From 0fc4032a712c5d8fec5cbab87ace8786f0fc42c7 Mon Sep 17 00:00:00 2001
From: tanhaoan333 <tanhaoan@huawei.com>
Date: Wed, 3 Jun 2026 21:34:47 +0800
Subject: [PATCH] Update qwen3_tts_code2wav.py (#4075)

Signed-off-by: tanhaoan333 <tanhaoan@huawei.com>
---
 .../npu/models/qwen3_tts_code2wav.py          | 48 +++++++++++++++++++
 vllm_omni/platforms/npu/platform.py           |  8 ++++
 2 files changed, 56 insertions(+)
 create mode 100644 vllm_omni/platforms/npu/models/qwen3_tts_code2wav.py

diff --git a/vllm_omni/platforms/npu/models/qwen3_tts_code2wav.py b/vllm_omni/platforms/npu/models/qwen3_tts_code2wav.py
new file mode 100644
index 000000000..282689bca
--- /dev/null
+++ b/vllm_omni/platforms/npu/models/qwen3_tts_code2wav.py
@@ -0,0 +1,48 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Monkey-patch ``Qwen3TTSCode2Wav.__init__`` for NPU Code2Wav runtime knobs."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+
+if TYPE_CHECKING:
+    pass
+
+logger = init_logger(__name__)
+
+_PATCHED = False
+_original_init = None
+
+
+def _prepare_npu_code2wav_runtime() -> None:
+    from vllm_omni.platforms import current_omni_platform
+
+    if not current_omni_platform.is_npu():
+        return
+    torch.npu.config.allow_internal_format = False
+    torch.npu.set_compile_mode(jit_compile=False)
+
+
+def _patched_init(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
+    _prepare_npu_code2wav_runtime()
+    assert _original_init is not None
+    _original_init(self, vllm_config=vllm_config, prefix=prefix)
+
+
+def apply_qwen3_tts_code2wav_patch() -> None:
+    global _PATCHED, _original_init
+    if _PATCHED:
+        return
+
+    from vllm_omni.model_executor.models.qwen3_tts.qwen3_tts_code2wav import Qwen3TTSCode2Wav
+
+    _original_init = Qwen3TTSCode2Wav.__init__
+    Qwen3TTSCode2Wav.__init__ = _patched_init  # type: ignore[method-assign]
+    _PATCHED = True
+    logger.debug("Applied NPU patch for Qwen3TTSCode2Wav.__init__")
diff --git a/vllm_omni/platforms/npu/platform.py b/vllm_omni/platforms/npu/platform.py
index faf69c1b8..3b9f3d15f 100644
--- a/vllm_omni/platforms/npu/platform.py
+++ b/vllm_omni/platforms/npu/platform.py
@@ -31,6 +31,14 @@ class NPUOmniPlatform(OmniPlatform, NPUPlatform):
     _omni_enum = OmniPlatformEnum.NPU
     dist_backend: str = "hccl"
 
+    # conv2d convolution operator in the code2wav module of Qwen3-TTS not being able to run on Aclnn
+    def __init__(self) -> None:
+        from vllm_omni.platforms.npu.models.patch_qwen3_tts_code2wav import (
+            apply_qwen3_tts_code2wav_patch,
+        )
+
+        apply_qwen3_tts_code2wav_patch()
+
     @classmethod
     def set_device(cls, device: torch.device) -> None:
         super().set_device(device)