As part of the development we would love to start bring community meetings that relates to vertical application topics of the areas that our community members are interested in. As a first step, we would like to start with the topic of running image generation models such as stable diffusion models.
We welcome topics related, but not limited to, approaches to optimize the diffusion and image generation models. Possible ideas and use-cases (e.g. LoRA) for related topics or anything related to the topic.
Please feel free to add to the agenda doc on topics you are interested in, or replying to this thread.
The meeting will be held on April 18th online virtually at 6 pm PDT. Please also feel free to join other related development meetings and let us know other possible model verticals you are interested in.
Possibility to compile the entire scheduler loop in SD, to remove app-level python boilerplate and make deployment significantly easier
How to work with auto-regressive model like GPT, that appends the output of one inference to the input of the next sequence. Input seq len increases by 1 on every inference for the new token. Also, does TVM have any story for the decoding step (beam search etc) to make deployment easier. ONNXRuntime has BeamSearch, GreedySearch, Sampling op etc for this purpose.
Thanks @masahi for great comments, the first is a great topic to add to the agenda. and the second one would be a great candidate for next vertical focused meeting