Pranav
Pranav
Yes, this works. Thank you Alex.
Hi @AlexCheema, - [ ] Stable diffusion (text to image) => [Stable Diffusion v2-1](https://huggingface.co/stabilityai/stable-diffusion-2-1-base) - [ ] MLX distributed inference - [ ] Tinygrad distributed inference This is my understanding...
There are no changes to that part. It's how the partition algorithm splits the shards across the devices.
Yeah, layers are non-uniform, so the split memory isn't exactly proportional to the number of layers. Can we split using the number of params?
Unet does have couple larger layers because of upsampled dims and clip text encoder has comparatively smaller layers as it can be easily split similar to llms, made of transformer...
Resolved!
For chat history, we can implement it as follows: - select a previous image from chat history or upload an image - then perform image-to-image generation using the image and...
Hi Alex, It now supports chat history 1. You can select the previously generated image from chat history by clicking on it 2. Then add an additional prompt to perform...
I don't think it is ignoring the previous image here, but it creates a latent using the encoded image and encoded prompt to generate a new image. So, it is...
Merged latest and resolved latest changes. I will try to refine it more. Regarding special cases, - I will try to remove the hardcoding from model loading - I think...