From cf328eb8e8ebd913d889a9f1142ee2cb801e77d2 Mon Sep 17 00:00:00 2001
From: Juan Pablo Zuluaga <46724788+JuanPZuluaga@users.noreply.github.com>
Date: Mon, 16 Mar 2026 20:23:09 +0100
Subject: [PATCH] [Bug][Qwen3TTS][Streaming] remove dynamic initial chunk and
 only compute on initial request (#1930)

Signed-off-by: JuanPZuluaga <juanz9312@gmail.com>
---
 .../chunk_transfer_adapter.py                 |  7 +++++
 .../stage_input_processors/qwen3_tts.py       | 26 ++++++++-----------
 2 files changed, 18 insertions(+), 15 deletions(-)

diff --git a/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py b/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py
index d96ef74db..8e9939d38 100644
--- a/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py
+++ b/vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py
@@ -238,6 +238,9 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
 
         if is_finished:
             self.code_prompt_token_ids.pop(request_id, None)
+            cached_ic = getattr(self, "_cached_ic", None)
+            if cached_ic is not None:
+                cached_ic.pop(request_id, None)
 
     ########################################################################
     # Cleanup
@@ -273,6 +276,10 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
         self.request_payload.pop(external_req_id, None)
         self.code_prompt_token_ids.pop(external_req_id, None)
 
+        cached_ic = getattr(self, "_cached_ic", None)
+        if cached_ic is not None:
+            cached_ic.pop(external_req_id, None)
+
     ########################################################################
     # Schedule Helper
     ########################################################################
diff --git a/vllm_omni/model_executor/stage_input_processors/qwen3_tts.py b/vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
index 435e45f03..c01517066 100644
--- a/vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+++ b/vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
@@ -115,22 +115,18 @@ def talker2code2wav_async_chunk(
             initial_chunk_size = int(entry.list_data[0])
             per_request_override = True
 
-    # Dynamic IC: always derived from chunk_size, recomputed every call to adapt to load.
+    # Dynamic IC: cache per request so boundaries stay stable for its lifetime.
     if not per_request_override:
-        max_ic = max_ic_for_chunk_size(chunk_size)
-        active = sum(1 for v in transfer_manager.code_prompt_token_ids.values() if len(v) > 0)
-        capacity = getattr(transfer_manager, "scheduler_max_num_seqs", 1)
-        initial_chunk_size = compute_dynamic_initial_chunk_size(active, capacity, max_ic)
-        logger.debug(
-            "Dynamic IC: active=%d, capacity=%d, max_ic=%d, ic=%d, cs=%d, req=%s, keys=%d",
-            active,
-            capacity,
-            max_ic,
-            initial_chunk_size,
-            chunk_size,
-            request_id,
-            len(transfer_manager.code_prompt_token_ids),
-        )
+        _ic_cache = getattr(transfer_manager, "_cached_ic", None)
+        if _ic_cache is None:
+            _ic_cache = {}
+            transfer_manager._cached_ic = _ic_cache
+        if request_id not in _ic_cache:
+            max_ic = max_ic_for_chunk_size(chunk_size)
+            active = sum(1 for v in transfer_manager.code_prompt_token_ids.values() if len(v) > 0)
+            capacity = getattr(transfer_manager, "scheduler_max_num_seqs", 1)
+            _ic_cache[request_id] = compute_dynamic_initial_chunk_size(active, capacity, max_ic)
+        initial_chunk_size = _ic_cache[request_id]
 
     if chunk_size <= 0 or left_context_size_config < 0 or initial_chunk_size < 0:
         raise ValueError(