Closed
Description
Describe the bug
scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.
Reproduction
import torch
from diffusers import FluxPipeline
pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16, revision='refs/pr/1')
#pipe.enable_model_cpu_offload()
pipe.to(device='mps')
prompt = "A cat holding a sign that says hello world"
out = pipe(
prompt=prompt,
guidance_scale=0.,
height=768,
width=1360,
num_inference_steps=4,
max_sequence_length=256,
).images[0]
out.save("image.png")
it also doesn't work with cpu offload.
Logs
scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.
System Info
Git master