Param issues

fastapi_app.py

@@ -87,36 +87,32 @@ def get_model(model_name):
                 logging.info(f"Converting Xenova model {model_name} to {hf_model_name}")
             
             try:
-                # Load the model using TensorFlow compatibility mode
-                import tensorflow as tf
-                logging.info(f"TensorFlow version: {tf.__version__}")
-                
-                # Configure pipeline with correct parameters to avoid conflicts
+                # Load the model using a simpler approach without problematic parameters
                 models[model_name] = pipeline(
                     "automatic-speech-recognition", 
                     model=hf_model_name,
                     chunk_length_s=30,
-                    stride_length_s=5,
-                    framework="tf",  # Explicitly use TensorFlow
-                    model_kwargs={
-                        "attention_mask": True,  # Explicitly set attention mask
-                        "use_cache": True,
-                    }
+                    stride_length_s=5
                 )
-                logging.info(f"Model {model_name} loaded successfully with TensorFlow")
+                logging.info(f"Model {model_name} loaded successfully with default settings")
             except Exception as e:
-                logging.error(f"Error loading model with TensorFlow: {str(e)}")
+                logging.error(f"Error loading model with default settings: {str(e)}")
                 try:
-                    # Fallback to PyTorch with more specific configurations
+                    # Fallback to explicit model loading if necessary
                     from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
+                    import torch
+                    
+                    # Check device
+                    device = "cuda" if torch.cuda.is_available() else "cpu"
+                    logging.info(f"Using device: {device}")
                     
-                    # First load processor and model separately to customize configs
+                    # First load processor and model separately
                     processor = AutoProcessor.from_pretrained(hf_model_name)
                     model = AutoModelForSpeechSeq2Seq.from_pretrained(
                         hf_model_name,
                         use_cache=True,
-                        attention_mask=True
-                    )
+                        low_cpu_mem_usage=True
+                    ).to(device)
                     
                     # Create the pipeline with the initialized model and processor
                     models[model_name] = pipeline(
@@ -126,11 +122,7 @@ def get_model(model_name):
                         feature_extractor=processor,
                         chunk_length_s=30,
                         stride_length_s=5,
-                        generate_kwargs={
-                            "task": "transcribe",
-                            "language": "english",
-                            # Don't set forced_decoder_ids here to avoid conflict
-                        }
+                        device=device
                     )
                     logging.info(f"Model {model_name} loaded successfully with PyTorch")
                 except Exception as e2:
@@ -156,14 +148,10 @@ def process_audio(audio_data, sample_rate=16000, model_name="openai/whisper-base
             
             logging.info(f"Processing chunk {i+1}/{total_chunks}")
             
-            # Use consistent parameters for model inference
+            # Use simpler parameters for model inference to avoid errors
             result = model(
                 chunk,
-                return_timestamps=True,
-                generate_kwargs={
-                    "task": "transcribe",
-                    "language": "english"
-                }
+                return_timestamps=True
             )
             
             if "chunks" in result and len(result["chunks"]) > 0: