This message was deleted.

trying to achieve the same as

upscaling_scheduler = EulerDiscreteScheduler.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", subfolder="scheduler")
pipe = StableDiffusionUpscalePipeline.from_pretrained(
    "stabilityai/stable-diffusion-x4-upscaler",
    torch_dtype=torch.float16,
    scheduler=upscaling_scheduler
)
bentoml.diffusers.save_model("upscale_model", pipe)

@larme would you mind taking a look at this?

when using picklable_model works as expected

Hi Amit, the best way to use bentoml.diffusers is:

bento_model = bentoml.diffusers.get("upscale_model:latest")
runner = bento_model.to_runner()

then the runner will automatically do optimization like using half precision and put model to GPU. It can be used as computation unit. You can test it locally by

runner.init_local()
runner.run(prompt="a bento box", ...)

Later you can use the runner inside a bentoml service, this will run the model in a separated process. So you can have multiple api workers handling image pre-processing and post-processing utilizing CPU resource, while runner process utilizing GPU resource for image generation/upscaling one example may related to your usage is at: https://github.com/bentoml/diffusers-examples/blob/main/sd2_with_upscaler/service.py