Skip to content

vllm_omni.platforms.npu.models.qwen3_tts ¶

Patch Qwen3-TTS NPU runtime setup and weights.

The code2wav runtime/weight fixes apply on every NPU. On A5 (Ascend 950 PR) torch.stft is unavailable, so the speaker-encoder mel front-end of the prompt-embeds builder runs on CPU and the result is moved back to NPU (same trade-off as the 310P path).

logger `module-attribute` ¶

logger = init_logger(__name__)

apply_qwen3_tts_code2wav_patch ¶

apply_qwen3_tts_code2wav_patch() -> None

apply_qwen3_tts_patches ¶

apply_qwen3_tts_patches() -> None

Install all Qwen3-TTS NPU patches.

The code2wav runtime/weight patch applies on every NPU; the A5 prompt-embeds builder swap is gated on is_a5() because it changes the speaker-encoder mel front-end (torch.stft is unsupported there).