# 1、拉取镜像
docker pull quay.io/ascend/vllm-omni:v0.20.0-a3

# 2、容器启动
export IMAGE=quay.io/ascend/vllm-omni:v0.20.0-a3
docker run -itd \
    --name Qwen3-TTS \
    --shm-size=1g \
    --net=host \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci_manager \
    --device /dev/devmm_svm \
    --device /dev/hisi_hdc \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /mnt/weight/:/mnt/weight/ \
    $IMAGE bash

# 3、进入容器
docker exec -it {container_id} /bin/bash

# 修改 /vllm-workspace/vllm-omni/vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_code2wav.py 文件
# 定位 _split_request_ids 方法，修改如下两行内容即可
# 具体可参照 PR：https://github.com/vllm-project/vllm-omni/pull/2823/changes

if seq_token_counts is not None and len(seq_token_counts) > 1:
--> if seq_token_counts is not None and len(seq_token_counts) >= 1:

boundaries.append(boundaries[-1] + count)
--> boundaries.append(boundaries[-1] + int(count))

modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local_dir ./Qwen3-TTS-12Hz-1.7B-CustomVoice
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --local_dir ./Qwen3-TTS-12Hz-1.7B-VoiceDesign
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-Base --local_dir ./Qwen3-TTS-12Hz-1.7B-Base

stages:
  - stage_id: 0
    max_num_seqs: 10
    gpu_memory_utilization: 0.6
    trust_remote_code: true

# 指定模型名称，自动下载模型
# Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
# Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
# Qwen/Qwen3-TTS-12Hz-1.7B-Base
export VLLM_USE_MODELSCOPE=true
vllm serve "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" \
    --host 0.0.0.0 \
    --port 8091 \
    --omni

# 指定模型下载路径，启动服务
vllm serve /mnt/weight/Qwen3-TTS-12Hz-1.7B-CustomVoice \
    --served-model-name "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" \
    --host 0.0.0.0 \
    --port 8091 \
    --omni 

curl -s -X POST http://0.0.0.0:8091/v1/audio/speech \
    -H "Content-Type: application/json" \
    -d '{
        "model": "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
        "input": "你好，我是通义千问。今天我们来做一段稍微长一点的语音生成测试，主要观察声音在连续表达时的稳定性、自然度和情绪一致性。一个好的语音模型，不仅要把每个字读清楚，还要在长句子里保持合适的停顿、语速和语调变化。比如在讲解复杂内容的时候，声音应该听起来从容、连贯，而不是忽快忽慢，或者在句子中间突然变得生硬。如果这段音频从开头到结尾都比较自然，发音清晰，节奏稳定，听起来没有明显的机械感，那就说明当前模型在中文长文本生成上的表现是比较可靠的。",
        "voice": "Vivian",
        "language": "Chinese"
    }' --output custom_output.wav

curl -s -X POST http://0.0.0.0:8091/v1/audio/speech \
    -H "Content-Type: application/json" \
    -d '{
        "model": "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
        "task_type": "VoiceDesign",
        "input": "今天我给你讲一个小笑话。有一只小猫第一次去上学，老师问它，一加一等于几呀？小猫想了半天，认真地说，等于两条小鱼！老师愣了一下，问它为什么。小猫眨眨眼说，因为我只要看到数字，就会想起午饭啦。虽然这个答案不太对，但是它说得太认真了，大家都忍不住笑了起来。",
        "instructions": "体现稚嫩的萝莉女声，音调偏高，语气活泼可爱，讲笑话时带一点俏皮和开心的感觉"
    }' --output design_output.wav

curl -s -X POST http://0.0.0.0:8091/v1/audio/speech \
    -H "Content-Type: application/json" \
    -d '{
        "model": "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
        "input": "Today I want to take a moment to reflect on how quickly technology is changing the way we communicate, learn, and create. A few years ago, speaking naturally with an intelligent system still felt like science fiction, but now it is becoming part of everyday life. What matters most is not only whether the voice sounds clear, but whether it feels steady, expressive, and comfortable to listen to over time. A good speech model should handle long sentences, short pauses, changes in rhythm, and subtle emotional tone without sounding mechanical or rushed. If this audio sounds smooth from beginning to end, with natural pacing and consistent pronunciation, then it is a strong sign that the generation quality is moving in the right direction.",
        "task_type": "Base",
        "ref_audio": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone_2.wav",
        "ref_text": "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."
    }' --output base_output.wav

# 压测命令参考
vllm bench serve --omni \
    --host 0.0.0.0 --port 8091 \
    --model /mnt/weight/Qwen3-TTS-12Hz-1.7B-CustomVoice \
    --served-model-name "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" \
    --backend openai-audio-speech \
    --endpoint /v1/audio/speech \
    --dataset-name seed-tts-text \
    --dataset-path /vllm-workspace/vllm-omni/benchmarks/build_dataset/seed_tts_smoke \
    --seed-tts-locale en \
    --num-prompts 10 \
    --num-warmups 2 \
    --extra-body '{"task_type":"CustomVoice"}' \
    --max-concurrency 1 --request-rate inf \
    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration \
    --save-result --result-dir ./results

# 压测命令参考
vllm bench serve --omni \
    --host 0.0.0.0 --port 8091 \
    --model /mnt/weight/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
    --served-model-name "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign" \
    --backend openai-audio-speech \
    --endpoint /v1/audio/speech \
    --dataset-name seed-tts-design \
    --dataset-path /vllm-workspace/vllm-omni/benchmarks/build_dataset/seed_tts_design \
    --seed-tts-locale en \
    --num-prompts 10 \
    --num-warmups 2 \
    --extra-body '{"task_type":"VoiceDesign"}' \
    --max-concurrency 1 --request-rate inf \
    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration \
    --save-result --result-dir ./results

# 数据集下载
wget --no-check-certificate https://modelscope.cn/datasets/CowboyZ/seed-tts-eval/resolve/master/seedtts_testset.tar

# 解压至数据集目录: /vllm-workspace/vllm-omni/benchmarks/build_dataset/
tar -xf seedtts_testset.tar

# 压测命令参考
vllm bench serve --omni \
    --host 0.0.0.0 --port 8000 \
    --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
    --backend openai-audio-speech \
    --endpoint /v1/audio/speech \
    --dataset-name seed-tts \
    --dataset-path /vllm-workspace/vllm-omni/benchmarks/build_dataset/seedtts_testset \
    --seed-tts-locale en \
    --num-prompts 10 --num-warmups 2 \
    --extra-body '{"task_type":"Base"}' \
    --max-concurrency 1 --request-rate inf \
    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration \
    --save-result --result-dir ./results