app.py

from __future__ import annotations

import os
import time
import urllib.request

import gradio as gr

from pyharp import *
from gradio_client import Client, handle_file


_BACKEND_SPACE = "2cylu2/woosh"
_BACKEND_API_NAME = "/generate"
_BACKEND_TOKEN_ENV = "HF_TOKEN"
_ACCEPT_USER_TOKEN = False
# How many times to wake+retry a sleeping backend, and how long to wait for
# it to boot (a free Space cold start can take a few minutes).
_CALL_RETRIES = int(os.environ.get("BACKEND_CALL_RETRIES", "4"))
_WAKE_TIMEOUT = float(os.environ.get("BACKEND_WAKE_TIMEOUT", "420"))
_client = None


def _backend_client():
    # Lazily create and cache one warm connection using this Space's own
    # token (from the HF_TOKEN secret) or anonymous if none is set. User
    # tokens are NOT cached here -- they get a fresh per-call connection.
    global _client
    if _client is None:
        _token = os.environ.get(_BACKEND_TOKEN_ENV) or None
        _client = Client(_BACKEND_SPACE, hf_token=_token)
    return _client


def _reset_client():
    # Drop the cached connection so the next attempt reconnects to a Space
    # that has since finished waking.
    global _client
    _client = None


def _make_conn(tok):
    tok = (tok or '').strip()
    if tok:
        return Client(_BACKEND_SPACE, hf_token=tok)
    return _backend_client()


def _space_url(space):
    slug = space.strip().lower().replace('/', '-').replace('_', '-')
    return f'https://{slug}.hf.space/'


def _is_cold_start(message):
    # Errors that mean 'the backend was asleep/booting', worth waking+retrying
    # (vs. a real application error, which we surface immediately).
    _low = (message or '').lower()
    return any(s in _low for s in (
        'read operation timed out', 'timed out', 'timeout', 'starting',
        'building', 'not ready', 'no application', 'connection', '503', '502',
    ))


def _wake_backend():
    # A sleeping Space boots when its URL is hit; poll until it answers (or
    # the budget expires) so the retried call lands on a running backend.
    _url = _space_url(_BACKEND_SPACE)
    _deadline = time.time() + _WAKE_TIMEOUT
    _delay = 5.0
    while time.time() < _deadline:
        try:
            _req = urllib.request.Request(_url, headers={'User-Agent': 'harp-frontend'})
            with urllib.request.urlopen(_req, timeout=30) as _resp:
                if getattr(_resp, 'status', 200) < 500:
                    return True
        except Exception:
            pass
        time.sleep(_delay)
        _delay = min(_delay * 1.5, 30.0)
    return False


def _quota_hint(message):
    # Turn a backend ZeroGPU quota error into an actionable message.
    # NOTE: 'message' is the backend's error text; it never contains our token.
    _low = (message or "").lower()
    if "quota" in _low or "zerogpu" in _low:
        if _ACCEPT_USER_TOKEN:
            return (
                "The backend's ZeroGPU quota is exhausted for the identity making "
                "this call. Paste your own Hugging Face token in the token field "
                "(read scope) so usage is attributed to your account."
            )
        return (
            "The backend's ZeroGPU quota is exhausted. This Space's calls are "
            "anonymous unless an HF_TOKEN secret is set (Settings -> Variables "
            "and secrets); use a token from a PRO account or a ZeroGPU-enabled org."
        )
    return message or "Backend call failed."


model_card = ModelCard(
    name="Woosh-DFlow (Text-to-Audio SFX)",
    description="Generate a ~5s, 48kHz sound effect from a text prompt using Sony AI's Woosh-DFlow, the distilled (4-step) text-to-audio model from the Woosh sound-effect foundation model family. This is a thin HARP frontend that proxies to a Woosh-DFlow backend Space over its /generate API; the heavy model (Python 3.12, torch 2.8, Gradio 6) runs there, unmodified. Open weights are CC-BY-NC 4.0 (non-commercial).",
    author="Sony AI (Hadjeres, Ferras, Koutini, Weck, Bittar, Hummel, Lahrichi, Missoum, Serra, Mitsufuji)",
    tags=["text-to-audio", "sound-effects", "sfx", "generative-audio"],
)


def process_fn(prompt, cfg_scale, seed):
    _tok = ''
    # Call the backend, waking it and retrying if it was asleep (a cold
    # start otherwise fails the first hit with 'read operation timed out').
    _raw = None
    for _attempt in range(_CALL_RETRIES + 1):
        try:
            _conn = _make_conn(_tok)
            _raw = _conn.predict(
                prompt,
                float(cfg_scale),
                int(seed),
                api_name="/generate",
            )
            break
        except Exception as _exc:  # never surfaces the token
            if _attempt < _CALL_RETRIES and _is_cold_start(str(_exc)):
                _reset_client()
                _wake_backend()
                continue
            raise gr.Error(_quota_hint(str(_exc)))
    _values = list(_raw) if isinstance(_raw, (list, tuple)) else [_raw]
    _detail = " | ".join(str(_v) for _v in _values if isinstance(_v, str) and _v.strip())
    _out_audio = _values[0] if len(_values) > 0 else None
    if not _out_audio:
        raise gr.Error(_detail or "The backend Space returned no 'audio' output. Check the backend Space's logs; if it uses ZeroGPU it may need a moment to warm up.")
    return _out_audio


with gr.Blocks() as demo:
    input_components = [
        gr.Textbox(label="Prompt", info="Describe the sound effect to generate, e.g. 'sportscar engine revving and driving away quickly'."),
        gr.Slider(minimum=0.0, maximum=15.0, step=0.1, value=4.5, label="CFG scale", info="Classifier-free guidance strength: higher follows the prompt more closely."),
        gr.Number(value=-1, label="Seed", info="Random seed; use -1 for a new random result each run."),
    ]
    output_components = [
        gr.Audio(type="filepath", label="Generated sound effect"),
    ]
    build_endpoint(
        model_card=model_card,
        input_components=input_components,
        output_components=output_components,
        process_fn=process_fn,
    )

demo.queue().launch(share=True, show_error=False, pwa=True)