Update handler.py

handler.py

@@ -1,25 +1,26 @@
 # handler.py
-import io
-import torch
-from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline
 
 class EndpointHandler:
     def __init__(self, path=""):
-        # Load ASR model/tokenizer via official helper
+        # lazy import only inside init
+        from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline
+        
+        # Load the omniASR pipeline
         self.pipeline = ASRInferencePipeline(model_card="facebook/omniASR-LLM-7B")
 
     def __call__(self, data):
-        # data dict may contain bytes, multipart upload, etc.
-        # Here we assume audio bytes in "inputs"
+        # import here so it doesn’t trigger during module load
+        import io
+        import soundfile as sf
+
+        # read raw bytes
         audio_bytes = data.get("inputs")
         if not audio_bytes:
             return {"error": "no audio provided"}
 
-        # convert bytes to temp file, path etc.
-        import soundfile as sf
         f = io.BytesIO(audio_bytes)
         audio, sr = sf.read(f)
 
-        # Run transcription
+        # run transcription
         result = self.pipeline.transcribe([audio], batch_size=1)
         return {"text": result}