Fix Dynamo lru_cache warnings during torch.compile (#13384)

* fix compile issue

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* compile friendly

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* add comments

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

---------

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

Author: jiqing-feng

src/diffusers/models/attention_dispatch.py

@@ -423,7 +423,9 @@ def dispatch_attention_fn(
         **attention_kwargs,
         "_parallel_config": parallel_config,
     }
-    if is_torch_version(">=", "2.5.0"):
+    # Equivalent to `is_torch_version(">=", "2.5.0")` — use module-level constant to avoid
+    # Dynamo tracing into the lru_cache-wrapped `is_torch_version` during torch.compile.
+    if _CAN_USE_FLEX_ATTN:
         kwargs["enable_gqa"] = enable_gqa
 
     if _AttentionBackendRegistry._checks_enabled:

src/diffusers/utils/torch_utils.py

@@ -347,7 +347,17 @@ def outer_wrapper(fn: Callable[P, T]):
 
         @functools.wraps(fn)
         def inner_wrapper(*args: P.args, **kwargs: P.kwargs):
-            if torch.compiler.is_exporting():
+            compiler = getattr(torch, "compiler", None)
+            is_exporting = bool(compiler and hasattr(compiler, "is_exporting") and compiler.is_exporting())
+            is_compiling = bool(compiler and hasattr(compiler, "is_compiling") and compiler.is_compiling())
+
+            # Fallback for older builds where compiler.is_compiling is unavailable.
+            if not is_compiling:
+                dynamo = getattr(torch, "_dynamo", None)
+                if dynamo is not None and hasattr(dynamo, "is_compiling"):
+                    is_compiling = dynamo.is_compiling()
+
+            if is_exporting or is_compiling:
                 return fn(*args, **kwargs)
             return cached(*args, **kwargs)