[ci] fix compilation CI.

src/diffusers/models/attention_dispatch.py

@@ -418,7 +418,8 @@ def dispatch_attention_fn(
         # variable), or we use a custom backend based on whether user is using the `attention_backend` context manager
         backend_name, backend_fn = _AttentionBackendRegistry.get_active_backend()
     else:
-        backend_name = AttentionBackendName(backend)
+        # Skip the `Enum` call for members: older torch versions cannot always trace it in a full graph.
+        backend_name = backend if isinstance(backend, AttentionBackendName) else AttentionBackendName(backend)
         backend_fn = _AttentionBackendRegistry._backends.get(backend_name)
 
     kwargs = {

tests/models/autoencoders/test_models_kandinsky6_vocoder.py

@@ -17,14 +17,11 @@
 import torch
 
 from diffusers import MMAudioVocoder
+from diffusers.utils import is_torch_version
 from diffusers.utils.torch_utils import randn_tensor
 
 from ...testing_utils import enable_full_determinism, torch_device
-from ..testing_utils import (
-    BaseModelTesterConfig,
-    MemoryTesterMixin,
-    ModelTesterMixin,
-)
+from ..testing_utils import BaseModelTesterConfig, MemoryTesterMixin, ModelTesterMixin, TorchCompileTesterMixin
 
 
 enable_full_determinism()
@@ -104,3 +101,28 @@ def test_group_offloading_with_layerwise_casting(self, record_stream, offload_ty
     )
     def test_group_offloading_with_disk(self, tmp_path, record_stream, offload_type):
         super().test_group_offloading_with_disk(tmp_path, record_stream, offload_type)
+
+
+class TestMMAudioVocoderTorchCompile(MMAudioVocoderTesterConfig, TorchCompileTesterMixin):
+    @pytest.mark.xfail(
+        condition=is_torch_version("<", "2.12"),
+        reason=(
+            'With deterministic algorithms on, torch < 2.12 routes `F.pad(mode="replicate")` through `importlib`, '
+            "which Dynamo cannot trace in a full graph."
+        ),
+        raises=torch._dynamo.exc.Unsupported,
+        strict=True,
+    )
+    def test_torch_compile_recompilation_and_graph_break(self):
+        super().test_torch_compile_recompilation_and_graph_break()
+
+    @pytest.mark.xfail(
+        reason=(
+            "Block-level offloading hooks each `ups[i]` `ModuleList`, but `forward` calls `ups[i][0]` directly, so the "
+            "hook never onloads the weights."
+        ),
+        raises=RuntimeError,
+        strict=True,
+    )
+    def test_compile_with_group_offloading(self):
+        super().test_compile_with_group_offloading()

tests/models/latent_upscaler/test_models_kandinsky6_latent_upscaler.py

@@ -17,14 +17,11 @@
 import torch
 
 from diffusers import Kandinsky6SRLatentUpscalerBank
+from diffusers.utils import is_torch_version
 from diffusers.utils.torch_utils import randn_tensor
 
 from ...testing_utils import enable_full_determinism, torch_device
-from ..testing_utils import (
-    BaseModelTesterConfig,
-    MemoryTesterMixin,
-    ModelTesterMixin,
-)
+from ..testing_utils import BaseModelTesterConfig, MemoryTesterMixin, ModelTesterMixin, TorchCompileTesterMixin
 
 
 enable_full_determinism()
@@ -116,3 +113,19 @@ def test_disk_offload_without_safetensors(self, base_model_output, tmp_path):
     @DISK_OFFLOAD_NUMERICS
     def test_disk_offload_with_safetensors(self, base_model_output, tmp_path):
         super().test_disk_offload_with_safetensors(base_model_output, tmp_path)
+
+
+class TestKandinsky6SRLatentUpscalerBankTorchCompile(
+    Kandinsky6SRLatentUpscalerBankTesterConfig, TorchCompileTesterMixin
+):
+    @pytest.mark.xfail(
+        condition=is_torch_version("<", "2.12"),
+        reason=(
+            'With deterministic algorithms on, torch < 2.12 routes `F.pad(mode="replicate")` through `importlib`, '
+            "which Dynamo cannot trace in a full graph."
+        ),
+        raises=torch._dynamo.exc.Unsupported,
+        strict=True,
+    )
+    def test_torch_compile_recompilation_and_graph_break(self):
+        super().test_torch_compile_recompilation_and_graph_break()

tests/models/testing_utils/quantization.py

@@ -2074,6 +2074,12 @@ def test_bfl_rejects_unquantized_checkpoint(self):
 @is_torch_compile
 @require_accelerator
 @require_accelerate
+# Both schemes store E4M3 tensors (fp8r weights, nvfp4 block scales) and compute in bf16; Triton cannot compile either
+# below compute capability 8.9 (e.g. T4).
+@pytest.mark.skipif(
+    torch_device == "cuda" and torch.cuda.get_device_capability() < (8, 9),
+    reason="torch.compile of BFL models needs FP8 support in Triton (compute capability 8.9+)",
+)
 class BFLCompileTesterMixin(BFLConfigMixin, QuantizationCompileTesterMixin):
     """
     Mixin class for testing torch.compile with BFL (`fp8r` / `nvfp4`) quantized models.

tests/models/transformers/test_models_transformer_kandinsky6_sr.py

@@ -60,10 +60,11 @@ def get_init_dict(self) -> dict:
             "out_visual_dim": 8,
             "time_dim": 16,
             "patch_size": (1, 2, 2),
-            "model_dim": 24,
+            "model_dim": 32,
             "ff_dim": 32,
             "num_visual_blocks": 2,
-            "axes_dims": (4, 4, 4),
+            # head_dim = sum(axes_dims); Inductor lowers flex attention only for head dims of at least 16
+            "axes_dims": (4, 6, 6),
         }
 
     def _build_dummy_inputs(self, batch_size: int, num_frames: int, height: int, width: int) -> dict:
@@ -106,6 +107,17 @@ def different_shapes_for_compilation(self):
     def get_dummy_inputs(self, height: int = 16, width: int = 16) -> dict:
         return self._build_dummy_inputs(batch_size=1, num_frames=2, height=height, width=width)
 
+    @pytest.mark.xfail(
+        reason=(
+            "The token grid is split into 8x8 blocks and `sliding_tile_mask` sizes its table by `max(...)` of the block "
+            "counts, so the compiled graph specializes on the input shape and recompiles."
+        ),
+        raises=torch._dynamo.exc.RecompileError,
+        strict=True,
+    )
+    def test_compile_on_different_shapes(self):
+        super().test_compile_on_different_shapes()
+
 
 @pytest.mark.skipif(torch_device == "cpu", reason="FlexAttention does not support backward on CPU.")
 class TestKandinsky6SRTransformerTraining(Kandinsky6SRTransformerTesterConfig, TrainingTesterMixin):