fix cuda bug

configs/inference/inference.yaml

@@ -101,7 +101,7 @@ enable_zero_snr: True
 stage1_ckpt_dir: "./exp_output/stage1/"
 
 single_inference_times: 10
-inference_steps: 40
+inference_steps: 20 # improve inference speed
 cfg_scale: 3.5
 
 seed: 42

scripts/inference.py

@@ -347,7 +347,8 @@ def inference(cfg, image_processor, audio_processor, pipeline, audioproj, save_d
 
     times = audio_emb.shape[0] // clip_length
     tensor_result = []
-    generator = torch.manual_seed(42)
+    # generator = torch.manual_seed(42)
+    generator = torch.cuda.manual_seed_all(42) # use cuda seed all
     for t in range(times):
         print(f"[{t+1}/{times}]")