obadx / muaalem-v3_2-torchscript

Muaalem Model – TorchScript versions

This repository contains the exported TorchScript versions of the obadx/muaalem-model-v3_2 model.

Files

• model_fp32.pt – full precision (float32)
• model_fp16.pt – half precision (float16)
• model_bf16.pt – bfloat16 precision

Model Output

The model returns a tuple whose first element is a dictionary of logits for each CTC head. The dictionary keys correspond to the heads defined in the configuration: ['ghonna', 'hams_or_jahr', 'istitala', 'itbaq', 'phonemes', 'qalqla', 'safeer', 'shidda_or_rakhawa', 'tafashie', 'tafkheem_or_taqeeq', 'tikraar']

For example, to get phoneme logits: output[0]['phonemes'] .

Loading the model in your code

Automatic precision selection based on GPU

import torch
from huggingface_hub import hf_hub_download

def load_optimized_model(repo_id="obadx/muaalem-v3_2-torchscript"):
    if torch.cuda.is_available():
        major, minor = torch.cuda.get_device_capability()
        if major >= 8:  # Ampere or newer
            try:
                model_file = hf_hub_download(repo_id=repo_id, filename="model_bf16.pt")
                dtype = torch.bfloat16
            except:
                model_file = hf_hub_download(repo_id=repo_id, filename="model_fp16.pt")
                dtype = torch.float16
        else:
            model_file = hf_hub_download(repo_id=repo_id, filename="model_fp16.pt")
            dtype = torch.float16
    else:
        model_file = hf_hub_download(repo_id=repo_id, filename="model_fp32.pt")
        dtype = torch.float32

    model = torch.jit.load(model_file)
    model = model.to(dtype)
    model.eval()
    return model, dtype

# Load processor from the same repo
from transformers import AutoFeatureExtractor
processor = AutoFeatureExtractor.from_pretrained(repo_id, subfolder="processor")
model, dtype = load_optimized_model()
print(f"Loaded model with dtype: {dtype}")