[tests] fix audioldm2 tests. (#13293)

fix audioldm2 tests.

.github/workflows/release_tests_fast.yml

@@ -4,6 +4,7 @@
 name: (Release) Fast GPU Tests on main
 
 on:
+  workflow_dispatch:
   push:
     branches:
       - "v*.*.*-release"
@@ -33,6 +34,7 @@ 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
       - name: Environment
         run: |
           python utils/print_env.py
@@ -74,6 +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
       - name: Environment
         run: |
           python utils/print_env.py
@@ -125,6 +128,7 @@ jobs:
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
     - name: Environment
       run: |
@@ -175,6 +179,7 @@ jobs:
           uv pip install -e ".[quality]"
           uv pip install peft@git+https://github.com/huggingface/peft.git
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
 
       - name: Environment
         run: |
@@ -232,6 +237,7 @@ 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
     - name: Environment
       run: |
         python utils/print_env.py
@@ -274,6 +280,7 @@ 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
     - name: Environment
       run: |
         python utils/print_env.py
@@ -316,6 +323,7 @@ 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
 
     - name: Environment
       run: |

docs/source/en/optimization/attention_backends.md

@@ -143,6 +143,7 @@ Refer to the table below for a complete list of available attention backends and
 | `flash_varlen` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention |
 | `flash_varlen_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention from kernels |
 | `aiter` | [AI Tensor Engine for ROCm](https://github.com/ROCm/aiter) | FlashAttention for AMD ROCm |
+| `flash_4_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-4 |
 | `_flash_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 |
 | `_flash_varlen_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention-3 |
 | `_flash_3_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 from kernels |

src/diffusers/models/attention_dispatch.py

@@ -229,6 +229,7 @@ class AttentionBackendName(str, Enum):
     FLASH_HUB = "flash_hub"
     FLASH_VARLEN = "flash_varlen"
     FLASH_VARLEN_HUB = "flash_varlen_hub"
+    FLASH_4_HUB = "flash_4_hub"
     _FLASH_3 = "_flash_3"
     _FLASH_VARLEN_3 = "_flash_varlen_3"
     _FLASH_3_HUB = "_flash_3_hub"
@@ -358,6 +359,11 @@ _HUB_KERNELS_REGISTRY: dict["AttentionBackendName", _HubKernelConfig] = {
         function_attr="sageattn",
         version=1,
     ),
+    AttentionBackendName.FLASH_4_HUB: _HubKernelConfig(
+        repo_id="kernels-staging/flash-attn4",
+        function_attr="flash_attn_func",
+        version=0,
+    ),
 }
 
 
@@ -521,6 +527,7 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
         AttentionBackendName._FLASH_3_HUB,
         AttentionBackendName._FLASH_3_VARLEN_HUB,
         AttentionBackendName.SAGE_HUB,
+        AttentionBackendName.FLASH_4_HUB,
     ]:
         if not is_kernels_available():
             raise RuntimeError(
@@ -531,6 +538,11 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
                 f"Backend '{backend.value}' needs to be used with a `kernels` version of at least 0.12. Please update with `pip install -U kernels`."
             )
 
+        if backend == AttentionBackendName.FLASH_4_HUB and not is_kernels_available(">=", "0.12.3"):
+            raise RuntimeError(
+                f"Backend '{backend.value}' needs to be used with a `kernels` version of at least 0.12.3. Please update with `pip install -U kernels`."
+            )
+
     elif backend == AttentionBackendName.AITER:
         if not _CAN_USE_AITER_ATTN:
             raise RuntimeError(
@@ -2676,6 +2688,37 @@ def _flash_attention_3_varlen_hub(
     return (out, lse) if return_lse else out
 
 
+@_AttentionBackendRegistry.register(
+    AttentionBackendName.FLASH_4_HUB,
+    constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+    supports_context_parallel=False,
+)
+def _flash_attention_4_hub(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: torch.Tensor | None = None,
+    scale: float | None = None,
+    is_causal: bool = False,
+    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 flash-attn 4.")
+
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_4_HUB].kernel_fn
+    out = func(
+        q=query,
+        k=key,
+        v=value,
+        softmax_scale=scale,
+        causal=is_causal,
+    )
+    if isinstance(out, tuple):
+        return (out[0], out[1]) if return_lse else out[0]
+    return out
+
+
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_VARLEN_3,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],

src/diffusers/pipelines/audioldm2/pipeline_audioldm2.py

@@ -324,17 +324,18 @@ class AudioLDM2Pipeline(DiffusionPipeline):
             `inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, hidden_size)`):
                 The sequence of generated hidden-states.
         """
-        cache_position_kwargs = {}
-        if is_transformers_version("<", "4.52.1"):
-            cache_position_kwargs["input_ids"] = inputs_embeds
-        else:
-            cache_position_kwargs["seq_length"] = inputs_embeds.shape[0]
-            cache_position_kwargs["device"] = (
-                self.language_model.device if getattr(self, "language_model", None) is not None else self.device
-            )
-        cache_position_kwargs["model_kwargs"] = model_kwargs
         max_new_tokens = max_new_tokens if max_new_tokens is not None else self.language_model.config.max_new_tokens
-        model_kwargs = self.language_model._get_initial_cache_position(**cache_position_kwargs)
+        if hasattr(self.language_model, "_get_initial_cache_position"):
+            cache_position_kwargs = {}
+            if is_transformers_version("<", "4.52.1"):
+                cache_position_kwargs["input_ids"] = inputs_embeds
+            else:
+                cache_position_kwargs["seq_length"] = inputs_embeds.shape[0]
+                cache_position_kwargs["device"] = (
+                    self.language_model.device if getattr(self, "language_model", None) is not None else self.device
+                )
+            cache_position_kwargs["model_kwargs"] = model_kwargs
+            model_kwargs = self.language_model._get_initial_cache_position(**cache_position_kwargs)
 
         for _ in range(max_new_tokens):
             # prepare model inputs

tests/lora/test_lora_layers_wanvace.py

@@ -28,7 +28,6 @@ from diffusers.utils.import_utils import is_peft_available
 
 from ..testing_utils import (
     floats_tensor,
-    is_flaky,
     require_peft_backend,
     require_peft_version_greater,
     skip_mps,
@@ -46,7 +45,6 @@ from .utils import PeftLoraLoaderMixinTests  # noqa: E402
 
 @require_peft_backend
 @skip_mps
-@is_flaky(max_attempts=10, description="very flaky class")
 class WanVACELoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = WanVACEPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
@@ -73,8 +71,8 @@ class WanVACELoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
         "base_dim": 3,
         "z_dim": 4,
         "dim_mult": [1, 1, 1, 1],
-        "latents_mean": torch.randn(4).numpy().tolist(),
-        "latents_std": torch.randn(4).numpy().tolist(),
+        "latents_mean": [-0.7571, -0.7089, -0.9113, -0.7245],
+        "latents_std": [2.8184, 1.4541, 2.3275, 2.6558],
         "num_res_blocks": 1,
         "temperal_downsample": [False, True, True],
     }