Update app.py

app.py

@@ -10,18 +10,13 @@ MODEL_NAME = 'sarvamai/shuka_v1'
 SAMPLE_RATE = 16000
 MAX_NEW_TOKENS = 256
 
-# Load the ShukaPipeline
-def load_pipeline():
-    model = transformers.AutoModel.from_pretrained(MODEL_NAME, trust_remote_code=True)
-    pipeline = transformers.pipeline(
-        "shuka-pipeline",
-        model=model,
-        torch_dtype=torch.float16,
-        device=0 if torch.cuda.is_available() else -1,
-    )
-    return pipeline
-
-pipe = load_pipeline()
+# Load the pipeline
+pipe = transformers.pipeline(
+    model=MODEL_NAME,
+    trust_remote_code=True,
+    device=0,
+    torch_dtype='bfloat16'
+)
 
 def create_conversation_turns(prompt: str) -> List[Dict[str, str]]:
     return [
@@ -45,7 +40,7 @@ def transcribe_and_respond(audio: np.ndarray) -> str:
         }
         
         # Generate response
-        response = pipe(inputs, max_new_tokens=MAX_NEW_TOKENS, temperature=0.7, repetition_penalty=1.1)
+        response = pipe(inputs, max_new_tokens=MAX_NEW_TOKENS)
         
         return response
     except Exception as e: