Upload ltx_loadonce.py with huggingface_hub

ltx_loadonce.py

@@ -148,9 +148,19 @@ def _patched_gpu_model(model):
     finally:
         v0 = torch.cuda.memory_allocated() // 1024**3
         torch.cuda.synchronize()
-        model.to("meta")
+        # to("meta") is a known non-fix: meta device has no storage so GPU allocs stay.
+        # to("cpu") actually transfers parameters + registered buffers (incl. CausalConv3d
+        # temporal caches) to system RAM, releasing CUDA allocations.
+        model.to("cpu")
+        # Also null any causal/temporal cache buffers that survive to("cpu") as CUDA refs
+        for _m in model.modules():
+            for _attr in ("cache", "_cache", "temporal_cache", "causal_cache"):
+                if getattr(_m, _attr, None) is not None:
+                    setattr(_m, _attr, None)
+        gc.collect()
+        torch.cuda.empty_cache()
         v1 = torch.cuda.memory_allocated() // 1024**3
-        print(f"[DIAG-gpu_model] {type(model).__name__} to(meta): {v0}GB -> {v1}GB", flush=True)
+        print(f"[DIAG-gpu_model] {type(model).__name__} to(cpu): {v0}GB -> {v1}GB", flush=True)
         _patched_cleanup()
 
 _gm_module.gpu_model = _patched_gpu_model