video: gate 'auto' attention on diffusers' real Sage usability

The 'auto' attention-backend resolver picked 'sage' whenever the
sageattention package was merely importable. But diffusers requires a
specific version with compiled CUDA kernels (0.38 needs >=2.1.1); an old
v1 wheel is importable yet rejected, which would silently drop to plain
SDPA instead of Flash. Gate on diffusers' own _CAN_USE_SAGE_ATTN (with an
is_sageattention_version fallback) so 'auto' chooses sage only when it can
actually dispatch, else flash, else SDPA.

SageAttention 2.2.0 is now built from source in the venv (CUDA kernels,
TORCH_CUDA_ARCH_LIST=8.6), so 'auto' resolves to sage on this host.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RdMufYvtTbtGDWsiZVoXce

codai/api/video.py

@@ -871,7 +871,12 @@ def _resolve_attn_backend(model_cfg: dict):
         return None
     if val == 'auto':
         import importlib.util as _u
-        if _u.find_spec('sageattention') is not None:
+        # Prefer Sage ONLY when diffusers can actually use it: it requires a
+        # specific SageAttention version with compiled CUDA kernels (e.g. 0.38
+        # needs >=2.1.1). Merely having the package importable is NOT enough — an
+        # old v1 wheel would be rejected by diffusers and silently drop us to
+        # plain SDPA instead of Flash. So gate on diffusers' own capability flag.
+        if _sage_usable():
             return 'sage'
         if _u.find_spec('flash_attn') is not None:
             return 'flash'
@@ -879,6 +884,21 @@ def _resolve_attn_backend(model_cfg: dict):
     return val
 
 
+def _sage_usable() -> bool:
+    """True only when diffusers can actually dispatch to SageAttention (correct
+    version + compiled kernels importable)."""
+    try:
+        from diffusers.models.attention_dispatch import _CAN_USE_SAGE_ATTN
+        return bool(_CAN_USE_SAGE_ATTN)
+    except Exception:
+        try:
+            from diffusers.utils import import_utils as _iu
+            return bool(_iu.is_sageattention_available()
+                        and _iu.is_sageattention_version(">=", "2.1.1"))
+        except Exception:
+            return False
+
+
 def _apply_attention_backend(pipe, model_cfg: dict) -> None:
     """Switch the Wan transformer(s) to a faster attention backend (diffusers 0.38+
     dispatcher). Best-effort and per-component: if a backend isn't available it