Merge branch 'video-wan' into video-hunyuan-gate

studio/backend/core/inference/video.py

@@ -1074,10 +1074,10 @@ class VideoBackend:
                 vae_tiling = vae_tiling,
                 memory_mode = plan.requested_mode,
                 speed_mode = effective_speed,
-                # Only the optimisations that actually engaged: apply_speed_optims
-                # returns every flag with True/False, and iterating the dict raw
-                # would report disabled ones as active in /video/status.
-                speed_optims = tuple(k for k, v in (speed_optims or {}).items() if v),
+                # Already filtered above to only the optimisations that engaged;
+                # apply_speed_optims returns every flag True/False and the view
+                # loop keeps just the True names.
+                speed_optims = speed_optims,
                 backend_flags = backend_flags,
                 attention_backend = attention_engaged,
                 transformer_cache = cache_engaged,

studio/backend/tests/test_video_backend.py

@@ -486,6 +486,25 @@ def test_load_generate_unload_gguf(fake_runtime, tmp_path):
     assert status["loaded"] is False
 
 
+def test_load_records_engaged_speed_optims(fake_runtime, tmp_path, monkeypatch):
+    # Regression: the load tail once re-ran the already-filtered speed_optims
+    # tuple through ``.items()`` as if it were still the raw applied dict, so
+    # every real-GPU load (where at least channels_last engages) crashed with
+    # 'tuple' object has no attribute 'items'. Fake runtime forces every optim
+    # False, so this only reproduces when one is made to engage.
+    from core.inference import video as video_mod
+
+    monkeypatch.setattr(
+        video_mod,
+        "apply_speed_optims",
+        lambda *a, **k: {"channels_last": True, "cudnn_benchmark": False},
+    )
+    backend = VideoBackend()
+    status = _load_gguf(backend, tmp_path)
+    assert status["loaded"] is True
+    assert status["speed_optims"] == ["channels_last"]
+
+
 def test_generate_defaults_from_variant(fake_runtime, tmp_path):
     # A distilled GGUF pick defaults to the few-step no-CFG schedule.
     (tmp_path / "ltx-2.3-22b-distilled-1.1-Q4_K_M.gguf").write_bytes(b"w")