.

AR/models/t2s_model_abc.py

@@ -449,7 +449,9 @@ class CUDAGraphCacheABC(ABC):
     def assign_graph(self, session: Any):
         if self.graph is None:
             args, kwds = self.decoder.pre_forward(session)
-            graph = self.decoder.capture(self.input_pos, self.xy_pos, self.xy_dec, *args, **kwds)
+            graph = self.decoder.capture(
+                self.input_pos, self.xy_pos, self.xy_dec, kv_caches=self.kv_cache, *args, **kwds
+            )
             self.graph = graph
 
         if self.assigned is False:

AR/models/t2s_model_flash_attn.py

@@ -239,7 +239,7 @@ class CUDAGraphCache(CUDAGraphCacheABC):
         session.input_pos = self.input_pos.clone().copy_(session.input_pos)
 
         args, kwds = self.decoder.pre_forward(session)
-        graph = self.decoder.capture(self.input_pos, self.xy_pos, self.xy_dec, *args, **kwds)
+        graph = self.decoder.capture(self.input_pos, self.xy_pos, self.xy_dec, kv_caches=self.kv_cache, *args, **kwds)
         session.graph = graph
 
 

inference_webui.py

@@ -38,6 +38,7 @@ logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
 logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
 logging.getLogger("python_multipart.multipart").setLevel(logging.ERROR)
 logging.getLogger("split_lang.split.splitter").setLevel(logging.ERROR)
+logging.getLogger("filelock").setLevel(logging.INFO)
 
 os.makedirs("pretrained_models", exist_ok=True)