app.py

# app.py
import os
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"

import gradio as gr
import spaces
from transformers import pipeline

# ——————————————————————————————
# Model loaded ONCE at startup (global)
# ——————————————————————————————
MODEL_NAME = "palli23/whisper-large-v3-is-samromur-20-40s-3x-final"

@spaces.GPU(duration=180)
def get_pipe():
    return pipeline(
        "automatic-speech-recognition",
        model=MODEL_NAME,
        torch_dtype="float16",
        device=0,                       # T4 GPU
        token=os.getenv("HF_TOKEN"),    # Remove line if model is public
    )

pipe = get_pipe()  # ← Loaded once when Space starts


# ——————————————————————————————
# Transcription function (super fast now)
# ——————————————————————————————
def transcribe_3min(audio_path):
    if not audio_path:
        return "Hladdu upp hljóðskrá"
    
    result = pipe(
        audio_path,
        chunk_length_s=30,
        stride_length_s=(6, 0),
        batch_size=8,
        return_timestamps=False,
    )
    return result["text"]


# ——————————————————————————————
# Gradio UI
# ——————————————————————————————
with gr.Blocks(title="Íslenskt ASR – 3 mín") as demo:
    gr.Markdown("# Íslenskt ASR – 3 mínútur")
    gr.Markdown("**Whisper · mjög lágur WER · allt að 5 mín hljóð**")
    
    audio_in = gr.Audio(
        type="filepath",
        label="Hladdu upp .mp3 / .wav (max 5 mín)"
    )
    btn = gr.Button("Transcribe", variant="primary", size="lg")
    output = gr.Textbox(lines=30, label="Útskrift")

    btn.click(fn=transcribe_3min, inputs=audio_in, outputs=output)


# ——————————————————————————————
# Stable launch (no more infinite Building)
# ——————————————————————————————
demo.launch(
    auth=("beta", "beta2025"),
    ssr_mode=False,
    show_error=True,
    server_name="0.0.0.0",
    server_port=7860,
    quiet=False
)