off-by-one error in pndm scheduler

[rokbok]

There appears to be an off-by-one error on this line:

diffusers/src/diffusers/schedulers/scheduling_pndm.py

Line 338 in 25a51b6

alpha_prod_t = self.alphas_cumprod[timestep + 1 - self._offset]

I got IndexError: index 1000 is out of bounds for dimension 0 with size 1000

Repro script:

from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
from PIL import Image

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", 
    torch_dtype = torch.float16,
    revision = 'fp16',
    use_auth_token = True,
).to("cuda")

with torch.autocast("cuda"):
        image = pipe(prompt="astronaut or doesn't really matter", num_inference_steps=3)

[natolambert]

We are discussing this in #444 and is being solved in #466 soon!

[rokbok]

Thanks for the heads-up