feat(parakeet): use bf16 for inference

handler.py

@@ -3,6 +3,7 @@ import zlib
 from functools import partial
 from io import BytesIO
 
+import torch
 from hfendpoints.openai import Context, run
 from hfendpoints.openai.audio import AutomaticSpeechRecognitionEndpoint, SegmentBuilder, Segment, \
     TranscriptionRequest, TranscriptionResponse, TranscriptionResponseKind, VerboseTranscription
@@ -40,6 +41,7 @@ class NemoAsrHandler(Handler):
     def __init__(self, config: EndpointConfig):
         logger.info(config.repository)
         self._model = ASRModel.from_pretrained(model_name=str(config.repository)).eval()
+        self._model = self._model.to(torch.bfloat16)
 
     async def __call__(self, request: TranscriptionRequest, ctx: Context) -> TranscriptionResponse:
         with logger.contextualize(request_id=ctx.request_id):