refactor: enhance isolation handling in notify_execution_graph and process_latent_in methods

comfy/isolation/__init__.py

@@ -188,22 +188,33 @@ async def notify_execution_graph(needed_class_types: Set[str]) -> None:
         scan_shm_forensics("ISO:stop_extension", refresh_model_context=True)
 
     scan_shm_forensics("ISO:notify_graph_start", refresh_model_context=True)
+    isolated_class_types_in_graph = needed_class_types.intersection(
+        {spec.node_name for spec in _ISOLATED_NODE_SPECS}
+    )
+    graph_uses_isolation = bool(isolated_class_types_in_graph)
     logger.debug(
         "%s ISO:notify_graph_start running=%d needed=%d",
         LOG_PREFIX,
         len(_RUNNING_EXTENSIONS),
         len(needed_class_types),
     )
-    for ext_name, extension in list(_RUNNING_EXTENSIONS.items()):
-        ext_class_types = _get_class_types_for_extension(ext_name)
+    if graph_uses_isolation:
+        for ext_name, extension in list(_RUNNING_EXTENSIONS.items()):
+            ext_class_types = _get_class_types_for_extension(ext_name)
 
-        # If NONE of this extension's nodes are in the execution graph → evict
-        if not ext_class_types.intersection(needed_class_types):
-            await _stop_extension(
-                ext_name,
-                extension,
-                "isolated custom_node not in execution graph, evicting",
-            )
+            # If NONE of this extension's nodes are in the execution graph -> evict.
+            if not ext_class_types.intersection(needed_class_types):
+                await _stop_extension(
+                    ext_name,
+                    extension,
+                    "isolated custom_node not in execution graph, evicting",
+                )
+    else:
+        logger.debug(
+            "%s ISO:notify_graph_skip_evict running=%d reason=no isolated nodes in graph",
+            LOG_PREFIX,
+            len(_RUNNING_EXTENSIONS),
+        )
 
     # Isolated child processes add steady VRAM pressure; reclaim host-side models
     # at workflow boundaries so subsequent host nodes (e.g. CLIP encode) keep headroom.
@@ -217,7 +228,7 @@ async def notify_execution_graph(needed_class_types: Set[str]) -> None:
                 _WORKFLOW_BOUNDARY_MIN_FREE_VRAM_BYTES,
             )
             free_before = model_management.get_free_memory(device)
-            if free_before < required and _RUNNING_EXTENSIONS:
+            if free_before < required and _RUNNING_EXTENSIONS and graph_uses_isolation:
                 for ext_name, extension in list(_RUNNING_EXTENSIONS.items()):
                     await _stop_extension(
                         ext_name,

comfy/isolation/model_patcher_proxy_registry.py

@@ -1082,15 +1082,32 @@ class ModelPatcherRegistry(BaseRegistry[Any]):
     async def process_latent_in(
         self, instance_id: str, args: tuple, kwargs: dict
     ) -> Any:
-        return self._run_operation_with_lease(
-            instance_id,
-            "process_latent_in",
-            lambda: detach_if_grad(
-                self._get_instance(instance_id).model.process_latent_in(
-                    *args, **kwargs
-                )
-            ),
-        )
+        import torch
+
+        def _invoke() -> Any:
+            instance = self._get_instance(instance_id)
+            result = detach_if_grad(instance.model.process_latent_in(*args, **kwargs))
+
+            # DynamicVRAM + isolation: returning CUDA tensors across RPC can stall
+            # at the transport boundary. Marshal dynamic-path results as CPU and let
+            # the proxy restore placement when needed.
+            is_dynamic_fn = getattr(instance, "is_dynamic", None)
+            if callable(is_dynamic_fn) and is_dynamic_fn():
+                def _to_cpu(obj: Any) -> Any:
+                    if torch.is_tensor(obj):
+                        return obj.detach().cpu() if obj.device.type != "cpu" else obj
+                    if isinstance(obj, dict):
+                        return {k: _to_cpu(v) for k, v in obj.items()}
+                    if isinstance(obj, list):
+                        return [_to_cpu(v) for v in obj]
+                    if isinstance(obj, tuple):
+                        return tuple(_to_cpu(v) for v in obj)
+                    return obj
+
+                return _to_cpu(result)
+            return result
+
+        return self._run_operation_with_lease(instance_id, "process_latent_in", _invoke)
 
     async def process_latent_out(
         self, instance_id: str, args: tuple, kwargs: dict

comfy/model_management.py

@@ -831,7 +831,11 @@ def archive_model_dtypes(model):
 def cleanup_models():
     to_delete = []
     for i in range(len(current_loaded_models)):
-        if current_loaded_models[i].real_model() is None:
+        real_model_ref = current_loaded_models[i].real_model
+        if real_model_ref is None:
+            to_delete = [i] + to_delete
+            continue
+        if callable(real_model_ref) and real_model_ref() is None:
             to_delete = [i] + to_delete
 
     for i in to_delete: