Throttle Gemini TTS podcast generation

.gitattributes

@@ -1,3 +1,5 @@
+/.gitattributes text eol=lf
+*.sh text eol=lf
 *.7z filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text

start-hf.sh

@@ -20,6 +20,9 @@ if [ -n "${VERTEX_SERVICE_ACCOUNT_JSON_B64:-}" ]; then
   export GOOGLE_APPLICATION_CREDENTIALS=/data/open-notebook/secrets/vertex-sa.json
 fi
 
+export TTS_BATCH_SIZE="${TTS_BATCH_SIZE:-1}"
+export TTS_BATCH_DELAY_SECONDS="${TTS_BATCH_DELAY_SECONDS:-65}"
+
 /app/.venv/bin/python - <<'PY'
 from pathlib import Path
 import re
@@ -277,6 +280,42 @@ if "_is_gemini_tts_model" not in src:
 print(f"Patched Vertex TTS provider: {target}")
 PY
 
+/app/.venv/bin/python - <<'PY'
+from pathlib import Path
+import re
+
+import podcast_creator
+
+target = Path(podcast_creator.__file__).resolve().parent / "nodes.py"
+src = target.read_text(encoding="utf-8")
+
+src, batch_count = re.subn(
+    r'batch_size = int\(os\.getenv\("TTS_BATCH_SIZE", "5"\)\)',
+    'batch_size = int(os.getenv("TTS_BATCH_SIZE", "1"))',
+    src,
+    count=1,
+)
+
+if "TTS_BATCH_DELAY_SECONDS" not in src:
+    src, delay_count = re.subn(
+        r"        # Small delay between batches to be extra safe with API limits\n"
+        r"        if batch_end < total_segments:\n"
+        r"            await asyncio\.sleep\(1\)",
+        "        # Delay between batches to respect Gemini-TTS per-minute quotas.\n"
+        "        if batch_end < total_segments:\n"
+        '            await asyncio.sleep(float(os.getenv("TTS_BATCH_DELAY_SECONDS", "65")))',
+        src,
+        count=1,
+    )
+    if delay_count != 1:
+        raise RuntimeError("Could not patch podcast_creator TTS batch delay; source layout changed")
+
+if batch_count > 0 or "TTS_BATCH_DELAY_SECONDS" in src:
+    target.write_text(src, encoding="utf-8")
+
+print(f"Patched podcast_creator TTS rate limiting: {target}")
+PY
+
 (
 /app/.venv/bin/python - <<'PY'
 import asyncio