Yang Fan
Yang Fan
@Luobots @chenyzh28 微调base且不微调embedding的情况下，无法学到两个特殊token，可能对性能有一定影响，具体有多大影响我们这边暂时没有详细的数据；我们正在开发代码以解决base模型无法微调embedding问题。
@DarkLight1337 @xiangxinhello I will take a look at it.
@xiangxinhello Hi, you set dtype to 'float32' in your example code. I want to confirm that which dtype do you use in vllm and transformers?
> @xiangxinhello Hi, you set dtype to 'float32' in your example code. I want to confirm that which dtype do you use in vllm and transformers? If vllm using fp32...
> Hi, @fyabc, Do you support Qwen-VL-Chat? @xiangxinhello #8029 already supported Qwen-VL-Chat, you can try latest vllm-0.6.1.
@xiangxinhello Hi, I have tested Qwen2-VL-7B-Instruct fp16/fp32 on vllm and HF, and got the same outputs `'击掌(529,516),(583,594)'` in all three different seeds. Can you provide your GPU & environment information,...
@xiangxinhello Hi, can you add `print(model.generation_config)` into you HF example script and show the result? I want to confirm HF & vllm are using the same generation hyperparameters.
> > @xiangxinhello Hi, can you add `print(model.generation_config)` into you HF example script and show the result? I want to confirm HF & vllm are using the same generation hyperparameters....
> @fyabc are you interested in implementing this? Hi, our team are developing on Qwen2-Audio vllm support, please check [this branch](https://github.com/faychu/vllm/tree/qwen2-audio), and @faychu will take effort on it.
> Sorry I don't have time to review in detail tonight, but from a quick glance, can you add this model to the following pages? > > - Supported Models...