feat:打印precision和修改chunk_length

fish_speech/models/text2semantic/inference.py

@@ -362,6 +362,8 @@ def generate(
 def init_model(checkpoint_path, device, precision, compile=False):
     model = DualARTransformer.from_pretrained(checkpoint_path, load_weights=True)
 
+    logger.info(f"precision: {precision.__class__.__name__}")
+
     model = model.to(device=device, dtype=precision)
     logger.info(f"Restored model from checkpoint")
 

tools/server/model_manager.py

@@ -83,7 +83,8 @@ class ModelManager:
             references=[],
             reference_id=None,
             max_new_tokens=1024,
-            chunk_length=200,
+            # chunk_length=200,
+            chunk_length=150,
             top_p=0.7,
             repetition_penalty=1.2,
             temperature=0.7,