Eric BuehlerandGitHub efbe4a69fc Our first Diffusion model: FLUX (#758) ...

* Sketch the diffusion pipeline, loader, model, and FLUX model

* Connect response part to openai server

* Sketch the flux loader

* Handle 256 limitiation of timestep distilliation model

* Work on loading

* Handle raw prompts for image gen models

* Should build now

* Clippy

* Add diffusion plain model

* Typos

* Sketch a server example

* Add parse_diffusion_arch

* Add debug bound

* Use as_any

* Some fixes

* Fix config

* Loads now

* Avoid empty prompt err

* Fix overflow

* Fix overflow

* Use the correct processor

* Add debugs

* Cast to correct dtype

* Squeeze

* Use sharded and downcast repo

* Better loading

* Dont modify cache

* Fix response

* Use big t5 for now

* Only load relevant tensors

* Clamp for f16

* Do some things

* Use sharded and downcast repo

* Diffusion model case

* Oops

* Oops

* Hotload t5xxl

* Use only encoder parts

* Don't cast it

* Fix files

* Format

* Cast every layer

* Updated resolution and prompt

* Fix issue with dual running

* Reduce lifetime of hotloaded t5 xxl

* Offloading t5 layers to cpu

* Add offloading config

* Add Python, Rust support, docs

* Add configurable height and width

2024-09-24 13:14:05 -04:00

..

main.rs

Our first Diffusion model: FLUX (#758)

2024-09-24 13:14:05 -04:00