handler.py

import torch
import soundfile as sf
from nemo.collections.speechlm2.models import SALM

class EndpointHandler():
    def __init__(self, path=""):
        # Load the model from Hugging Face directory (ASR + LLM model)
        self.model = SALM.from_pretrained(path)
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model.to(self.device)

    def __call__(self, data):
        mode = data.get("mode")
        
        if mode == "asr":
            # For ASR, we expect `audio` to be a list of floats or bytes
            audio = data.get("audio")  # base64 / list of floats / bytes

            if isinstance(audio, list):
                audio_tensor = torch.tensor(audio).unsqueeze(0).to(self.device)
            else:
                # fallback: assume bytes
                audio_data, sr = sf.read(audio)
                audio_tensor = torch.tensor(audio_data).unsqueeze(0).to(self.device)

            transcripts = self.model.transcribe(audio_tensor)
            return {"text": transcripts}

        elif mode == "llm":
            # For LLM, we expect "inputs" key
            prompt = data.get("inputs", "")
            outputs = self.model.generate([prompt])
            return {"text": outputs[0]}
        
        else:
            return {"error": "Please specify mode as 'asr' or 'llm'."}