merge_qwen.py

from transformers import Qwen2Model, Qwen2ForCausalLM, Qwen2_5_VLPreTrainedModel, Qwen2_5_VLForConditionalGeneration, AutoProcessor, AutoTokenizer, AddedToken
import torch
from qwen_vl_utils import process_vision_info


qwen25_model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", device_map="auto", torch_dtype=torch.bfloat16)
llm_device = qwen25_model.model.device
deepseek_model = Qwen2ForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B").to(torch.bfloat16).to(llm_device)
qwen25_model.model.load_state_dict(deepseek_model.model.state_dict()) 
qwen25_model.lm_head.load_state_dict(deepseek_model.lm_head.state_dict())


qwen25_model = qwen25_model.to(torch.bfloat16)
min_pixels = 256*28*28
max_pixels = 1280*28*28
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels, use_fast=False)
ID_TO_NEW_TOKEN = {
    151643: "<｜end▁of▁sentence｜>",
    151644: "<｜User｜>",
    151645: "<｜Assistant｜>",
    151646: "<｜begin▁of▁sentence｜>",
    151648: "<think>",
    151649: "</think>",
}

# The reverse mapping: new text -> old ID
NEW_TOKEN_TO_ID = {v: k for k, v in ID_TO_NEW_TOKEN.items()}

for old_id, text in ID_TO_NEW_TOKEN.items():
    # Create an AddedToken that won't get split
    # 'special=True' ensures it is recognized as one piece
    # 'normalized=False' means "do not lowercase or strip it"
    # so it is preserved exactly.
    tok = AddedToken(
        text,
        special=True,
        normalized=False,
        lstrip=False,
        rstrip=False,
        single_word=False
    )
    # Register in the slow tokenizer's internal data structures:
    #   _added_tokens_decoder: maps ID -> AddedToken object
    #   _added_tokens_encoder: maps text -> ID
    # Then update the trie so that it can match them in raw text.
    processor.tokenizer._added_tokens_decoder[old_id] = tok
    processor.tokenizer._added_tokens_encoder[text] = old_id

processor.tokenizer._update_trie()


print("Model loaded and move to GPU")
repo_name = "ahmedheakl/vlm-r1-base2"
qwen25_model.push_to_hub(repo_name)
processor.push_to_hub(repo_name)



# messages = [
#     {
#         "role": "user",
#         "content": [
#             # {
#             #     "type": "image",
#             #     "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
#             # },
#             {"type": "text", "text": "What is the integration of cos^2(x)"},
#         ],
#     }
# ]

# text = processor.apply_chat_template(
#     messages, tokenize=False, add_generation_prompt=True
# )
# image_inputs, video_inputs = process_vision_info(messages)
# inputs = processor(
#     text=[text],
#     images=image_inputs,
#     videos=video_inputs,
#     padding=True,
#     return_tensors="pt",
# )
# inputs = inputs.to("cuda")

# # Inference: Generation of the output
# generated_ids = qwen25_model.generate(**inputs, max_new_tokens=1000)
# generated_ids_trimmed = [
#     out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
# ]
# output_text = processor.batch_decode(
#     generated_ids_trimmed, skip_special_tokens=False, clean_up_tokenization_spaces=False
# )
# print(output_text[0])