Fix bug in half precision for DPMSolverMultistepScheduler

[rtaori]

The following code fails for the new DPMSolverMultistepScheduler when used in half precision due to torch.quantile() expecting inputs to be in full precision. This slight patch fixes this bug and allows for the following use case:

pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_mega",
                                         scheduler=DPMSolverMultistepScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule='scaled_linear'),
                                         torch_dtype=torch.float16, revision="fp16", safety_checker=None, use_auth_token=True).to('cuda')
pipe.text2img('a picture of a stealth fighter', num_images_per_prompt=10, num_inference_steps=50, output_type='numpy')

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[patrickvonplaten]

Thanks a lot for the fix! Could you maybe add one test to:

diffusers/tests/test_scheduler.py

Line 787 in 78a6eed

class DPMSolverMultistepSchedulerTest(SchedulerCommonTest):

to showcase your PR fixes a bug? :-)

[rtaori]

Added a test!

[patrickvonplaten]

Thanks a lot @rtaori !