Possible to use LLM online API?

[SunnyPai0413]

Hi HunyuanVideoWrapper Team,
Sorry for disturbing by opening an issue, but I am not able to find the discussion selections of this repository.
I saw a llava-llama-3-8b-text-encoder-tokenizer when loading the encoder. As for my low vram (P100-16G), it takes me for 0.5hrs to run the TextEncode step and just 2min for video sampler. I really wonder that can I use online api instead to reduce the vram usage, so that I can save 90% of the time when running a batch generation?
Thank you again for your contributing, and answering our questions. Wish you have a nice day!
SunnyPai

[DocShotgun]

I'm curious what resolution/frames/steps settings you are using that would only require 2 minutes for the video sampler. For me, the time spent during generation is probably like 95% on video sampling, 4% on VAE decoding, and <1% on text encoding.

[kurttu4]

maybe make support for quantum versions of the GGuf model https://huggingface.co/IbnAbdeen/llava-llama-3-8b-text-encoder-tokenizer-Q4_K_M-GGUF

[SunnyPai0413]

I'm curious what resolution/frames/steps settings you are using that would only require 2 minutes for the video sampler. For me, the time spent during generation is probably like 95% on video sampling, 4% on VAE decoding, and <1% on text encoding.

just using with 96*160 resolution, 30 steps, 21frames/9fps

[SunnyPai0413]

maybe make support for quantum versions of the GGuf model https://huggingface.co/IbnAbdeen/llava-llama-3-8b-text-encoder-tokenizer-Q4_K_M-GGUF

Thank you. I'll have a try for that Q4 LLM. I'm trying to make a model for movements generation, and planning to build a dataset generated by this model.