From 2c157fb670598cf5a88e74304a9a93aef861569e Mon Sep 17 00:00:00 2001
From: Sy03 <1370724210@qq.com>
Date: Tue, 25 Aug 2026 02:34:57 +0800
Subject: [PATCH] [Perf][Qwen3-TTS] Enable chunk ramp by default

Signed-off-by: Sy03 <1370724210@qq.com>
---
 tests/config/test_config_factory.py              | 12 ++++++++++++
 vllm_omni/deploy/qwen3_tts.yaml                  | 13 +++++++------
 vllm_omni/deploy/qwen3_tts_high_concurrency.yaml | 13 +++++++------
 3 files changed, 26 insertions(+), 12 deletions(-)

diff --git a/tests/config/test_config_factory.py b/tests/config/test_config_factory.py
index 374a6d212..1c69b21ad 100644
--- a/tests/config/test_config_factory.py
+++ b/tests/config/test_config_factory.py
@@ -1854,6 +1854,18 @@ class TestQwen3TTSPipeline:
         # Stage 1 uses its per-stage override
         assert stages[1].yaml_engine_args["model_arch"] == "Qwen3TTSCode2Wav"
 
+    @pytest.mark.parametrize(
+        "deploy_name",
+        ["qwen3_tts.yaml", "qwen3_tts_high_concurrency.yaml"],
+    )
+    def test_default_async_deploy_enables_chunk_ramp(self, deploy_name):
+        deploy = load_deploy_config(Path(get_deploy_config_path(deploy_name)))
+
+        assert deploy.async_chunk is True
+        assert deploy.connectors is not None
+        extra = deploy.connectors["connector_of_shared_memory"]["extra"]
+        assert extra["codec_chunk_ramp"] == [4, 4, 8, 16, 25]
+
     def test_subtalker_sampling_params_deep_merge_preserves_base_keys(self):
         """Verify subtalker sampling params participate in stage deep-merge."""
         base = {
diff --git a/vllm_omni/deploy/qwen3_tts.yaml b/vllm_omni/deploy/qwen3_tts.yaml
index 58b191f77..cd0f7c480 100644
--- a/vllm_omni/deploy/qwen3_tts.yaml
+++ b/vllm_omni/deploy/qwen3_tts.yaml
@@ -26,13 +26,14 @@ connectors:
       codec_left_context_frames: 72
       # Emit only the first audio chunk early, then return to codec_chunk_frames.
       initial_codec_chunk_frames: 1
-      # Chunk ramp (opt-in): gradual size increase for first N chunks to
-      # eliminate the IC→steady cliff (1→25). Uncomment to enable.
+      # Gradual size increase for the first N chunks eliminates the
+      # IC-to-steady cliff (1->25) and improves streaming continuity.
       # Each entry is the size of that chunk index; indices past the table
-      # fall back to codec_chunk_frames. When enabled, ramp overrides
-      # initial_codec_chunk_frames (including the per-request API field).
-      # Example:
-      #   codec_chunk_ramp: [4, 4, 8, 16, 25]
+      # fall back to codec_chunk_frames. Ramp overrides
+      # initial_codec_chunk_frames, including the per-request API field.
+      # Set this to null in a deployment overlay to restore the legacy
+      # initial_codec_chunk_frames behavior.
+      codec_chunk_ramp: [4, 4, 8, 16, 25]
       # Decoder CUDA Graph sizing:
       #   * async_chunk=true derives all stateful ICL/xvec shapes from
       #     codec_chunk_ramp when enabled, otherwise from
diff --git a/vllm_omni/deploy/qwen3_tts_high_concurrency.yaml b/vllm_omni/deploy/qwen3_tts_high_concurrency.yaml
index ff086beb5..8129df166 100644
--- a/vllm_omni/deploy/qwen3_tts_high_concurrency.yaml
+++ b/vllm_omni/deploy/qwen3_tts_high_concurrency.yaml
@@ -31,13 +31,14 @@ connectors:
       ref_code_context_frames: 72
       # Emit only the first audio chunk early, then return to codec_chunk_frames.
       initial_codec_chunk_frames: 1
-      # Chunk ramp (opt-in): gradual size increase for first N chunks to
-      # eliminate the IC→steady cliff (1→25). Uncomment to enable.
+      # Gradual size increase for the first N chunks eliminates the
+      # IC-to-steady cliff (1->25) and improves streaming continuity.
       # Each entry is the size of that chunk index; indices past the table
-      # fall back to codec_chunk_frames. When enabled, ramp overrides
-      # initial_codec_chunk_frames (including the per-request API field).
-      # Example:
-      #   codec_chunk_ramp: [4, 4, 8, 16, 25]
+      # fall back to codec_chunk_frames. Ramp overrides
+      # initial_codec_chunk_frames, including the per-request API field.
+      # Set this to null in a deployment overlay to restore the legacy
+      # initial_codec_chunk_frames behavior.
+      codec_chunk_ramp: [4, 4, 8, 16, 25]
       # async_chunk derives stateful ICL/xvec graph shapes from the ramp when
       # enabled, otherwise from the fixed chunk settings above, and ignores
       # decode_cudagraph_capture_sizes. Keep B>1