add inference mode for lm generate

Signed-off-by: weedge <weege007@gmail.com>

kimia_infer/api/kimia.py

@@ -47,6 +47,7 @@ class KimiAudio(object):
         self.kimia_text_audiodelaytokens = 6
         self.eod_ids = [self.extra_tokens.msg_end, self.extra_tokens.media_end]
 
+    @torch.inference_mode()
     def _generate_loop(
         self,
         audio_input_ids: torch.Tensor,  # input audio tokens
@@ -204,6 +205,7 @@ class KimiAudio(object):
         )
         return return_audio_tokens, return_text_tokens
 
+    @torch.inference_mode()
     def generate(
         self,
         chats: list[dict],