ahmedheakl commited on
Commit
9be0b9d
·
verified ·
1 Parent(s): acafc25

Create merge_qwen.py

Browse files
Files changed (1) hide show
  1. merge_qwen.py +93 -0
merge_qwen.py ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import Qwen2Model, Qwen2ForCausalLM, Qwen2_5_VLPreTrainedModel, Qwen2_5_VLForConditionalGeneration, AutoProcessor, AutoTokenizer, AddedToken
2
+ import torch
3
+ from qwen_vl_utils import process_vision_info
4
+
5
+
6
+ qwen25_model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", device_map="auto", torch_dtype=torch.bfloat16)
7
+ llm_device = qwen25_model.model.device
8
+ deepseek_model = Qwen2ForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B").to(torch.bfloat16).to(llm_device)
9
+ qwen25_model.model.load_state_dict(deepseek_model.model.state_dict())
10
+ qwen25_model.lm_head.load_state_dict(deepseek_model.lm_head.state_dict())
11
+
12
+
13
+ qwen25_model = qwen25_model.to(torch.bfloat16)
14
+ min_pixels = 256*28*28
15
+ max_pixels = 1280*28*28
16
+ processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels, use_fast=False)
17
+ ID_TO_NEW_TOKEN = {
18
+ 151643: "<|end▁of▁sentence|>",
19
+ 151644: "<|User|>",
20
+ 151645: "<|Assistant|>",
21
+ 151646: "<|begin▁of▁sentence|>",
22
+ 151648: "<think>",
23
+ 151649: "</think>",
24
+ }
25
+
26
+ # The reverse mapping: new text -> old ID
27
+ NEW_TOKEN_TO_ID = {v: k for k, v in ID_TO_NEW_TOKEN.items()}
28
+
29
+ for old_id, text in ID_TO_NEW_TOKEN.items():
30
+ # Create an AddedToken that won't get split
31
+ # 'special=True' ensures it is recognized as one piece
32
+ # 'normalized=False' means "do not lowercase or strip it"
33
+ # so it is preserved exactly.
34
+ tok = AddedToken(
35
+ text,
36
+ special=True,
37
+ normalized=False,
38
+ lstrip=False,
39
+ rstrip=False,
40
+ single_word=False
41
+ )
42
+ # Register in the slow tokenizer's internal data structures:
43
+ # _added_tokens_decoder: maps ID -> AddedToken object
44
+ # _added_tokens_encoder: maps text -> ID
45
+ # Then update the trie so that it can match them in raw text.
46
+ processor.tokenizer._added_tokens_decoder[old_id] = tok
47
+ processor.tokenizer._added_tokens_encoder[text] = old_id
48
+
49
+ processor.tokenizer._update_trie()
50
+
51
+
52
+ print("Model loaded and move to GPU")
53
+ repo_name = "ahmedheakl/vlm-r1-base2"
54
+ qwen25_model.push_to_hub(repo_name)
55
+ processor.push_to_hub(repo_name)
56
+
57
+
58
+
59
+ # messages = [
60
+ # {
61
+ # "role": "user",
62
+ # "content": [
63
+ # # {
64
+ # # "type": "image",
65
+ # # "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
66
+ # # },
67
+ # {"type": "text", "text": "What is the integration of cos^2(x)"},
68
+ # ],
69
+ # }
70
+ # ]
71
+
72
+ # text = processor.apply_chat_template(
73
+ # messages, tokenize=False, add_generation_prompt=True
74
+ # )
75
+ # image_inputs, video_inputs = process_vision_info(messages)
76
+ # inputs = processor(
77
+ # text=[text],
78
+ # images=image_inputs,
79
+ # videos=video_inputs,
80
+ # padding=True,
81
+ # return_tensors="pt",
82
+ # )
83
+ # inputs = inputs.to("cuda")
84
+
85
+ # # Inference: Generation of the output
86
+ # generated_ids = qwen25_model.generate(**inputs, max_new_tokens=1000)
87
+ # generated_ids_trimmed = [
88
+ # out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
89
+ # ]
90
+ # output_text = processor.batch_decode(
91
+ # generated_ids_trimmed, skip_special_tokens=False, clean_up_tokenization_spaces=False
92
+ # )
93
+ # print(output_text[0])