Deploy HARP wrapper via model agent

.harp/manifest.json

@@ -0,0 +1,21 @@
+{
+  "backend_space": "REPLACE_WITH_YOUR_BACKEND_SPACE",
+  "deploy_mode": "remote-backend",
+  "entry": "app.py",
+  "framework": "gradio_client",
+  "generated": true,
+  "io": {
+    "inputs": [
+      "textbox",
+      "slider",
+      "number"
+    ],
+    "outputs": [
+      "audio"
+    ]
+  },
+  "repo_id": "SonyResearch/Woosh",
+  "source": "recipe",
+  "space_layout": "huggingface-gradio",
+  "task": "text-to-audio"
+}

README.md

@@ -1,13 +1,19 @@
 ---
-title: Woosh Dflow
-emoji: 🐠
-colorFrom: pink
-colorTo: blue
+title: "Woosh-DFlow (Text-to-Audio SFX)"
+colorFrom: indigo
+colorTo: gray
 sdk: gradio
-sdk_version: 6.19.0
-python_version: '3.13'
+sdk_version: 5.28.0
 app_file: app.py
 pinned: false
+license: "cc-by-nc-4.0"
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Woosh-DFlow (Text-to-Audio SFX)
+
+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).
+
+- Inputs: textbox, slider, number
+- Outputs: audio
+
+Generated by the HARP model agent from a recipe.

app.py

@@ -0,0 +1,91 @@
+from __future__ import annotations
+
+import os
+
+import gradio as gr
+
+from pyharp import *
+from gradio_client import Client, handle_file
+
+
+_BACKEND_SPACE = "REPLACE_WITH_YOUR_BACKEND_SPACE"
+_BACKEND_API_NAME = "/generate"
+_BACKEND_TOKEN_ENV = "HF_TOKEN"
+_ACCEPT_USER_TOKEN = False
+_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 _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):
+    _conn = _backend_client()
+    try:
+        _raw = _conn.predict(
+            prompt,
+            float(cfg_scale),
+            int(seed),
+            api_name="/generate",
+        )
+    except Exception as _exc:  # surface a token-aware hint, never the token
+        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)

requirements.txt

@@ -0,0 +1,3 @@
+git+https://github.com/TEAMuP-dev/pyharp.git@v0.3.0
+gradio>=4.0
+gradio_client