fix: ZONOS2 only generating 1024 tokens max

1 files changed, 11 insertions, 0 deletions

diff --git a/app/backends/sglomni/catalog.py b/app/backends/sglomni/catalog.py
index c89e865..5ff0c81 100644
--- a/app/backends/sglomni/catalog.py
+++ b/app/backends/sglomni/catalog.py
@@ -60,6 +60,16 @@ class ModelEntry:
     fp8_moe: bool = False
     fp8_min_compute_capability: Optional[Tuple[int, int]] = None
     bf16_config: Optional[str] = None
+    # max_new_tokens sent with every /v1/audio/speech request (None = send
+    # nothing and leave the server default in charge). Zonos2's AR engine
+    # defaults to 1024 audio frames (44100 Hz / 512 hop = 86.13 fps), which
+    # caps one request at ~12 s of speech and silently truncates longer
+    # text. 12288 frames (~143 s) covers a full 250-word CHUNK_SIZE
+    # sub-chunk (~8600 frames) and stays under the KV-pool admission check
+    # on every GPU that can host the model (all are >= 24 GB with
+    # >= 28083-token pools; the scheduler rejects prompt + max_new_tokens
+    # above it, and zonos2 requests are not auto-clamped).
+    max_new_tokens: Optional[int] = None
 
 
 # The Qwen3-TTS CustomVoice speaker table — the same built-in speakers the
@@ -197,6 +207,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         fp8_moe=True,
         fp8_min_compute_capability=(8, 9),
         bf16_config="zonos2_bf16.yaml",
+        max_new_tokens=12288,
     ),
 )