Retry transient VoxCPM accelerator errors

app.py

@@ -19,6 +19,7 @@ import asyncio
 import json as _json
 import os
 import threading
+import time
 
 # ZeroGPU requires the spaces shim to be imported before torch. Locally, or on
 # non-ZeroGPU hardware, this falls back to a no-op decorator.
@@ -412,24 +413,40 @@ async def qwen_tts(request: Request):
     return Response(wav, media_type="audio/wav", headers={"Cache-Control": "no-store"})
 
 
-def _voxcpm_tts(text, instruct):
+def _voxcpm_predict(api_name, *args):
     from gradio_client import Client
     client = Client(VOXCPM_SPACE, token=HF_TOKEN or None)
-    result = client.predict(text, instruct or "A clear, natural voice at a moderate pace.", api_name="/synthesize")
+    last_err = None
+    for attempt in range(3):
+        try:
+            return client.predict(*args, api_name=api_name)
+        except Exception as e:  # noqa: BLE001
+            last_err = e
+            msg = str(e).lower()
+            if attempt == 2 or not any(s in msg for s in ("accelerator", "queue", "gpu", "timeout", "temporarily")):
+                raise
+            time.sleep(1.5 * (attempt + 1))
+    raise last_err
+
+
+def _voxcpm_tts(text, instruct):
+    result = _voxcpm_predict(
+        "/synthesize",
+        text,
+        instruct or "A clear, natural voice at a moderate pace.",
+    )
     path = result[0] if isinstance(result, (tuple, list)) else result
     with open(os.fspath(path), "rb") as f:
         return f.read()
 
 
 def _voxcpm_clone(text, ref_audio_b64, ref_text, instruct):
-    from gradio_client import Client
-    client = Client(VOXCPM_SPACE, token=HF_TOKEN or None)
-    result = client.predict(
+    result = _voxcpm_predict(
+        "/clone",
         text,
         ref_audio_b64,
         ref_text or "",
         instruct or "",
-        api_name="/clone",
     )
     path = result[0] if isinstance(result, (tuple, list)) else result
     with open(os.fspath(path), "rb") as f: