fix(ace15): handle missing lm_metadata in memory estimation during checkpoint export Comfy-Org#12669 (Comfy-Org#12686)

Author: fappaz

comfy/text_encoders/ace15.py

@@ -328,14 +328,14 @@ def load_sd(self, sd):
                 return getattr(self, self.lm_model).load_sd(sd)
 
     def memory_estimation_function(self, token_weight_pairs, device=None):
-        lm_metadata = token_weight_pairs["lm_metadata"]
+        lm_metadata = token_weight_pairs.get("lm_metadata", {})
         constant = self.constant
         if comfy.model_management.should_use_bf16(device):
             constant *= 0.5
 
         token_weight_pairs = token_weight_pairs.get("lm_prompt", [])
         num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
-        num_tokens += lm_metadata['min_tokens']
+        num_tokens += lm_metadata.get("min_tokens", 0)
         return num_tokens * constant * 1024 * 1024
 
 def te(dtype_llama=None, llama_quantization_metadata=None, lm_model="qwen3_2b"):