = model(**inputs_dict)["sample"]

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_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/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

@@ -1,7 +1,6 @@
 import unittest
 
 import numpy as np
-import pytest
 import torch
 from transformers import AutoTokenizer
 from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig, T5GemmaModuleConfig
@@ -11,17 +10,11 @@ from diffusers.models import AutoencoderDC, AutoencoderKL
 from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel
 from diffusers.pipelines.prx.pipeline_prx import PRXPipeline
 from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
-from diffusers.utils import is_transformers_version
 
 from ..pipeline_params import TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin
 
 
-@pytest.mark.xfail(
-    condition=is_transformers_version(">", "4.57.1"),
-    reason="See https://github.com/huggingface/diffusers/pull/12456#issuecomment-3424228544",
-    strict=False,
-)
 class PRXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = PRXPipeline
     params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
@@ -99,7 +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,
@@ -263,3 +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)

tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3.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, FlowMatchEulerDiscreteScheduler, SD3Transformer2DModel, StableDiffusion3Pipeline
 
@@ -72,7 +79,9 @@ class StableDiffusion3PipelineFastTests(unittest.TestCase, PipelineTesterMixin):
         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/stable_diffusion_3/test_pipeline_stable_diffusion_3_img2img.py

@@ -4,7 +4,14 @@ 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 StableDiffusion3Img2ImgPipelineFastTests(PipelineLatentTesterMixin, unitte
         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/stable_diffusion_3/test_pipeline_stable_diffusion_3_inpaint.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 StableDiffusion3InpaintPipelineFastTests(PipelineLatentTesterMixin, unitte
         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")