Deploy HARP wrapper via model agent

.harp/manifest.json

@@ -0,0 +1,26 @@
+{
+  "backend_space": "ASLP-lab/DiffRhythm2",
+  "deploy_mode": "remote-backend",
+  "entry": "app.py",
+  "framework": "gradio_client",
+  "generated": true,
+  "io": {
+    "inputs": [
+      "textbox",
+      "audio",
+      "textbox",
+      "slider",
+      "checkbox",
+      "slider",
+      "slider",
+      "dropdown"
+    ],
+    "outputs": [
+      "audio"
+    ]
+  },
+  "repo_id": "ASLP-lab/DiffRhythm2",
+  "source": "recipe",
+  "space_layout": "huggingface-gradio",
+  "task": "custom"
+}

README.md

@@ -1,13 +1,19 @@
 ---
-title: DiffRhythm2
-emoji: 🐢
-colorFrom: purple
-colorTo: green
+title: "Diffrhythm2"
+colorFrom: indigo
+colorTo: gray
 sdk: gradio
-sdk_version: 6.20.0
-python_version: '3.13'
+sdk_version: 5.28.0
 app_file: app.py
 pinned: false
+license: "other"
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Diffrhythm2
+
+TODO: describe this model.
+
+- Inputs: textbox, audio, textbox, slider, checkbox, slider, slider, dropdown
+- Outputs: audio
+
+Generated by the HARP model agent from a recipe.

app.py

@@ -0,0 +1,187 @@
+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 = "ASLP-lab/DiffRhythm2"
+_BACKEND_API_NAME = "/infer_music"
+_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 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."
+        )
+    # Opaque backend failure: the Space raised an exception it refuses to
+    # expose (it runs with show_error=False), so all we get is a generic
+    # 'Internal Gradio error'. The frontend can't fix a server-side crash --
+    # point the user at where the real cause lives.
+    if (
+        "internal gradio error" in _low
+        or "internal server error" in _low
+        or "apperror" in _low
+        or _low.strip() in ("", "none")
+    ):
+        _hint = (
+            "The backend Space raised an error it did not expose (it runs with "
+            "show_error disabled), so the real cause is only in the backend "
+            "Space's Logs tab (" + _BACKEND_SPACE + ")."
+        )
+        if _ACCEPT_USER_TOKEN:
+            _hint += (
+                " If it is a ZeroGPU Space, an anonymous call can fail this way -- "
+                "paste a Hugging Face token in the token field and retry."
+            )
+        return _hint
+    return message or "Backend call failed."
+
+
+model_card = ModelCard(
+    name="Diffrhythm2",
+    description="TODO: describe this model.",
+    author="ASLP-lab",
+    tags=[],
+)
+
+
+def process_fn(lrc, audio_prompt, text_prompt, seed, randomize_seed, steps, cfg_strength, file_type):
+    _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(
+                lrc,
+                (handle_file(audio_prompt) if audio_prompt else None),
+                text_prompt,
+                seed,
+                randomize_seed,
+                steps,
+                cfg_strength,
+                file_type,
+                'euler',
+                api_name="/infer_music",
+            )
+            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_result = _values[0] if len(_values) > 0 else None
+    if not _out_audio_result:
+        raise gr.Error(_detail or "The backend Space returned no 'audio_result' output. Check the backend Space's logs; if it uses ZeroGPU it may need a moment to warm up.")
+    return _out_audio_result
+
+
+with gr.Blocks() as demo:
+    input_components = [
+        gr.Textbox(label="Lyrics", value="[start]\n[intro]\n[verse]\nThought I heard your voice yesterday\nWhen I turned around to say\nThat I loved you baby\nI realize it was juss my mind\nPlayed tricks on me\nAnd it seems colder lately at night\nAnd I try to sleep with the lights on\nEvery time the phone rings\nI pray to God it's you\nAnd I just can't believe\nThat we're through\n[chorus]\nI miss you\nThere's no other way to say it\nAnd I can't deny it\nI miss you\nIt's so easy to see\nI miss you and me\n[verse]\nIs it turning over this time\nHave we really changed our minds about each other's love\nAll the feelings that we used to share\nI refuse to believe\nThat you don't care\n[chorus]\nI miss you\nThere's no other way to say it\nAnd I and I can't deny it\nI miss you\n[verse]\nIt's so easy to see\nI've got to gather myself as together\nI've been through worst kinds of weather\nIf it's over now\n[outro]"),
+        gr.Audio(type="filepath", label="Audio Prompt"),
+        gr.Textbox(label="Text Prompt", value="Pop, Piano, Bass, Drums, Happy"),
+        gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0, label="Seed"),
+        gr.Checkbox(value=True, label="Randomize seed"),
+        gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=16, label="Diffusion Steps"),
+        gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1.3, label="CFG Strength"),
+        gr.Dropdown(choices=["TODO_option_1", "TODO_option_2"], value="mp3", label="Output Format"),
+    ]
+    output_components = [
+        gr.Audio(type="filepath", label="Audio Result"),
+    ]
+    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@develop
+gradio>=4.0
+gradio_client