Merge 8d53b96264 into 35c1470935

comfy/model_patcher.py

@@ -165,6 +165,23 @@ def low_vram_patch_estimate_vram(model, key):
 
     return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR
 
+def _collect_quant_synthetic_keys():
+    # Synthetic per-parameter suffixes a QuantizedTensor surfaces in state_dict(), gathered from
+    # the core quant algo table so new quant types are covered automatically. "comfy_quant" is the
+    # metadata marker key attached alongside quantized weights.
+    keys = {"comfy_quant"}
+    try:
+        import comfy.quant_ops
+        for algo in getattr(comfy.quant_ops, "QUANT_ALGOS", {}).values():
+            keys.update(algo.get("parameters", ()))
+    except Exception:
+        keys.update({"weight_scale", "weight_scale_2", "input_scale"})
+    return keys
+
+
+_QUANT_SYNTHETIC_KEYS = _collect_quant_synthetic_keys()
+
+
 def get_key_weight(model, key):
     set_func = None
     convert_func = None
@@ -817,6 +834,14 @@ class ModelPatcher:
             if filter_prefix is not None:
                 if not k.startswith(filter_prefix):
                     continue
+            # Quantized weights (Mixed Precision Quantization) flatten into synthetic
+            # state_dict keys (e.g. *.weight_scale / *.input_scale + the comfy_quant marker)
+            # that are components of the QuantizedTensor in *.weight, not real module
+            # attributes. The *.weight key carries the convert_weight dequant path, so the
+            # synthetic sub-keys are skipped here and merging uses *.weight.
+            op_keys = k.rsplit('.', 1)
+            if len(op_keys) == 2 and op_keys[1] in _QUANT_SYNTHETIC_KEYS:
+                continue
             bk = self.backup.get(k, None)
             hbk = self.hook_backup.get(k, None)
             weight, set_func, convert_func = get_key_weight(self.model, k)