Fix tie_word_embedding issue for llava_onevision model

src/transformers/models/llava_next_video/configuration_llava_next_video.py

@@ -162,5 +162,11 @@ def __init__(
 
         super().__init__(**kwargs)
 
+        # Due to a mismatch at model addition-time, the `tie_word_embeddings` was saved in the text config, even
+        # though it concerns the main model, while it was set to False by default in the main model... So we hardcode a fix here
+        if not self.tie_word_embeddings and self.text_config.tie_word_embeddings:
+            self.tie_word_embeddings = True
+            self.text_config.tie_word_embeddings = False
+
 
 __all__ = ["LlavaNextVideoConfig"]

src/transformers/models/llava_next_video/modular_llava_next_video.py

@@ -181,6 +181,12 @@ def __init__(
 
         super().__init__(**kwargs)
 
+        # Due to a mismatch at model addition-time, the `tie_word_embeddings` was saved in the text config, even
+        # though it concerns the main model, while it was set to False by default in the main model... So we hardcode a fix here
+        if not self.tie_word_embeddings and self.text_config.tie_word_embeddings:
+            self.tie_word_embeddings = True
+            self.text_config.tie_word_embeddings = False
+
 
 class LlavaNextVideoModelOutputWithPast(LlavaNextModelOutputWithPast):
     r"""

src/transformers/models/llava_onevision/configuration_llava_onevision.py

@@ -190,5 +190,10 @@ def __init__(
 
         super().__init__(**kwargs)
 
+        # Due to a mismatch at model addition-time, the `tie_word_embeddings` was saved in the text config, even
+        # though it concerns the main model, while it was set to False by default in the main model... So we hardcode a fix here
+        self.tie_word_embeddings = self.tie_word_embeddings or self.text_config.tie_word_embeddings
+        self.text_config.tie_word_embeddings = False
+
 
 __all__ = ["LlavaOnevisionConfig"]