Add inference pipeline for VL models

[irexyc]

Motivation

Add vision language chat pipeline

Examples

from lmdeploy.vl import load_image_from_url
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig

model_path = 'liuhaotian/llava-v1.6-vicuna-7b'
model_name = 'vicuna'
# model_path = 'liuhaotian/llava-v1.5-7b'
# model_name = 'vicuna'
pipe = pipeline(model_path, 
                backend_config=TurbomindEngineConfig(session_len=8192), 
                chat_template_config=ChatTemplateConfig(model_name=model_name))

im = load_image_from_url('https://img11.kcimg.cn/officialpic/0/3/onzc.jpg').convert('RGB')
pipe(('描述这个图片', [im]))

from lmdeploy.vl import load_image_from_url
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig

model_path = 'Qwen/Qwen-VL-Chat'
model_name = 'qwen-7b'
pipe = pipeline(model_path,
                backend_config=TurbomindEngineConfig(session_len=8192),
                chat_template_config=ChatTemplateConfig(model_name=model_name))

im1 = load_image_from_url('https://raw.githubusercontent.com/QwenLM/Qwen-VL/master/assets/mm_tutorial/Shanghai.jpg').convert('RGB')
im2 = load_image_from_url('https://raw.githubusercontent.com/QwenLM/Qwen-VL/master/assets/mm_tutorial/Chongqing.jpeg').convert('RGB')
pipe(('对比一下这两个图片', [im1, im2]))

[AllentDan]

Seems lmdeploy list | grep vl can only get yi models

[AllentDan]

docstring and typehint

[lvhan028]

就用一个吧。name='yi-vl'

[AllentDan]

LGTM