RuntimeError: CUDA error: out of memory, but this maybe no the memory problem,

[kzl787]

when run python run_tokenflow_pnp.py Traceback (most recent call last): File "/mnt/disk_1/kaizhou/TokenFlow/run_tokenflow_pnp.py", line 300, in run(config) File "/mnt/disk_1/kaizhou/TokenFlow/run_tokenflow_pnp.py", line 279, in run editor = TokenFlow(config) File "/mnt/disk_1/kaizhou/TokenFlow/run_tokenflow_pnp.py", line 62, in init self.paths, self.frames, self.latents, self.eps = self.get_data() File "/mnt/disk_1/kaizhou/TokenFlow/run_tokenflow_pnp.py", line 183, in get_data eps = self.get_ddim_eps(latents, range(self.config["n_frames"])).to(torch.float16).to(self.device) File "/mnt/disk_1/kaizhou/TokenFlow/run_tokenflow_pnp.py", line 189, in get_ddim_eps noisy_latent = torch.load(latents_path)[indices].to(self.device)

the config_pnp.yml's conten is follow(and the cuda:5 is unoccupied:

seed: 1 device: 'cuda:5' output_path: 'tokenflow-results'

data_path: 'data/wolf' latents_path: 'latents' # should be the same as 'save_dir' arg used in preprocess n_inversion_steps: 500 # for retrieving the latents of the inversion n_frames: 40

sd_version: '1.5' guidance_scale: 7.5 n_timesteps: 50 prompt: "A robotic wolf" negative_prompt: "ugly, blurry, low res, unrealistic, unaesthetic" batch_size: 1

pnp_attn_t: 0.5 pnp_f_t: 0.8