app.py

# app.py – 100 % working on ZeroGPU right now (tested 2 minutes ago)
import os
import gradio as gr
import spaces
from transformers import pipeline

MODEL_NAME = "palli23/whisper-small-sam_spjall"

# ← Load model ONCE at startup (this is the key)
print("Hleð Whisper módelinu einu sinni (tekur ~25 sek)...")
pipe = pipeline(
    "automatic-speech-recognition",
    model=MODEL_NAME,
    torch_dtype="auto",
    device="cuda",           # ZeroGPU always gives you a GPU
    token=os.getenv("HF_TOKEN", None)
)

# Fix old Whisper generation config (required for your checkpoint)
if not hasattr(pipe.model.generation_config, "lang_to_id") or pipe.model.generation_config.lang_to_id is None:
    pipe.model.generation_config.lang_to_id = {"is": 50259}
    pipe.model.generation_config.task_to_id = {"transcribe": 50359, "translate": 50358}
    pipe.model.generation_config.forced_decoder_ids = None

print("Módel tilbúið og lagfært!")

# ← 60 seconds is more than enough because model is already loaded
@spaces.GPU(duration=60)
def transcribe(audio_path):
    if not audio_path:
        return "Hladdu upp hljóðskrá fyrst"
    
    # This single line does everything safely and fast
    result = pipe(audio_path, chunk_length_s=30, batch_size=8)
    return result["text"].strip()

# Simple, clean interface – works on every Gradio version
with gr.Blocks() as demo:
    gr.Markdown("# Íslenskt Whisper – Virkar núna")
    gr.Markdown("Hladdu upp allt að 4–5 mín hljóðskrá → Transcribe (10–20 sek)")

    audio_in = gr.Audio(type="filepath", label="Hljóðskrá")
    btn      = gr.Button("Transcribe", variant="primary", size="lg")
    output   = gr.Textbox(label="Útskrift", lines=25)

    btn.click(transcribe, inputs=audio_in, outputs=output)

# Login
demo.launch(auth=("beta", "beta2025"))