From b3dd45874a750f7edfa39bb02262804228e5ff7b Mon Sep 17 00:00:00 2001
From: Tianyao Wu <54675599+twu3202@users.noreply.github.com>
Date: Wed, 9 Sep 2026 03:27:38 +0800
Subject: [PATCH] [Bugfix] Build the forced-aligner prompt without a chat
 template (word timestamps one bin late) (#7240)

Signed-off-by: Tianyao Wu <rayroy31@gmail.com>
---
 .../utils/test_qwen3_force_align_processor.py | 16 ++++++++++++++-
 .../deploy/qwen3_tts_forced_aligner.yaml      |  5 +++--
 .../utils/qwen3_force_align_processor.py      | 20 +++++++++----------
 3 files changed, 28 insertions(+), 13 deletions(-)

diff --git a/tests/utils/test_qwen3_force_align_processor.py b/tests/utils/test_qwen3_force_align_processor.py
index 4a928d0e5..f846ac7a8 100644
--- a/tests/utils/test_qwen3_force_align_processor.py
+++ b/tests/utils/test_qwen3_force_align_processor.py
@@ -10,7 +10,21 @@ def test_build_prompt_has_boundary_timestamp_markers():
 
     assert prompt.count("<timestamp>") == 4
     assert "hello<timestamp><timestamp>world" in prompt
-    assert prompt.endswith("<|im_start|>assistant\n")
+    # Official aligner format: audio placeholder + words, no chat template
+    # (a leading "<|im_start|>user\n" shifts the predicted markers one bin late).
+    assert prompt == f"{processor.AUDIO_PLACEHOLDER}hello<timestamp><timestamp>world<timestamp><timestamp>"
+    assert "<|im_start|>" not in prompt
+    assert processor.build_prompt([]) == f"{processor.AUDIO_PLACEHOLDER}<timestamp><timestamp>"
+
+
+def test_build_prompt_matches_official_encode_timestamp():
+    pytest.importorskip("qwen_asr")
+    from qwen_asr.inference.qwen3_forced_aligner import Qwen3ForceAlignProcessor
+
+    text = "It's 3 o'clock, 你好 world."
+    word_list, official_prompt = Qwen3ForceAlignProcessor().encode_timestamp(text, "english")
+
+    assert processor.build_prompt(list(word_list)) == official_prompt
 
 
 @pytest.mark.parametrize(
diff --git a/vllm_omni/deploy/qwen3_tts_forced_aligner.yaml b/vllm_omni/deploy/qwen3_tts_forced_aligner.yaml
index 2aecdbf9b..b600ff1e1 100644
--- a/vllm_omni/deploy/qwen3_tts_forced_aligner.yaml
+++ b/vllm_omni/deploy/qwen3_tts_forced_aligner.yaml
@@ -5,8 +5,9 @@
 # --forced-aligner sets just the model path.
 #
 # The stage runs the model with runner="pooling" + pooling_task="token_classify".
-# Prompt tokens, chat template and word segmentation are Qwen-specific and live
-# in qwen3_force_align_processor (not configurable here).
+# Prompt tokens and word segmentation are Qwen-specific and live in
+# qwen3_force_align_processor (not configurable here); the prompt is the same
+# string qwen_asr's encode_timestamp builds, with no chat template.
 
 forced_aligner:
   model: Qwen/Qwen3-ForcedAligner-0.6B
diff --git a/vllm_omni/utils/qwen3_force_align_processor.py b/vllm_omni/utils/qwen3_force_align_processor.py
index e70d270db..76c5b9084 100644
--- a/vllm_omni/utils/qwen3_force_align_processor.py
+++ b/vllm_omni/utils/qwen3_force_align_processor.py
@@ -65,18 +65,18 @@ _LANG_ALIASES = {
 
 
 def build_prompt(words: list[str]) -> str:
-    """Wrap segmented words in the Qwen3 aligner prompt template.
+    """Build the Qwen3 aligner prompt exactly as qwen_asr's
+    ``Qwen3ForceAlignProcessor.encode_timestamp`` does: the audio placeholder
+    followed by the words, each with two trailing ``<timestamp>`` markers
+    (start + end) that the model classifies into audio time bins.
 
-    Each word gets two trailing ``<timestamp>`` markers (start + end); the
-    model classifies each marker into an audio time bin.
+    No chat template. The official aligner feeds this string straight to the
+    tokenizer; wrapping it in ``<|im_start|>user ... <|im_start|>assistant``
+    puts three tokens in front of the audio and shifts almost every predicted
+    marker one 80 ms bin later than the official output.
     """
-    if not words:
-        # Pad with one timestamp so the decoder always has something to
-        # read; an empty result still surfaces as "[]" upstream.
-        body = TIMESTAMP_TOKEN
-    else:
-        body = f"{TIMESTAMP_TOKEN}{TIMESTAMP_TOKEN}".join(words) + f"{TIMESTAMP_TOKEN}{TIMESTAMP_TOKEN}"
-    return f"<|im_start|>user\n{AUDIO_PLACEHOLDER}{body}<|im_end|>\n<|im_start|>assistant\n"
+    body = f"{TIMESTAMP_TOKEN}{TIMESTAMP_TOKEN}".join(words) + f"{TIMESTAMP_TOKEN}{TIMESTAMP_TOKEN}"
+    return f"{AUDIO_PLACEHOLDER}{body}"
 
 
 # --- word segmentation (port of Qwen3ForceAlignProcessor) ----------------