gpt comments

studio/backend/core/inference/audio_codecs.py

@@ -311,10 +311,16 @@ class AudioCodecManager:
     ) -> Tuple[bytes, int]:
         """Unified decode — dispatches to the right codec decoder."""
         if audio_type == "snac":
+            if not token_ids:
+                raise ValueError("SNAC decoding requires token_ids")
             return self.decode_snac(torch.tensor([token_ids], dtype = torch.long), device)
         elif audio_type == "bicodec":
+            if not text:
+                raise ValueError("BiCodec decoding requires text")
             return self.decode_bicodec(text, device)
         elif audio_type == "dac":
+            if not text:
+                raise ValueError("DAC decoding requires text")
             return self.decode_dac(text, device)
         raise ValueError(f"Cannot decode audio_type: {audio_type}")
 

studio/backend/core/inference/llama_cpp.py

@@ -732,8 +732,17 @@ class LlamaCppBackend:
             self._hf_repo = None
             self._hf_variant = None
             self._is_vision = False
+            self._is_audio = False
+            self._audio_type = None
             self._port = None
             self._healthy = False
+            # Free audio codec GPU memory
+            if LlamaCppBackend._codec_mgr is not None:
+                LlamaCppBackend._codec_mgr.unload()
+                LlamaCppBackend._codec_mgr = None
+                import torch
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
             return True
 
     def _kill_process(self):
@@ -1079,6 +1088,7 @@ class LlamaCppBackend:
         temperature: float = 0.6,
         top_p: float = 0.95,
         top_k: int = 50,
+        min_p: float = 0.0,
         max_new_tokens: int = 2048,
         repetition_penalty: float = 1.1,
     ) -> tuple:
@@ -1098,6 +1108,7 @@ class LlamaCppBackend:
             "temperature": temperature,
             "top_p": top_p,
             "top_k": top_k if top_k >= 0 else 0,
+            "min_p": min_p,
             "repeat_penalty": repetition_penalty,
         }
         if stop:

studio/backend/routes/inference.py

@@ -159,9 +159,7 @@ async def load_model(
             from utils.models import is_audio_input_type
 
             _gguf_audio = llama_backend.detect_audio_type()
-            _gguf_is_audio = _gguf_audio is not None and _gguf_audio not in (
-                "audio_vlm",
-            )
+            _gguf_is_audio = _gguf_audio in ("snac", "bicodec", "dac")
             llama_backend._is_audio = _gguf_is_audio
             llama_backend._audio_type = _gguf_audio
             if _gguf_is_audio:
@@ -547,11 +545,9 @@ async def generate_audio(
     if llama_backend.is_loaded and getattr(llama_backend, "_is_audio", False):
         model_name = llama_backend.model_identifier
         gen = lambda: llama_backend.generate_audio_response(
-            text = text,
-            audio_type = llama_backend._audio_type,
-            temperature = payload.temperature,
-            top_p = payload.top_p,
-            top_k = payload.top_k,
+            text = text, audio_type = llama_backend._audio_type,
+            temperature = payload.temperature, top_p = payload.top_p,
+            top_k = payload.top_k, min_p = payload.min_p,
             max_new_tokens = payload.max_tokens or 2048,
             repetition_penalty = payload.repetition_penalty,
         )