handler.py

from typing import Dict, List, Any
from transformers import pipeline
from transformers import (
    AutomaticSpeechRecognitionPipeline,
    WhisperForConditionalGeneration,
    WhisperTokenizer,
    WhisperProcessor,
)
# import faster_whisper
# import json
import logging
from peft import PeftModel, PeftConfig
import torch

logger = logging.getLogger(__name__)

class EndpointHandler():
    def __init__(self, path=""):
        peft_model_id = "Awaz-e-Sehat/whisper-fine-tune-new-LoRA" # Use the same model ID as before.
        peft_config = PeftConfig.from_pretrained(peft_model_id)
        model = WhisperForConditionalGeneration.from_pretrained(peft_config.base_model_name_or_path,load_in_8bit=True,device_map="auto")
        # self.model = faster_whisper.WhisperModel(path, device = "cuda")
        language = "Urdu"
        task = "transcribe"
        model = PeftModel.from_pretrained(model, peft_model_id)
        tokenizer = WhisperTokenizer.from_pretrained(peft_config.base_model_name_or_path, language=language, task=task)
        processor = WhisperProcessor.from_pretrained(peft_config.base_model_name_or_path, language=language, task=task)
        feature_extractor = processor.feature_extractor
        self.forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task=task)
        self.pipe = AutomaticSpeechRecognitionPipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor, chunk_length_s = 30, stride_length_s = 5)
        logger.info("Model Initialized")

    def __call__(self, data: Any) -> str:
        """
       data args:
            inputs (:obj: `str`)
            date (:obj: `str`)
      Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        # get inputs
        logger.info("In inference")
        logger.info(data)
        inputs = data.pop("inputs",data)
        logger.info("Data pop")
        logger.info(inputs)
        # segments, _ = self.model.transcribe(inputs, language = "ur", task = "transcribe")
        # logger.info("model transcribe")
        # segments = list(segments)
        # logger.info("Actual transcribed")
        # prediction = ''
        # for i in segments:
        #     prediction += i[4]
        # return prediction
        
        with torch.cuda.amp.autocast():
            text = self.pipe(inputs, generate_kwargs={"forced_decoder_ids": self.forced_decoder_ids}, max_new_tokens=255)["text"]
        return text