From 7215fc8e3466a9129ddd340d7a31abce75954276 Mon Sep 17 00:00:00 2001
From: liminfei-amd <91481003+liminfei-amd@users.noreply.github.com>
Date: Thu, 25 Jun 2026 11:28:50 +0800
Subject: [PATCH 1/2] model_management: tolerate cleanup_models finalizer
 re-entrancy in free_memory

free_memory iterates current_loaded_models while a weakref finalizer
(cleanup_models) can fire during garbage collection on the same thread and pop
entries, leaving the cached absolute index stale (IndexError at the is_dynamic
check, single-threaded). Iterate a snapshot, carry the LoadedModel object
instead of the index, and remove unloaded entries by identity so a re-entrant
pop is tolerated.

Signed-off-by: liminfei-amd <91481003+liminfei-amd@users.noreply.github.com>
---
 comfy/model_management.py | 31 +++++++++++++++++++------------
 1 file changed, 19 insertions(+), 12 deletions(-)

diff --git a/comfy/model_management.py b/comfy/model_management.py
index b15d08ba1..c33388199 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -808,30 +808,37 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins
     can_unload = []
     unloaded_models = []
 
-    for i in range(len(current_loaded_models) -1, -1, -1):
-        shift_model = current_loaded_models[i]
+    # Iterate a snapshot: free_memory can re-enter itself on this thread when a
+    # weakref finalizer (cleanup_models) fires during GC mid-loop and pops from
+    # current_loaded_models. Carry the LoadedModel object (not a stale absolute
+    # index) and remove unloaded entries by identity so a concurrent pop is tolerated.
+    for shift_model in list(current_loaded_models):
         if device is None or shift_model.device == device:
             if shift_model not in keep_loaded and not shift_model.is_dead():
-                can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
+                can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), shift_model))
                 shift_model.currently_used = False
 
-    can_unload_sorted = sorted(can_unload)
+    can_unload_sorted = sorted(can_unload, key=lambda a: a[:3])
     for x in can_unload_sorted:
-        i = x[-1]
+        shift_model = x[-1]
         memory_to_free = 1e32
         if not DISABLE_SMART_MEMORY or device is None:
             memory_to_free = 0 if device is None else memory_required - get_free_memory(device)
-            if current_loaded_models[i].model.is_dynamic() and for_dynamic:
+            if shift_model.model.is_dynamic() and for_dynamic:
                 #don't actually unload dynamic models for the sake of other dynamic models
                 #as that works on-demand.
-                memory_required -= current_loaded_models[i].model.loaded_size()
+                memory_required -= shift_model.model.loaded_size()
                 memory_to_free = 0
-        if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
-            logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-            unloaded_model.append(i)
+        if memory_to_free > 0 and shift_model.model_unload(memory_to_free):
+            logging.debug(f"Unloading {shift_model.model.model.__class__.__name__}")
+            unloaded_model.append(shift_model)
 
-    for i in sorted(unloaded_model, reverse=True):
-        unloaded_models.append(current_loaded_models.pop(i))
+    for shift_model in unloaded_model:
+        unloaded_models.append(shift_model)
+        for _idx in range(len(current_loaded_models) - 1, -1, -1):
+            if current_loaded_models[_idx] is shift_model:
+                current_loaded_models.pop(_idx)
+                break
 
     if not for_dynamic and pins_required > 0:
         ensure_pin_budget(pins_required)

From 3711b4d01f7f79d0fd50019c04d157f66bd16953 Mon Sep 17 00:00:00 2001
From: liminfei-amd <91481003+liminfei-amd@users.noreply.github.com>
Date: Mon, 29 Jun 2026 03:08:00 +0000
Subject: [PATCH 2/2] free_memory: drop explanatory comment per review

Signed-off-by: liminfei-amd <91481003+liminfei-amd@users.noreply.github.com>
---
 comfy/model_management.py | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/comfy/model_management.py b/comfy/model_management.py
index c33388199..570084f86 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -808,10 +808,6 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins
     can_unload = []
     unloaded_models = []
 
-    # Iterate a snapshot: free_memory can re-enter itself on this thread when a
-    # weakref finalizer (cleanup_models) fires during GC mid-loop and pops from
-    # current_loaded_models. Carry the LoadedModel object (not a stale absolute
-    # index) and remove unloaded entries by identity so a concurrent pop is tolerated.
     for shift_model in list(current_loaded_models):
         if device is None or shift_model.device == device:
             if shift_model not in keep_loaded and not shift_model.is_dead():