jokercurry

``` local: INFO tokio-runtime-worker ThreadId(02) ballista_scheduler::state::executor_manager: Reserved 0 executor slots in 588.435µs etcd: INFO tokio-runtime-worker ThreadId(02) ballista_scheduler::state::executor_manager: Reserved 0 executor slots in 77.631762ms ``` These days, I am testing the...

> could you explain the reason of using multiple schedulers? Is it just for HA or worried about the performance of single scheduler for task scheduling? Hi @yahoNanJing At present,...

@jerqi @leixm @zuston @advancedxy If you have time, PTAL.

If it's ok, I will go ahead. Gently ping. @jerqi @zuston @kaijchen @advancedxy

> Great work. I think cli is a good way for admin to manage cluster. But do we need to trigger updating conf by CLI, the update latency is acceptable...

> > If it's ok, I will go ahead. Gently ping. @jerqi @zuston @kaijchen @advancedxy > > I think cli is a good way to manage shuffle clusters. I create...

> That sucks, why was it yanked? There may be bugs? I don’t know the specifics either.

看起来是process_vision_info方法没有处理type为text的input导致的, 看起来只支持传入video或image ``` def process_vision_info( conversations: list[dict] | list[list[dict]], ) -> tuple[list[Image.Image] | None, list[torch.Tensor | list[Image.Image]] | None]: vision_infos = extract_vision_info(conversations) ## Read images or videos image_inputs = []...

@JJJYmmm 那模型支持文本+视频抽帧的混合输入吗, 目前格式如下, 是用process_vision_info单独处理image, text用tokenizer处理吗 ``` { 'role': 'user', 'content': [{ 'image': 'frame_0000.png', 'type': 'image' }, { 'image': 'frame_0001.png', 'type': 'image' }, { 'text': 'xxxxxxxx', 'type': 'text' }] } ```