Spaces:

Zhengyi

/

LLaMA-Mesh

like 114

Running on Zero

I am trying to find a solution to fix it. This may be because the GPU quota for each people is limited and our model takes a relative long time to generate a shape.

Thanks a lot for your support !

I am familiar with GPU quotas on Huggingface but normally they at least allow me to try something 2 or 3 times before forcing me to pause, and that includes models that are too large to run on a 24GB 4090.

cc: @hysts

I looked into the issue, and it turned out that some bugs of gradio and spaces are involved:

1. This Space is using lazy caching, which is the recommended way on ZeroGPU, but due to this bug of gradio, it doesn't work as expected.
2. There's a known bug of spaces and gr.ChatInterface where signed-in users are treated as not signed-in, so ZeroGPU quota on this Space is shorter than expected.

The first issue should be avoided by setting cache_examples=False, but I don't think there's a way to avoid the second point for now.

So maybe one of the solution is not using gr.ChatInterface? Do you have suggestions for replacing gr.ChatInterface with other chatting blocks?

I tried again and at first it seemed to be working but it did error out much like yesterday and stopped generating before the end - it had listed all vertices, and had begun listing polygon faces when it stopped with the "error" message. Again no extra detail was provided with that "error" message.

When I tried to run the demo again, just after, I got this message, which is very similar to the one I got yesterday:

"Error
You have exceeded your GPU quota (120s requested vs. 44s left). Create a free account to get more usage quota."

Let me know if there is anything I can do to help you fix this problem, and thanks again for your support.

@augmentedrealitycat I think this is because I use gr.ChatInterface and this will cause the users have less GPU quota than usual. Maybe a solution is to replace this block.

@Zhengyi

So maybe one of the solution is not using gr.ChatInterface? Do you have suggestions for replacing gr.ChatInterface with other chatting blocks?

Yeah, good idea! Replacing gr.ChatInterface with more basic gradio components would probably fix the issue.

You can use gr.Chatbot and gr.Textbox to make a simple chatbot like this:

import gradio as gr


def fn(message, history):
    history.append({"role": "user", "content": message})
    history.append({"role": "assistant", "content": message * 2})
    return history


with gr.Blocks() as demo:
    chatbot = gr.Chatbot(type="messages")
    text = gr.Textbox(submit_btn=True)
    gr.Examples(examples=[["hey"], ["hello"]], inputs=text)
    text.submit(
        fn=fn,
        inputs=[text, chatbot],
        outputs=chatbot,
    ).success(
        fn=lambda: None,
        outputs=text,
    )
demo.launch()

Same issue. Single run could not finish. Any updates?