norm description

app.py

@@ -140,6 +140,8 @@ def generate_song(lyric, description=None, prompt_audio=None, genre=None, cfg_co
         description = None
     elif description is not None and description != "":
         genre = None
+        if description[-1] != ".":
+            description = description + "."
 
     progress(0.0, "Start Generation")
     start = time.time()

codeclm/models/codeclm_gen.py

@@ -278,12 +278,12 @@ class CodecLM_gen:
                 if type_info is not None:
                     attr["text"]["type_info"] = type_info[i]
             conditions.append(attr)
-            print("conditions", conditions)
         if prepare_null_condition:
             cfg_inference = ClassifierFreeGuidanceDropoutInference() 
             null_conditions = cfg_inference(conditions, condition_types=["audio", "text"], 
                                             customized=None)
             conditions = conditions + null_conditions
+        print("conditions", conditions)
         tokenized_conditions = self.condition_provider.tokenize(conditions)
         # import pdb; pdb.set_trace()
         condition_tensors = self.condition_provider(tokenized_conditions)

levo_inference.py

@@ -48,7 +48,7 @@ class LeVoInference(torch.nn.Module):
             enforce_eager=True,
             dtype="bfloat16",  
             gpu_memory_utilization=0.65,
-            max_num_seqs=4,
+            max_num_seqs=8,
             tokenizer=None,
             skip_tokenizer_init=True,
             enable_prompt_embeds=True,
@@ -82,7 +82,7 @@ class LeVoInference(torch.nn.Module):
             pmt_wav = None
             melody_is_wav = True
 
-        description = description if description else '.'
+        description = description.lower() if description else '.'
         description = '[Musicality-very-high]' + ', ' + description
         generate_inp = {
             'descriptions': [lyric.replace("  ", " ")],
@@ -106,11 +106,8 @@ class LeVoInference(torch.nn.Module):
         )
         # 拆成现支持的batch 3 CFG形式
         prompts = [{"prompt_embeds": embed} for embed in fused_input]
-        promptss = []
-        for _ in range(2):
-            promptss+=prompts
-        uncondi = prompts[1]
-        promptss = promptss[::2] + [uncondi]
+        condi, uncondi = prompts[0], prompts[1]
+        promptss = [condi, condi, uncondi]
         outputs = self.llm.generate(promptss, sampling_params=sampling_params)
         token_ids_CFG = torch.tensor(outputs[1].outputs[0].token_ids)
         token_ids_CFG = token_ids_CFG[:-1].unsqueeze(0).unsqueeze(0)