102 lines
3.2 KiB
Markdown
102 lines
3.2 KiB
Markdown
|
---
|
||
|
|
frameworks:
|
||
|
|
- Pytorch
|
||
|
|
license: Apache License 2.0
|
||
|
|
tasks:
|
||
|
|
- image-text-to-text
|
||
|
|
|
||
|
|
model-type:
|
||
|
|
#如 gpt、phi、llama、chatglm、baichuan 等
|
||
|
|
- qwen
|
||
|
|
|
||
|
|
domain:
|
||
|
|
#如 nlp、cv、audio、multi-modal
|
||
|
|
- multi-modal
|
||
|
|
|
||
|
|
language:
|
||
|
|
#语言代码列表 https://help.aliyun.com/document_detail/215387.html?spm=a2c4g.11186623.0.0.9f8d7467kni6Aa
|
||
|
|
- en
|
||
|
|
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen3-8B
|
||
|
|
- Qwen/Qwen2.5-VL-7B-Instruct
|
||
|
|
#metrics:
|
||
|
|
##如 CIDEr、Blue、ROUGE 等
|
||
|
|
#- CIDEr
|
||
|
|
|
||
|
|
#tags:
|
||
|
|
##各种自定义，包括 pretrained、fine-tuned、instruction-tuned、RL-tuned 等训练方法和其他
|
||
|
|
#- pretrained
|
||
|
|
|
||
|
|
#tools:
|
||
|
|
##如 vllm、fastchat、llamacpp、AdaSeq 等
|
||
|
|
#- vllm
|
||
|
|
---
|
||
|
|
|
||
|
|
Simple-VL-8B is a vision-language (VL) model trained by integrating the language modeling capabilities of Qwen3-8B with the visual understanding architecture of Qwen2.5-VL-7B-Instruct .
|
||
|
|
|
||
|
|
The model is trained under [ms-swift](https://github.com/modelscope/ms-swift/tree/main) framework, the SOP process document can be found [here](https://swift.readthedocs.io/en/latest/BestPractices/Rapidly-Training-VL-model.html)
|
||
|
|
|
||
|
|
Base Models :
|
||
|
|
- [Qwen2.5-VL-7B-Instruct](https://www.modelscope.cn/models/Qwen/Qwen2.5-VL-7B-Instruct)
|
||
|
|
- [Qwen3-8B](https://www.modelscope.cn/models/Qwen/Qwen3-8B)
|
||
|
|
|
||
|
|
The Simple-VL-8B model was created through a two-stage fine-tuning process:
|
||
|
|
|
||
|
|
1. Architecture Modification : The original Qwen2.5-VL-7B-Instruct model's LLM component was replaced with weights from Qwen3-8B. Several key parameters in the configuration were updated to match Qwen3-8B's structure.
|
||
|
|
2. Two-Stage Training :
|
||
|
|
1. Stage 1 : Only the vision-to-language aligner (merger layer) was trained while keeping the ViT and LLM components frozen.
|
||
|
|
2. Stage 2 : All components were unfrozen and jointly fine-tuned to enhance overall performance.
|
||
|
|
|
||
|
|
Here we show a code snippet to show you how to use the chat model
|
||
|
|
|
||
|
|
```python
|
||
|
|
from modelscope import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
|
||
|
|
from qwen_vl_utils import process_vision_info
|
||
|
|
|
||
|
|
# default: Load the model on the available device(s)
|
||
|
|
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
||
|
|
"swift/Simple-VL-8B", torch_dtype="auto", device_map="auto"
|
||
|
|
)
|
||
|
|
|
||
|
|
# default processer
|
||
|
|
processor = AutoProcessor.from_pretrained("swift/Simple-VL-8B")
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{
|
||
|
|
"role": "user",
|
||
|
|
"content": [
|
||
|
|
{
|
||
|
|
"type": "image",
|
||
|
|
"image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
|
||
|
|
},
|
||
|
|
{"type": "text", "text": "Describe this image."},
|
||
|
|
],
|
||
|
|
}
|
||
|
|
]
|
||
|
|
|
||
|
|
# Preparation for inference
|
||
|
|
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 = model.generate(**inputs, max_new_tokens=128)
|
||
|
|
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=True, clean_up_tokenization_spaces=False
|
||
|
|
)
|
||
|
|
print(output_text)
|
||
|
|
|
||
|
|
```
|