import numpy as np; np.trapezoid = getattr(np,"trapezoid",np.trapz) import time, torch, control from diffusers import DDPMScheduler import paths from nonlinear.testing.test import load_checkpoint_cfg, tf_to_coeffs_vector, MAX_ORDER from utils.models.ConditionalDiffusion_mlp import ReverseDiffusionModel from utils.samplers import sample_diffusion dev="cuda" ckpt_p, ds_p = paths.nonlinear_default() ck=torch.load(ckpt_p, map_location=dev, weights_only=False) mcfg=load_checkpoint_cfg(ck["cfg"]); mcfg.device=dev model=ReverseDiffusionModel(mcfg).to(dev); model.load_state_dict(ck["model"]); model.eval() ns=DDPMScheduler(num_train_timesteps=mcfg.num_timesteps, beta_schedule="squaredcos_cap_v2") sc=ck["dataset_scalers"]; G_scale=float(np.asarray(sc["G_scale"]).reshape(-1)[0]) g=tf_to_coeffs_vector(control.tf([1.0],[1.0,1.0]), max_len=MAX_ORDER+1) cond=torch.tensor(np.asarray(g)/G_scale,dtype=torch.float32,device=dev).unsqueeze(0).repeat(8,1) torch.manual_seed(0); torch.cuda.synchronize(); t0=time.perf_counter() s=sample_diffusion(model,cond,(8,mcfg.output_dim),ns,device=dev,use_cfg=False) torch.cuda.synchronize() print("SAMPLED", tuple(s.shape), "in %.2fs for 8 (%.1f ms/ctrl)"%(time.perf_counter()-t0,(time.perf_counter()-t0)/8*1000)) print("finite:", bool(torch.isfinite(s).all()))