RaushanTurganbay HF staff commited on
Commit
745cfbd
·
verified ·
1 Parent(s): 55769ad

update for chat template

Browse files
Files changed (1) hide show
  1. README.md +19 -0
README.md CHANGED
@@ -96,6 +96,25 @@ output = model.generate(**inputs, max_new_tokens=100)
96
  print(processor.decode(output[0], skip_special_tokens=True))
97
  ```
98
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
99
  ### Model optimization
100
 
101
  #### 4-bit quantization through `bitsandbytes` library
 
96
  print(processor.decode(output[0], skip_special_tokens=True))
97
  ```
98
 
99
+ -----------
100
+ From transformers>=v4.48, you can also pass image url or local path to the conversation history, and let the chat template handle the rest.
101
+ Chat template will load the image for you and return inputs in `torch.Tensor` which you can pass directly to `model.generate()`
102
+
103
+ ```python
104
+ messages = [
105
+ {
106
+ "role": "user",
107
+ "content": [
108
+ {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"}
109
+ {"type": "text", "text": "What is shown in this image?"},
110
+ ],
111
+ },
112
+ ]
113
+
114
+ inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors"pt")
115
+ output = model.generate(**inputs, max_new_tokens=50)
116
+ ```
117
+
118
  ### Model optimization
119
 
120
  #### 4-bit quantization through `bitsandbytes` library