From 5b7f4bc6c1c12111623261e03db480a50a90aa89 Mon Sep 17 00:00:00 2001 From: isaac-vidas <80056737+isaac-vidas@users.noreply.github.com> Date: Fri, 19 Jan 2024 07:31:25 -0500 Subject: [PATCH] [`Llava`] Fix convert_llava_weights_to_hf.py script (#28570) * Update convert_llava_weights_to_hf.py Fix call to `tokenizer.add_tokens` * Add special_tokens to tokenizer.add_tokens in convert_vipllava_weights_to_hf.py --- src/transformers/models/llava/convert_llava_weights_to_hf.py | 2 +- .../models/vipllava/convert_vipllava_weights_to_hf.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/src/transformers/models/llava/convert_llava_weights_to_hf.py b/src/transformers/models/llava/convert_llava_weights_to_hf.py index 65b58236db1..cd2f46e003f 100644 --- a/src/transformers/models/llava/convert_llava_weights_to_hf.py +++ b/src/transformers/models/llava/convert_llava_weights_to_hf.py @@ -55,7 +55,7 @@ def convert_llava_llama_to_hf(text_model_id, vision_model_id, output_hub_path, o text_config = AutoConfig.from_pretrained(text_model_id) tokenizer = AutoTokenizer.from_pretrained(text_model_id) - tokenizer.add_tokens(AddedToken("", special=True, normalized=False), special=True) + tokenizer.add_tokens(AddedToken("", special=True, normalized=False), special_tokens=True) tokenizer.add_special_tokens({"pad_token": ""}) image_processor = CLIPImageProcessor.from_pretrained(vision_model_id) diff --git a/src/transformers/models/vipllava/convert_vipllava_weights_to_hf.py b/src/transformers/models/vipllava/convert_vipllava_weights_to_hf.py index a96d56084ce..9c56cf74a7a 100644 --- a/src/transformers/models/vipllava/convert_vipllava_weights_to_hf.py +++ b/src/transformers/models/vipllava/convert_vipllava_weights_to_hf.py @@ -58,7 +58,7 @@ def convert_vipllava_llama_to_hf(text_model_id, vision_model_id, output_hub_path text_config = AutoConfig.from_pretrained(text_model_id) tokenizer = AutoTokenizer.from_pretrained(text_model_id) - tokenizer.add_tokens(AddedToken("", special=True, normalized=False)) + tokenizer.add_tokens(AddedToken("", special=True, normalized=False), special_tokens=True) tokenizer.add_special_tokens({"pad_token": ""}) image_processor = CLIPImageProcessor.from_pretrained(vision_model_id)