n_draws

audiocraft.py

@@ -70,7 +70,7 @@ class AudioGen(torch.nn.Module):
         with torch.autocast(device_type='cpu', dtype=torch.bfloat16):
             gen_tokens = self.lm.generate(
                 text_condition=[prompt] * N_REPEAT  + [''] * N_REPEAT,#['dogs', 'dogs...!', '', '']
-                max_tokens=int(.04 * duration / N_REPEAT * self.compression_model.frame_rate) + 12) # [bs, 4, 74*self.lm.n_draw]
+                max_tokens=max(int(duration / (N_REPEAT * self.lm.n_draw) * 50) + 5, 12))
 
         # OOM if vocode all tokens
         x = []