update

inspiremusic/llm/llm.py

@@ -365,7 +365,7 @@ class LLM(torch.nn.Module):
                 lm_input = torch.cat([lm_input, lm_cf_input], 0)
 
         # 4. cal min/max_length
-        min_len = duration_to_gen * token_rate
+        min_len = 0.9 * duration_to_gen * token_rate
         max_len = duration_to_gen * token_rate
         logging.info(
             f"LLM generation sequence length: {max_len}, generate audio length {duration_to_gen}s.")
@@ -388,6 +388,9 @@ class LLM(torch.nn.Module):
             logp = logits.log_softmax(dim=-1)
             logp = logp.squeeze(dim=0)
 
+            if i < int(min_len):
+                logp[self.audio_token_size] = torch.tensor(float('-inf'), dtype=torch.float16)
+
             if i < int(min_len):
                 logp[self.audio_token_size] = torch.tensor(float('-inf'), dtype=torch.float16)