From 091b256674afa1bae476553c35ab1d22d155608a Mon Sep 17 00:00:00 2001
From: ztyan <191593482+vivivi-111@users.noreply.github.com>
Date: Mon, 5 Oct 2026 20:20:25 +0800
Subject: [PATCH] [Model][Perf] IndexTTS2: drop per-step hidden D2H in latent
 mode (#8478)

Signed-off-by: ztyan <191593482+vivivi-111@users.noreply.github.com>
---
 .../models/indextts2/test_talker_v2_5.py        |  6 +++---
 .../models/indextts2/indextts2_talker.py        | 17 +++++++++--------
 2 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/tests/model_executor/models/indextts2/test_talker_v2_5.py b/tests/model_executor/models/indextts2/test_talker_v2_5.py
index c916354e3..67efc787a 100644
--- a/tests/model_executor/models/indextts2/test_talker_v2_5.py
+++ b/tests/model_executor/models/indextts2/test_talker_v2_5.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from types import SimpleNamespace
 
@@ -215,8 +215,8 @@ def test_mel_head_preserves_official_checkpoint_bias_and_hidden_payload_policy(
 
     assert seen["bias"] is True
     assert talker.mel_head.bias is not None
-    assert talker.omni_pooler_payload_include_hidden is use_gpt_latent
-    assert talker.omni_payload_at_request_end is (model_type == "indextts2_5")
+    assert talker.omni_pooler_payload_include_hidden is False
+    assert talker.omni_payload_at_request_end is True
 
 
 def test_v2_load_weights_loads_exact_mel_head_bias_parameter(monkeypatch):
diff --git a/vllm_omni/model_executor/models/indextts2/indextts2_talker.py b/vllm_omni/model_executor/models/indextts2/indextts2_talker.py
index f9eae0c6f..a70fd9df4 100644
--- a/vllm_omni/model_executor/models/indextts2/indextts2_talker.py
+++ b/vllm_omni/model_executor/models/indextts2/indextts2_talker.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """IndexTTS2 Stage 0: GPT-2 AR Talker with vLLM-native PagedAttention.
 
 Predicts mel codes autoregressively and collects hidden_states as latent
@@ -158,17 +158,18 @@ class IndexTTS2TalkerForConditionalGeneration(nn.Module):
         self.has_postprocess = True
         self.requires_raw_input_tokens = True
         self.enable_update_additional_information = True
-        # Stage 1 consumes the hidden-state stream only in the legacy
-        # ``use_gpt_latent`` mode.  IndexTTS 2.5 transports mel codes plus
-        # reference conditioning, so asking the generic runner to include
-        # hidden states would add a synchronous [num_tokens, model_dim] D2H
-        # copy on every decode step without adding anything to the payload.
-        self.omni_pooler_payload_include_hidden = self.use_gpt_latent
+        # S2Mel never reads the generic pooler ``hidden`` key: latent mode
+        # ships the aligned GPT latent as ``hidden_states.latent`` from
+        # make_omni_output, and no-latent mode needs no hidden state at all.
+        # Including it would add a synchronous [num_tokens, model_dim] D2H
+        # copy plus a CPU-side concat on every decode step, which also
+        # serializes async scheduling.
+        self.omni_pooler_payload_include_hidden = False
         # S2Mel consumes the complete sequence only at request end. In latent
         # mode the runner retains aligned code/latent deltas on GPU; no-latent
         # mode reuses the CPU output-token history and retains only conditioning
         # metadata here.
-        self.omni_payload_at_request_end = self.config.model_type == "indextts2_5"
+        self.omni_payload_at_request_end = True
         self.omni_request_end_token_ids = ()
         self._speaker_cache = get_speaker_cache()
         self.gpu_resident_buffer_keys: set[tuple[str, str]] = {