Fix convert_internvl_weights_to_hf.py to support local paths (#38264)

fix(internvl): add local path support to convert_internvl_weights_to_hf.py

src/transformers/models/internvl/convert_internvl_weights_to_hf.py

@@ -15,7 +15,7 @@ import argparse
 import gc
 import os
 import re
-from typing import Optional
+from typing import Literal, Optional
 
 import torch
 from einops import rearrange
@@ -124,6 +124,29 @@ chat_template = (
 CONTEXT_LENGTH = 8192
 
 
+def get_lm_type(path: str) -> Literal["qwen2", "llama"]:
+    """
+    Determine the type of language model (either 'qwen2' or 'llama') based on a given model path.
+    """
+    if path not in LM_TYPE_CORRESPONDENCE.keys():
+        base_config = AutoModel.from_pretrained(path, trust_remote_code=True).config
+
+        lm_arch = base_config.llm_config.architectures[0]
+
+        if lm_arch == "InternLM2ForCausalLM":
+            lm_type = "llama"
+        elif lm_arch == "Qwen2ForCausalLM":
+            lm_type = "qwen2"
+        else:
+            raise ValueError(
+                f"Architecture '{lm_arch}' is not supported. Only 'Qwen2ForCausalLM' and 'InternLM2ForCausalLM' are recognized."
+            )
+    else:
+        lm_type: Literal["qwen2", "llama"] = LM_TYPE_CORRESPONDENCE[path]
+
+    return lm_type
+
+
 def convert_old_keys_to_new_keys(state_dict_keys: Optional[dict] = None, path: Optional[str] = None):
     """
     This function should be applied only once, on the concatenated keys to efficiently rename using
@@ -138,7 +161,7 @@ def convert_old_keys_to_new_keys(state_dict_keys: Optional[dict] = None, path: O
         output_dict = dict(zip(old_text_vision.split("\n"), new_text.split("\n")))
         old_text_language = "\n".join([key for key in state_dict_keys if key.startswith("language_model")])
         new_text = old_text_language
-        if LM_TYPE_CORRESPONDENCE[path] == "llama":
+        if get_lm_type(path) == "llama":
             for pattern, replacement in ORIGINAL_TO_CONVERTED_KEY_MAPPING_TEXT_LLAMA.items():
                 new_text = re.sub(pattern, replacement, new_text)
         elif LM_TYPE_CORRESPONDENCE[path] == "qwen2":
@@ -177,7 +200,7 @@ def get_internvl_config(input_base_path):
     llm_config = base_config.llm_config.to_dict()
     vision_config = base_config.vision_config.to_dict()
     vision_config["use_absolute_position_embeddings"] = True
-    if LM_TYPE_CORRESPONDENCE[input_base_path] == "qwen2":
+    if get_lm_type(input_base_path) == "qwen2":
         image_token_id = 151667
         language_config_class = Qwen2Config
     else:
@@ -188,7 +211,7 @@ def get_internvl_config(input_base_path):
     # Force use_cache to True
     llm_config["use_cache"] = True
     # Force correct eos_token_id for InternVL3
-    if "InternVL3" in input_base_path and LM_TYPE_CORRESPONDENCE[input_base_path] == "qwen2":
+    if "InternVL3" in input_base_path and get_lm_type(input_base_path) == "qwen2":
         llm_config["eos_token_id"] = 151645
 
     vision_config = {k: v for k, v in vision_config.items() if k not in UNNECESSARY_CONFIG_KEYS}
@@ -299,7 +322,7 @@ def write_model(
         processor.push_to_hub(hub_dir, use_temp_dir=True)
 
     # generation config
-    if LM_TYPE_CORRESPONDENCE[input_base_path] == "llama":
+    if get_lm_type(input_base_path) == "llama":
         print("Saving generation config...")
         # in the original model, eos_token is not the same in the text_config and the generation_config
         # ("</s>" - 2 in the text_config and "<|im_end|>" - 92542 in the generation_config)
@@ -323,7 +346,7 @@ def write_model(
 def write_tokenizer(
     save_dir: str, push_to_hub: bool = False, path: Optional[str] = None, hub_dir: Optional[str] = None
 ):
-    if LM_TYPE_CORRESPONDENCE[path] == "qwen2":
+    if get_lm_type(path) == "qwen2":
         tokenizer = AutoTokenizer.from_pretrained(
             "Qwen/Qwen2.5-VL-7B-Instruct",
             return_token_type_ids=False,