src/model.py

import os
from transformers import WhisperProcessor, WhisperForConditionalGeneration

MODEL_NAME = "tarteel-ai/whisper-tiny-ar-quran"

# Use /tmp for all writable storage
CACHE_DIR = "/tmp/hf_cache"
LOCAL_SAVE_DIR = "/tmp/whisper-tiny-ar-quran-local"

# Make sure dirs exist
os.makedirs(CACHE_DIR, exist_ok=True)
os.makedirs(LOCAL_SAVE_DIR, exist_ok=True)

os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
os.environ["HF_HOME"] = CACHE_DIR
os.environ["HF_HUB_CACHE"] = CACHE_DIR

_model = None
_processor = None

def load_model():
    global _model, _processor

    if _model is not None and _processor is not None:
        return _model, _processor

    local_config_file = os.path.join(LOCAL_SAVE_DIR, "preprocessor_config.json")

    if os.path.exists(local_config_file):
        _processor = WhisperProcessor.from_pretrained(LOCAL_SAVE_DIR)
        _model = WhisperForConditionalGeneration.from_pretrained(LOCAL_SAVE_DIR)
    else:
        _processor = WhisperProcessor.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
        _model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)

        _model.save_pretrained(LOCAL_SAVE_DIR)
        _processor.save_pretrained(LOCAL_SAVE_DIR)

    # Patch configs
    if not hasattr(_model.generation_config, "lang_to_id"):
        _model.generation_config.lang_to_id = {"arabic": 50361}
    if not hasattr(_model.generation_config, "task_to_id"):
        _model.generation_config.task_to_id = {"transcribe": 0}

    _model.generation_config.language = "arabic"
    _model.generation_config.task = "transcribe"

    print(f"✅ Model and processor loaded from: {LOCAL_SAVE_DIR}")
    return _model, _processor