From 425952e7d33c62cf38d0956eb457ad776c3be779 Mon Sep 17 00:00:00 2001
From: Markus / Mark <46672778+marksverdhei@users.noreply.github.com>
Date: Sat, 31 Jan 2026 01:00:03 +0100
Subject: [PATCH] [BugFix] Fix Qwen3 TTS 0.6B profile run hang (#995) (#1082)

---
 vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py b/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py
index 927ec0a90..8514a725d 100644
--- a/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py
+++ b/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py
@@ -125,6 +125,14 @@ class Qwen3TTSModelForGeneration(nn.Module):
             if isinstance(value, list) and len(value) > 0:
                 runtime_additional_information[key] = value[0]
 
+        # During profile/warmup runs, text is empty and no real inputs exist.
+        # Cap generation steps so the full pipeline executes (preserving
+        # KV-cache profiling behaviour) but exits quickly even if the model
+        # cannot converge from degenerate dummy inputs.
+        if not text:
+            logger.info("Profile run detected (empty text). Capping max_new_tokens to 2.")
+            runtime_additional_information["max_new_tokens"] = 2
+
         # Call the appropriate generation method based on task_type
         if task_type == "CustomVoice":
             result = self.model.generate_custom_voice(