From 1c876f0cd105048fbf4f056700983ca75706e987 Mon Sep 17 00:00:00 2001 From: rein yang <73573651+R2-Y@users.noreply.github.com> Date: Thu, 28 May 2026 17:26:57 +0800 Subject: [PATCH] [Perf] Trim Code2Wav CUDA Graph buckets for Qwen3-TTS single-GPU deploy (#3932) Signed-off-by: Rein Yang --- vllm_omni/deploy/qwen3_tts.yaml | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm_omni/deploy/qwen3_tts.yaml b/vllm_omni/deploy/qwen3_tts.yaml index f54cca6ef..c39ab792e 100644 --- a/vllm_omni/deploy/qwen3_tts.yaml +++ b/vllm_omni/deploy/qwen3_tts.yaml @@ -28,6 +28,11 @@ connectors: codec_left_context_frames: 72 # Emit only the first audio chunk early, then return to codec_chunk_frames. initial_codec_chunk_frames: 1 + # Common Stage1 decode buckets: + # no-ref first/steady chunks: 25 / 97 frames + # Base ref-context first/steady chunks: 73 / 169 frames + # decoder internal non-streaming chunks: 325 frames + decode_cudagraph_capture_sizes: [25, 73, 97, 169, 325] stages: - stage_id: 0