* In MiniLM JAX example do not hardcode device
* In MiniLM JAX example don't use bytecode MLIR
---------
Co-authored-by: Boian Petkantchin <boian@nod-labs.com>
* Do not hardcode the name of the VM module in get_iree_module
* Add example JAX MiniLM inference
---------
Co-authored-by: Boian Petkantchin <boian@nod-labs.com>
python apps/stable_diffusion/scripts/outpaint.py --prompt="Face of a yellow cat, high resolution, sitting on a park bench" --img_path=test_imgs/overture-creations-5sI6fQgYIuo.png --import_mlir --hf_model_id="stabilityai/stable-diffusion-2-inpainting" --pixels=128 --mask_blur=8 --left --right --top --bottom --steps=20
* Revert "[SD] Modify the flags to use --iree-preprocessing-pass-pipeline (#914)"
This reverts commit a783c089a9.
* Revert "Fix iree flags due to the change in shark-runtime (#944)"
This reverts commit 1d38d49162.
The stable diffusion codebase has been reorganized to make it more
modular so that the same script can be used for web as well as cli,
instead of duplicating the whole codebase.
Signed-off-by: Gaurav Shukla <gaurav@nod-labs.com>
-- This commit uses `load_pipeline_from_original_stable_diffusion_ckpt`
as exposed due to [Diffusers PR](https://github.com/huggingface/diffusers/pull/2019).
-- It also adds a support for the end users to use `.safetensors` along
with `.ckpt` file.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
-- This commit includes two python modules as part of requirements.txt.
-- It also updates README.md to also inclue `--no-use_tuned` for users to
be able to try `hf_model_id` or `ckpt_loc` without any issue.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
-- This commit adds `batch_size` command-line arg.
-- It also involves replicating the prompt `batch_size` no. of times.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
-- Currently we require users to specify the base model on which the custom
model (.ckpt) is tuned on. Even for running a HuggingFace repo-id, we
require the users to go a tedious way of adding things to variants.json.
-- This commit aims to address the above issues and will be treated as a
starting point for a series of design changes which makes using SHARK's SD
easier.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
-- This commit fixes CKPT script to rely on the previous CKPT to Diffusers
script.
TODO: Let go of the script once the CKPT is included in next release
of diffusers.
-- It also adds many variants as part of `variants.json` and updates
`README.md` to reflect change in default `hf_model_id`.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
