Use system token to discriminate ChatGLM3 from ChatGLM2

li-plus · Mar 13, 2024 · 4040cb7 · 4040cb7
1 parent e539592
commit 4040cb7
Showing 1 changed file with 3 additions and 2 deletions.
diff --git a/chatglm_cpp/convert.py b/chatglm_cpp/convert.py
@@ -487,8 +487,9 @@ def convert(f: BinaryIO, model_name_or_path: str, lora_model_name_or_path: Optio
 
     if model.config.model_type == "chatglm":
         if hasattr(model.config, "multi_query_attention"):
-            # hack: ChatGLM3 shares the same architecture and model config with ChatGLM2, so I have to check transformers_version to discriminate one from the other
-            if model.config.transformers_version == "4.27.1" and model.config.seq_length <= 32768: # fix: Fixed a conversion error caused by a version recognition error
+            # ChatGLM3 shares the same architecture and model config with ChatGLM2, but its tokenizer further supports system prompts,
+            # so we can check system token to discriminate ChatGLM3 from ChatGLM2.
+            if "<|system|>" not in tokenizer.tokenizer.special_tokens:
                 ChatGLM2Converter.convert(f, model, tokenizer, ggml_type)
             else:
                 ChatGLM3Converter.convert(f, model, tokenizer, ggml_type)