the model is not trained using DPMSolverMultistepScheduler, it works in previous versions

[Alexandr1111111]

in version 0.16.1 the model was trained. In the new version diffusers, the DPM SolverMultistepScheduler algorithm has been changed. After model inference (image generation), the model stops learning. Changing settings does not change behavior.
Settings:
num_train_timesteps: int = 1000,
beta_start: float = 0.0001,
beta_end: float = 0.02,
beta_schedule: str = "linear",
trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
solver_order: int = 2,
prediction_type: str = "epsilon",
#prediction_type: str = "v_prediction",
thresholding: bool = False,
dynamic_thresholding_ratio: float = 0.995,
sample_max_value: float = 1.0,
algorithm_type: str = "dpmsolver++",
solver_type: str = "midpoint",
lower_order_final: bool = True,
euler_at_final: bool = False,
use_karras_sigmas: Optional[bool] = False,
use_lu_lambdas: Optional[bool] = False,
final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min"
lambda_min_clipped: float = -float("inf"),
variance_type: Optional[str] = 'learned',
timestep_spacing: str = "linspace",
steps_offset: int = 0,
rescale_betas_zero_snr: bool = False

How to fix it.

code generation:
noise_scheduler.set_timesteps(num_inference_steps=100)
x = torch.randn(1, 3, 256, 256).to('cuda')
for i, t in tqdm(enumerate(noise_scheduler.timesteps)):
with torch.no_grad():
t = torch.as_tensor([t]).cuda()
residual = model(x, t)
x = noise_scheduler.step(residual, t, x).prev_sample
x = x.detach().cpu().clip(-1, 1)[0]
img1=(x+1)/2