Add CUDA Graph inference acceleration support

2026-08-29 01:59:23 +02:00 · 2026-07-20 22:34:58 +08:00
parent f748ba4c48
commit b3f54df1be
26 changed files with 735 additions and 181 deletions
--- a/configs/config.py
+++ b/configs/config.py
@@ -10,6 +10,8 @@ from tools.file_io import read_text
 import torch
 import logging

+from tools.cuda_graph import configure_cuda_graph
+
 logger = logging.getLogger(__name__)


@@ -126,6 +128,12 @@ if infer_device.type != "cuda":
        )


+# Run a real capture/replay probe on the selected inference device.  Both
+# application entry points import this module, so downstream inference code
+# receives one consistent 0/1 switch without duplicating device checks.
+CUDA_GRAPH_AVAILABLE = configure_cuda_graph(infer_device)
+
+
 CONFIGS_DIR = Path(__file__).resolve().parent
 MODEL_CONFIG_FILES = (
    "v1/32k.json",
@@ -152,6 +160,7 @@ class Config:
        self.device = str(infer_device)
        self.dtype = infer_dtype
        self.is_half = infer_dtype == torch.float16
+        self.cuda_graph = CUDA_GRAPH_AVAILABLE
        self.n_cpu = 0
        self.gpu_name = None
        self.json_config = self.load_config_json()
--- a/i18n/locale/en_US.json
+++ b/i18n/locale/en_US.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Vocal extraction aggressiveness",
+{
+    "人声提取激进程度": "Vocal extraction aggressiveness",
    "%s → 成功": "%s → Success",
    "%s运行中，请先停止该任务": "%s is running; stop it before starting another task",
    "%s进程已终止": "The %s process has been terminated",
@@ -49,8 +49,8 @@
    "[索引训练][失败] 聚类失败，将使用原始特征继续\n%s": "[Index training][Failed] Clustering failed; continuing with the original features\n%s",
    "[索引训练][失败] 请先进行特征提取": "[Index training][Failed] Extract features first",
    "ckpt处理": "ckpt Processing",
-    "index文件路径不可包含中文": "The index file path cannot contain Chinese characters",
-    "pth文件路径不可包含中文": "The .pth file path cannot contain Chinese characters",
+    "index文件路径不可包含中文": "The index file path cannot contain Chinese characters",
+    "pth文件路径不可包含中文": "The .pth file path cannot contain Chinese characters",
    "rmvpe卡号配置：以-分隔输入使用的不同进程卡号,例如0-0-1使用在卡0上跑2个进程并在卡1上跑1个进程": "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1",
    "step1: 填写实验配置. 实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件. ": "Step 1: Fill in the experimental configuration. Experimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.",
    "step1:正在处理数据": "Step 1: Processing data",
@@ -71,8 +71,8 @@
    "任务": "Task",
    "伴奏人声分离&去混响&去回声": "Vocals/Accompaniment Separation & Reverberation Removal",
    "使用显卡：%s": "GPUs in use: %s",
-    "使用模型采样率": "Use model sample rate",
-    "使用设备采样率": "Use device sample rate",
+    "使用模型采样率": "Use model sample rate",
+    "使用设备采样率": "Use device sample rate",
    "保存名": "Save name:",
    "保存的文件名, 默认空为和源文件同名": "Save file name (default: same as the source file):",
    "保存的模型名不带后缀": "Saved model name (without extension):",
@@ -177,7 +177,7 @@
    "特征": "Features",
    "特征提取": "Feature extraction",
    "状态": "Status",
-    "独占 WASAPI 设备": "Exclusive WASAPI device",
+    "独占 WASAPI 设备": "Exclusive WASAPI device",
    "生成器预训练模型不存在，将不使用：assets/pretrained%s/%sG%s.pth": "Generator pretrained model not found; it will not be used: assets/pretrained%s/%sG%s.pth",
    "男转女推荐+12key, 女转男推荐-12key, 如果音域爆炸导致音色失真也可以自己调整到合适音域. ": "Recommended +12 key for male to female conversion, and -12 key for female to male conversion. If the sound range goes too far and the voice is distorted, you can also adjust it to the appropriate range by yourself.",
    "目标采样率": "Target sample rate:",
@@ -202,7 +202,7 @@
    "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log": "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.",
    "训练设备规则选择的精度：%s": "Training precision selected by device rules: %s",
    "训练轮次：{} [{:.0f}%]": "Training epoch: {} [{:.0f}%]",
-    "设备类型": "Device type",
+    "设备类型": "Device type",
    "请上传音频文件": "Upload an audio file",
    "请填写输出文件夹路径": "Enter the output folder path",
    "请指定说话人id": "Please specify the speaker/singer ID:",
@@ -233,7 +233,7 @@
    "选择模型": "Select model",
    "选择索引": "Select index",
    "选择音高提取算法": "Select the pitch extraction algorithm",
-    "采样率:": "Sample rate:",
+    "采样率:": "Sample rate:",
    "采样长度": "Sample length",
    "重载设备列表": "Reload device list",
    "错误信息：%s": "Error details: %s",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Index training] External index link already exists: %s",
    "[索引训练][跳过] added索引已存在：%s": "[Index training][Skipped] added index already exists: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Index training][Skipped] trained index already exists: %s",
-    "当前设备：%s | 推理精度：%s": "Current device: %s | Inference precision: %s"
+    "当前设备：%s | 推理精度：%s": "Current device: %s | Inference precision: %s",
+    "正在预热CUDA Graph": "Warming up CUDA Graph",
+    "CUDA Graph预热完成": "CUDA Graph warm-up complete"
 }
--- a/i18n/locale/es_ES.json
+++ b/i18n/locale/es_ES.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Intensidad de extracción vocal",
+{
+    "人声提取激进程度": "Intensidad de extracción vocal",
    "%s → 成功": "%s → Éxito",
    "%s运行中，请先停止该任务": "%s está en ejecución; deténgalo antes de iniciar otra tarea",
    "%s进程已终止": "El proceso %s ha finalizado",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Entrenamiento de índice] El enlace externo ya existe: %s",
    "[索引训练][跳过] added索引已存在：%s": "[Entrenamiento de índice][Omitido] El índice added ya existe: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Entrenamiento de índice][Omitido] El índice trained ya existe: %s",
-    "当前设备：%s | 推理精度：%s": "Dispositivo actual: %s | Precisión de inferencia: %s"
+    "当前设备：%s | 推理精度：%s": "Dispositivo actual: %s | Precisión de inferencia: %s",
+    "正在预热CUDA Graph": "Preparando CUDA Graph",
+    "CUDA Graph预热完成": "Preparación de CUDA Graph completada"
 }
--- a/i18n/locale/fr_FR.json
+++ b/i18n/locale/fr_FR.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Intensité de l’extraction vocale",
+{
+    "人声提取激进程度": "Intensité de l’extraction vocale",
    "%s → 成功": "%s → Réussite",
    "%s运行中，请先停止该任务": "%s est en cours ; arrêtez cette tâche avant d'en démarrer une autre",
    "%s进程已终止": "Le processus %s a été arrêté",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Entraînement de l’index] Le lien externe existe déjà : %s",
    "[索引训练][跳过] added索引已存在：%s": "[Entraînement de l’index][Ignoré] L’index added existe déjà : %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Entraînement de l’index][Ignoré] L’index trained existe déjà : %s",
-    "当前设备：%s | 推理精度：%s": "Périphérique actuel : %s | Précision d’inférence : %s"
+    "当前设备：%s | 推理精度：%s": "Périphérique actuel : %s | Précision d’inférence : %s",
+    "正在预热CUDA Graph": "Préchauffage de CUDA Graph",
+    "CUDA Graph预热完成": "Préchauffage de CUDA Graph terminé"
 }
--- a/i18n/locale/it_IT.json
+++ b/i18n/locale/it_IT.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Intensità dell'estrazione vocale",
+{
+    "人声提取激进程度": "Intensità dell'estrazione vocale",
    "%s → 成功": "%s → Successo",
    "%s运行中，请先停止该任务": "%s è in esecuzione; interromperlo prima di avviare un'altra attività",
    "%s进程已终止": "Il processo %s è stato terminato",
@@ -54,7 +54,7 @@
    "rmvpe卡号配置：以-分隔输入使用的不同进程卡号,例如0-0-1使用在卡0上跑2个进程并在卡1上跑1个进程": "Configurazione GPU RMVPE separata da trattini; ad esempio 0-0-1 avvia due processi sulla GPU 0 e uno sulla GPU 1",
    "step1: 填写实验配置. 实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件. ": "Passaggio 1: compilare la configurazione sperimentale. ",
    "step1:正在处理数据": "Passaggio 1: elaborazione dei dati",
-    "step2:正在提取音高&正在提取特征": "step2:Estrazione dell'intonazione e delle caratteristiche",
+    "step2:正在提取音高&正在提取特征": "step2:Estrazione dell'intonazione e delle caratteristiche",
    "step2a: 自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练. ": "Passaggio 2a: attraversa automaticamente tutti i file nella cartella di addestramento che possono essere decodificati in audio ed esegui la normalizzazione delle sezioni. ",
    "step2b: 使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号)": "Passaggio 2b: utilizzare la CPU per estrarre il tono (se il modello ha il tono), utilizzare la GPU per estrarre le caratteristiche (selezionare l'indice GPU):",
    "step3: 填写训练设置, 开始训练模型和索引": "Passaggio 3: compilare le impostazioni di addestramento e avviare l'addestramento del modello e dell'indice",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Addestramento indice] Il collegamento esterno esiste già: %s",
    "[索引训练][跳过] added索引已存在：%s": "[Addestramento indice][Saltato] L’indice added esiste già: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Addestramento indice][Saltato] L’indice trained esiste già: %s",
-    "当前设备：%s | 推理精度：%s": "Dispositivo corrente: %s | Precisione di inferenza: %s"
+    "当前设备：%s | 推理精度：%s": "Dispositivo corrente: %s | Precisione di inferenza: %s",
+    "正在预热CUDA Graph": "Riscaldamento di CUDA Graph",
+    "CUDA Graph预热完成": "Riscaldamento di CUDA Graph completato"
 }
--- a/i18n/locale/ja_JP.json
+++ b/i18n/locale/ja_JP.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "ボーカル抽出の強度",
+{
+    "人声提取激进程度": "ボーカル抽出の強度",
    "%s → 成功": "%s → 成功",
    "%s运行中，请先停止该任务": "%sを実行中です。先に停止してください",
    "%s进程已终止": "%sプロセスを終了しました",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[インデックス学習] 外部インデックスリンクは既に存在します：%s",
    "[索引训练][跳过] added索引已存在：%s": "[インデックス学習][スキップ] addedインデックスは既に存在します：%s",
    "[索引训练][跳过] trained索引已存在：%s": "[インデックス学習][スキップ] trainedインデックスは既に存在します：%s",
-    "当前设备：%s | 推理精度：%s": "現在のデバイス：%s | 推論精度：%s"
+    "当前设备：%s | 推理精度：%s": "現在のデバイス：%s | 推論精度：%s",
+    "正在预热CUDA Graph": "CUDA Graphをウォームアップしています",
+    "CUDA Graph预热完成": "CUDA Graphのウォームアップが完了しました"
 }
--- a/i18n/locale/ko_KR.json
+++ b/i18n/locale/ko_KR.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "보컬 추출 강도",
+{
+    "人声提取激进程度": "보컬 추출 강도",
    "%s → 成功": "%s → 성공",
    "%s运行中，请先停止该任务": "%s이(가) 실행 중입니다. 먼저 중지하세요",
    "%s进程已终止": "%s 프로세스가 종료되었습니다",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[인덱스 학습] 외부 인덱스 링크가 이미 있습니다: %s",
    "[索引训练][跳过] added索引已存在：%s": "[인덱스 학습][건너뜀] added 인덱스가 이미 있습니다: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[인덱스 학습][건너뜀] trained 인덱스가 이미 있습니다: %s",
-    "当前设备：%s | 推理精度：%s": "현재 장치: %s | 추론 정밀도: %s"
+    "当前设备：%s | 推理精度：%s": "현재 장치: %s | 추론 정밀도: %s",
+    "正在预热CUDA Graph": "CUDA Graph 워밍업 중",
+    "CUDA Graph预热完成": "CUDA Graph 워밍업 완료"
 }
--- a/i18n/locale/pt_BR.json
+++ b/i18n/locale/pt_BR.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Intensidade da extração vocal",
+{
+    "人声提取激进程度": "Intensidade da extração vocal",
    "%s → 成功": "%s → Sucesso",
    "%s运行中，请先停止该任务": "%s está em execução; interrompa antes de iniciar outra tarefa",
    "%s进程已终止": "O processo %s foi encerrado",
@@ -54,7 +54,7 @@
    "rmvpe卡号配置：以-分隔输入使用的不同进程卡号,例如0-0-1使用在卡0上跑2个进程并在卡1上跑1个进程": "Configuração do número do cartão rmvpe: Use - para separar os números dos cartões de entrada de diferentes processos. Por exemplo, 0-0-1 é usado para executar 2 processos no cartão 0 e 1 processo no cartão 1.",
    "step1: 填写实验配置. 实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件. ": "Etapa 1: Preencha a configuração experimental. Os dados experimentais são armazenados na pasta 'logs', com cada experimento tendo uma pasta separada. Digite manualmente o caminho do nome do experimento, que contém a configuração experimental, os logs e os arquivos de modelo treinados.",
    "step1:正在处理数据": "Etapa 1: Processamento de dados",
-    "step2:正在提取音高&正在提取特征": "step2:Extração de tom e características",
+    "step2:正在提取音高&正在提取特征": "step2:Extração de tom e características",
    "step2a: 自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练. ": "Etapa 2a: Percorra automaticamente todos os arquivos na pasta de treinamento que podem ser decodificados em áudio e execute a normalização da fatia. Gera 2 pastas wav no diretório do experimento. Atualmente, apenas o treinamento de um único cantor/palestrante é suportado.",
    "step2b: 使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号)": "Etapa 2b: Use a CPU para extrair o tom (se o modelo tiver tom), use a GPU para extrair recursos (selecione o índice da GPU):",
    "step3: 填写训练设置, 开始训练模型和索引": "Etapa 3: Preencha as configurações de treinamento e comece a treinar o modelo e o Index",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Treinamento de índice] O link externo já existe: %s",
    "[索引训练][跳过] added索引已存在：%s": "[Treinamento de índice][Ignorado] O índice added já existe: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Treinamento de índice][Ignorado] O índice trained já existe: %s",
-    "当前设备：%s | 推理精度：%s": "Dispositivo atual: %s | Precisão de inferência: %s"
+    "当前设备：%s | 推理精度：%s": "Dispositivo atual: %s | Precisão de inferência: %s",
+    "正在预热CUDA Graph": "Aquecendo o CUDA Graph",
+    "CUDA Graph预热完成": "Aquecimento do CUDA Graph concluído"
 }
--- a/i18n/locale/ru_RU.json
+++ b/i18n/locale/ru_RU.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Интенсивность выделения вокала",
+{
+    "人声提取激进程度": "Интенсивность выделения вокала",
    "%s → 成功": "%s → Успешно",
    "%s运行中，请先停止该任务": "%s уже выполняется; сначала остановите эту задачу",
    "%s进程已终止": "Процесс %s завершён",
@@ -54,7 +54,7 @@
    "rmvpe卡号配置：以-分隔输入使用的不同进程卡号,例如0-0-1使用在卡0上跑2个进程并在卡1上跑1个进程": "Введите номера графических процессоров, разделенные символом «-», например, 0-0-1, чтобы запустить два процесса на GPU 0 и один процесс на GPU 1:",
    "step1: 填写实验配置. 实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件. ": "Шаг 1. Конфигурирование модели. Данные обучения модели сохраняются в папку 'logs', и для каждой модели создаётся отдельная папка. Введите вручную путь к настройкам для модели, в которой находятся логи и тренировочные файлы.",
    "step1:正在处理数据": "Шаг 1. Переработка данных",
-    "step2:正在提取音高&正在提取特征": "step2:Извлечение высоты тона и признаков",
+    "step2:正在提取音高&正在提取特征": "step2:Извлечение высоты тона и признаков",
    "step2a: 自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练. ": "Шаг 2А. Автоматическая обработка исходных аудиозаписей для обучения и выполнение нормализации среза. Создаст 2 папки wav в папке модели. В данный момент поддерживается обучение только на одноголосных записях.",
    "step2b: 使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号)": "Шаг 2Б. Оценка и извлечение тональности в аудиофайлах с помощью процессора (если включена поддержка изменения высоты звука), извлечение черт с помощью GPU (выберите номер GPU):",
    "step3: 填写训练设置, 开始训练模型和索引": "Шаг 3. Заполнение дополнительных настроек обучения и запуск обучения модели и индекса",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Обучение индекса] Внешняя ссылка уже существует: %s",
    "[索引训练][跳过] added索引已存在：%s": "[Обучение индекса][Пропущено] Индекс added уже существует: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Обучение индекса][Пропущено] Индекс trained уже существует: %s",
-    "当前设备：%s | 推理精度：%s": "Текущее устройство: %s | Точность вывода: %s"
+    "当前设备：%s | 推理精度：%s": "Текущее устройство: %s | Точность вывода: %s",
+    "正在预热CUDA Graph": "Прогрев CUDA Graph",
+    "CUDA Graph预热完成": "Прогрев CUDA Graph завершён"
 }
--- a/i18n/locale/tr_TR.json
+++ b/i18n/locale/tr_TR.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "Vokal ayırma yoğunluğu",
+{
+    "人声提取激进程度": "Vokal ayırma yoğunluğu",
    "%s → 成功": "%s → Başarılı",
    "%s运行中，请先停止该任务": "%s çalışıyor; başka bir görev başlatmadan önce durdurun",
    "%s进程已终止": "%s işlemi sonlandırıldı",
@@ -54,7 +54,7 @@
    "rmvpe卡号配置：以-分隔输入使用的不同进程卡号,例如0-0-1使用在卡0上跑2个进程并在卡1上跑1个进程": "RMVPE GPU yapılandırmasını tirelerle ayırın; örneğin 0-0-1, GPU 0'da iki ve GPU 1'de bir işlem çalıştırır",
    "step1: 填写实验配置. 实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件. ": "Adım 1: Deneysel yapılandırmayı doldurun. Deneysel veriler 'logs' klasöründe saklanır ve her bir deney için ayrı bir klasör vardır. Deneysel adı yolu manuel olarak girin; bu yol, deneysel yapılandırmayı, günlükleri ve eğitilmiş model dosyalarını içerir.",
    "step1:正在处理数据": "Adım 1: Veri işleme",
-    "step2:正在提取音高&正在提取特征": "step2:Perde ve özellik çıkarımı",
+    "step2:正在提取音高&正在提取特征": "step2:Perde ve özellik çıkarımı",
    "step2a: 自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练. ": "Adım 2a: Eğitim klasöründe ses dosyalarını otomatik olarak gezinerek dilimleme normalizasyonu yapın. Deney dizini içinde 2 wav klasörü oluşturur. Şu anda sadece tek kişilik eğitim desteklenmektedir.",
    "step2b: 使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号)": "Adım 2b: Ses yüksekliği (Pitch) çıkartmak için CPU kullanın (eğer model ses yüksekliği içeriyorsa), özellikleri çıkartmak için GPU kullanın (GPU indeksini seçin):",
    "step3: 填写训练设置, 开始训练模型和索引": "Adım 3: Eğitim ayarlarını doldurun ve modeli ve dizini eğitmeye başlayın",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[Dizin eğitimi] Harici dizin bağlantısı zaten var: %s",
    "[索引训练][跳过] added索引已存在：%s": "[Dizin eğitimi][Atlandı] added dizini zaten var: %s",
    "[索引训练][跳过] trained索引已存在：%s": "[Dizin eğitimi][Atlandı] trained dizini zaten var: %s",
-    "当前设备：%s | 推理精度：%s": "Geçerli cihaz: %s | Çıkarım hassasiyeti: %s"
+    "当前设备：%s | 推理精度：%s": "Geçerli cihaz: %s | Çıkarım hassasiyeti: %s",
+    "正在预热CUDA Graph": "CUDA Graph ısınıyor",
+    "CUDA Graph预热完成": "CUDA Graph ısınması tamamlandı"
 }
--- a/i18n/locale/zh_CN.json
+++ b/i18n/locale/zh_CN.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "人声提取激进程度",
+{
+    "人声提取激进程度": "人声提取激进程度",
    "%s → 成功": "%s → 成功",
    "%s运行中，请先停止该任务": "%s运行中，请先停止该任务",
    "%s进程已终止": "%s进程已终止",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[索引训练] 外部索引链接已存在：%s",
    "[索引训练][跳过] added索引已存在：%s": "[索引训练][跳过] added索引已存在：%s",
    "[索引训练][跳过] trained索引已存在：%s": "[索引训练][跳过] trained索引已存在：%s",
-    "当前设备：%s | 推理精度：%s": "当前设备：%s | 推理精度：%s"
+    "当前设备：%s | 推理精度：%s": "当前设备：%s | 推理精度：%s",
+    "正在预热CUDA Graph": "正在预热CUDA Graph",
+    "CUDA Graph预热完成": "CUDA Graph预热完成"
 }
--- a/i18n/locale/zh_HK.json
+++ b/i18n/locale/zh_HK.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "人聲提取激進程度",
+{
+    "人声提取激进程度": "人聲提取激進程度",
    "%s → 成功": "%s → 成功",
    "%s运行中，请先停止该任务": "%s运行中，请先停止该任务",
    "%s进程已终止": "%s進程已终止",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[索引訓練] 外部索引連結已存在：%s",
    "[索引训练][跳过] added索引已存在：%s": "[索引訓練][跳過] added索引已存在：%s",
    "[索引训练][跳过] trained索引已存在：%s": "[索引訓練][跳過] trained索引已存在：%s",
-    "当前设备：%s | 推理精度：%s": "目前裝置：%s | 推理精度：%s"
+    "当前设备：%s | 推理精度：%s": "目前裝置：%s | 推理精度：%s",
+    "正在预热CUDA Graph": "正在預熱 CUDA Graph",
+    "CUDA Graph预热完成": "CUDA Graph 預熱完成"
 }
--- a/i18n/locale/zh_SG.json
+++ b/i18n/locale/zh_SG.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "人声提取激进程度",
+{
+    "人声提取激进程度": "人声提取激进程度",
    "%s → 成功": "%s → 成功",
    "%s运行中，请先停止该任务": "%s运行中，请先停止该任务",
    "%s进程已终止": "%s进程已终止",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[索引训练] 外部索引链接已存在：%s",
    "[索引训练][跳过] added索引已存在：%s": "[索引训练][跳过] added索引已存在：%s",
    "[索引训练][跳过] trained索引已存在：%s": "[索引训练][跳过] trained索引已存在：%s",
-    "当前设备：%s | 推理精度：%s": "当前设备：%s | 推理精度：%s"
+    "当前设备：%s | 推理精度：%s": "当前设备：%s | 推理精度：%s",
+    "正在预热CUDA Graph": "正在预热CUDA Graph",
+    "CUDA Graph预热完成": "CUDA Graph预热完成"
 }
--- a/i18n/locale/zh_TW.json
+++ b/i18n/locale/zh_TW.json
@@ -1,5 +1,5 @@
-{
-    "人声提取激进程度": "人聲提取激進程度",
+{
+    "人声提取激进程度": "人聲提取激進程度",
    "%s → 成功": "%s → 成功",
    "%s运行中，请先停止该任务": "%s运行中，请先停止该任务",
    "%s进程已终止": "%s進程已终止",
@@ -257,5 +257,7 @@
    "[索引训练] 外部索引链接已存在：%s": "[索引訓練] 外部索引連結已存在：%s",
    "[索引训练][跳过] added索引已存在：%s": "[索引訓練][跳過] added索引已存在：%s",
    "[索引训练][跳过] trained索引已存在：%s": "[索引訓練][跳過] trained索引已存在：%s",
-    "当前设备：%s | 推理精度：%s": "目前裝置：%s | 推論精度：%s"
+    "当前设备：%s | 推理精度：%s": "目前裝置：%s | 推論精度：%s",
+    "正在预热CUDA Graph": "正在預熱 CUDA Graph",
+    "CUDA Graph预热完成": "CUDA Graph 預熱完成"
 }
--- a/infer/fcpe.py
+++ b/infer/fcpe.py
@@ -1,5 +1,7 @@
 import torch

+from tools.cuda_graph import cuda_graph_enabled, run_cuda_graph
+

 def _is_directml_device(device):
    """Return whether *device* is the PrivateUse1 device registered by DirectML."""
@@ -35,6 +37,46 @@ class FCPEInfer:
            self.infer_model.model.to(device).eval()
        else:
            self.infer_model = spawn_bundled_infer_model(device)
+            if getattr(device, "type", None) == "cuda" or str(device).startswith(
+                "cuda"
+            ):
+                # torchfcpe creates this tensor on CPU and copies it to CUDA in
+                # every local-argmax decode.  Host-to-device copies are forbidden
+                # during CUDA Graph capture, so keep the immutable offsets on the
+                # target device and use the graph-safe equivalent decoder below.
+                self.local_offsets = torch.arange(
+                    9, device=device, dtype=torch.long
+                ).view(1, 1, 9)
+
+    def _graphable_model_infer(self, mel, decoder_mode, threshold):
+        """Run FCPE network and an exactly equivalent capture-safe decoder."""
+        model = self.infer_model.model
+        latent = model(mel)
+        batch, frames, _ = latent.shape
+        cents = model.cent_table[None, None, :].expand(batch, frames, -1)
+
+        if decoder_mode == "argmax":
+            confidence = torch.max(latent, dim=-1, keepdim=True).values
+            decoded = torch.sum(cents * latent, dim=-1, keepdim=True) / torch.sum(
+                latent, dim=-1, keepdim=True
+            )
+        elif decoder_mode == "local_argmax":
+            confidence, max_index = torch.max(latent, dim=-1, keepdim=True)
+            local_index = self.local_offsets + (max_index - 4)
+            local_index = local_index.clamp(0, model.out_dims - 1)
+            local_cents = torch.gather(cents, -1, local_index)
+            local_latent = torch.gather(latent, -1, local_index)
+            decoded = torch.sum(
+                local_cents * local_latent, dim=-1, keepdim=True
+            ) / torch.sum(local_latent, dim=-1, keepdim=True)
+        else:
+            raise ValueError("Unknown FCPE decoder mode: %s" % decoder_mode)
+
+        # Match torchfcpe's masking and cent-to-Hz formulas operation-for-operation.
+        confidence_mask = torch.ones_like(confidence)
+        confidence_mask.masked_fill_(confidence <= threshold, float("-inf"))
+        decoded = decoded * confidence_mask
+        return 10.0 * torch.pow(2.0, decoded / 1200.0)

    def _decode_on_cpu(self, latent, decoder_mode, threshold):
        """Decode DML network logits on CPU with torchfcpe's exact formulas.
@@ -79,11 +121,33 @@ class FCPEInfer:
        threshold=0.006,
    ):
        if not self.is_directml:
-            return self.infer_model.infer(
+            wav = wav.to(self.device)
+            if cuda_graph_enabled(wav.device):
+                # Wav2MelModule contains tensor-dependent Python conditionals and
+                # cannot be captured.  Running it eagerly also creates/caches its
+                # STFT window before the graph.  The much larger FCPE neural net
+                # and decoder form the stable-shape CUDA Graph boundary.
+                mel = self.infer_model.wav2mel(wav, sr)
+                return run_cuda_graph(
+                    self.infer_model.model,
+                    "fcpe-core-%s-%s" % (decoder_mode, threshold),
+                    lambda input_mel: self._graphable_model_infer(
+                        input_mel,
+                        decoder_mode,
+                        threshold,
+                    ),
+                    mel,
+                )
+            return run_cuda_graph(
+                self.infer_model,
+                "fcpe-%s-%s-%s" % (sr, decoder_mode, threshold),
+                lambda input_wav: self.infer_model.infer(
+                    input_wav,
+                    sr=sr,
+                    decoder_mode=decoder_mode,
+                    threshold=threshold,
+                ),
                wav,
-                sr=sr,
-                decoder_mode=decoder_mode,
-                threshold=threshold,
            )

        wav_cpu = wav.detach().to(device="cpu", dtype=torch.float32)
--- a/infer/hubert.py
+++ b/infer/hubert.py
@@ -4,7 +4,9 @@ from pathlib import Path

 import torch
 from torch import nn
-from transformers import AutoFeatureExtractor, HubertModel
+from transformers import AutoFeatureExtractor, HubertModel
+
+from tools.cuda_graph import run_cuda_graph


 logger = logging.getLogger(__name__)
@@ -63,7 +65,7 @@ def hubert_audio_requires_normalization():
    return bool(feature_extractor.do_normalize)


-def extract_hubert_features(model, source, version, padding_mask=None):
+def extract_hubert_features(model, source, version, padding_mask=None):
    """Return the RVC v1 (256-D) or v2 (768-D) HuBERT representation.

    Transformers hidden_states[N] is numerically equivalent to the source checkpoint's
@@ -77,20 +79,49 @@ def extract_hubert_features(model, source, version, padding_mask=None):
    if padding_mask is not None and bool(torch.any(padding_mask).item()):
        attention_mask = (~padding_mask.bool()).long()

-    if version == "v1":
-        outputs = model(
-            input_values=source,
-            attention_mask=attention_mask,
-            output_hidden_states=True,
-            return_dict=True,
-        )
-        features = outputs.hidden_states[9]
-        return model.final_proj(features)
-
-    outputs = model(
-        input_values=source,
-        attention_mask=attention_mask,
-        output_hidden_states=False,
-        return_dict=True,
-    )
-    return outputs.last_hidden_state
+    if version == "v1":
+        if attention_mask is None:
+            def forward(input_values):
+                outputs = model(
+                    input_values=input_values,
+                    attention_mask=None,
+                    output_hidden_states=True,
+                    return_dict=True,
+                )
+                return model.final_proj(outputs.hidden_states[9])
+
+            return run_cuda_graph(model, "hubert-v1-no-mask", forward, source)
+
+        def forward(input_values, mask):
+            outputs = model(
+                input_values=input_values,
+                attention_mask=mask,
+                output_hidden_states=True,
+                return_dict=True,
+            )
+            return model.final_proj(outputs.hidden_states[9])
+
+        return run_cuda_graph(
+            model, "hubert-v1-mask", forward, source, attention_mask
+        )
+
+    if attention_mask is None:
+        def forward(input_values):
+            return model(
+                input_values=input_values,
+                attention_mask=None,
+                output_hidden_states=False,
+                return_dict=True,
+            ).last_hidden_state
+
+        return run_cuda_graph(model, "hubert-v2-no-mask", forward, source)
+
+    def forward(input_values, mask):
+        return model(
+            input_values=input_values,
+            attention_mask=mask,
+            output_hidden_states=False,
+            return_dict=True,
+        ).last_hidden_state
+
+    return run_cuda_graph(model, "hubert-v2-mask", forward, source, attention_mask)
--- a/infer/module/models.py
+++ b/infer/module/models.py
@@ -67,8 +67,11 @@ class TextEncoder(nn.Module):
        )
        x = self.encoder(x * x_mask, x_mask)
        if skip_head is not None:
-            assert isinstance(skip_head, torch.Tensor)
-            head = int(skip_head.item())
+            head = (
+                int(skip_head.item())
+                if isinstance(skip_head, torch.Tensor)
+                else int(skip_head)
+            )
            x = x[:, :, head:]
            x_mask = x_mask[:, :, head:]
        stats = self.proj(x) * x_mask
@@ -231,8 +234,7 @@ class Generator(torch.nn.Module):
        n_res = None,
    ):
        if n_res is not None:
-            assert isinstance(n_res, torch.Tensor)
-            n = int(n_res.item())
+            n = int(n_res.item()) if isinstance(n_res, torch.Tensor) else int(n_res)
            if n != x.shape[-1]:
                x = F.interpolate(x, size=n, mode="linear")
        x = self.conv_pre(x)
@@ -501,8 +503,7 @@ class GeneratorNSF(torch.nn.Module):
        har_source, noi_source, uv = self.m_source(f0, self.upp)
        har_source = har_source.transpose(1, 2)
        if n_res is not None:
-            assert isinstance(n_res, torch.Tensor)
-            n = int(n_res.item())
+            n = int(n_res.item()) if isinstance(n_res, torch.Tensor) else int(n_res)
            if n * self.upp != har_source.shape[-1]:
                har_source = F.interpolate(har_source, size=n * self.upp, mode="linear")
            if n != x.shape[-1]:
@@ -674,12 +675,14 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
    ):
        g = self.emb_g(sid).unsqueeze(-1)
        if skip_head is not None and return_length is not None:
-            assert isinstance(skip_head, torch.Tensor)
-            assert isinstance(return_length, torch.Tensor)
-            head = int(skip_head.item())
-            length = int(return_length.item())
-            flow_head = torch.clamp(skip_head - 24, min=0)
-            dec_head = head - int(flow_head.item())
+            head = int(skip_head.item()) if isinstance(skip_head, torch.Tensor) else int(skip_head)
+            length = (
+                int(return_length.item())
+                if isinstance(return_length, torch.Tensor)
+                else int(return_length)
+            )
+            flow_head = max(head - 24, 0)
+            dec_head = head - flow_head
            m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths, flow_head)
            z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
            z = self.flow(z_p, x_mask, g=g, reverse=True)
@@ -862,12 +865,14 @@ class SynthesizerTrnMs256NSFsid_nono(nn.Module):
    ):
        g = self.emb_g(sid).unsqueeze(-1)
        if skip_head is not None and return_length is not None:
-            assert isinstance(skip_head, torch.Tensor)
-            assert isinstance(return_length, torch.Tensor)
-            head = int(skip_head.item())
-            length = int(return_length.item())
-            flow_head = torch.clamp(skip_head - 24, min=0)
-            dec_head = head - int(flow_head.item())
+            head = int(skip_head.item()) if isinstance(skip_head, torch.Tensor) else int(skip_head)
+            length = (
+                int(return_length.item())
+                if isinstance(return_length, torch.Tensor)
+                else int(return_length)
+            )
+            flow_head = max(head - 24, 0)
+            dec_head = head - flow_head
            m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths, flow_head)
            z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
            z = self.flow(z_p, x_mask, g=g, reverse=True)
--- a/infer/rmvpe.py
+++ b/infer/rmvpe.py
@@ -6,7 +6,9 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from librosa.util import normalize, pad_center, tiny
-from scipy.signal import get_window
+from scipy.signal import get_window
+
+from tools.cuda_graph import run_cuda_graph

 import logging

@@ -536,7 +538,7 @@ class RMVPE:
        cents_mapping = 20 * np.arange(360) + 1997.3794084376191
        self.cents_mapping = np.pad(cents_mapping, (4, 4))  # 368

-    def mel2hidden(self, mel):
+    def mel2hidden(self, mel):
        with torch.no_grad():
            n_frames = mel.shape[-1]
            n_pad = 32 * ((n_frames - 1) // 32 + 1) - n_frames
@@ -549,10 +551,30 @@ class RMVPE:
                    [onnx_outputs_names],
                    input_feed={onnx_input_name: mel.cpu().numpy()},
                )[0]
-            else:
-                mel = mel.half() if self.is_half else mel.float()
-                hidden = self.model(mel)
-            return hidden[:, :n_frames]
+            else:
+                mel = mel.half() if self.is_half else mel.float()
+                hidden = run_cuda_graph(
+                    self.model,
+                    "rmvpe-network",
+                    lambda input_mel: self.model(input_mel),
+                    mel,
+                )
+            return hidden[:, :n_frames]
+
+    def extract_mel(self, audio, center=True):
+        if not torch.is_tensor(audio):
+            audio = torch.from_numpy(audio)
+        audio = audio.float().to(self.device)
+        if audio.dim() == 1:
+            audio = audio.unsqueeze(0)
+        if "privateuseone" in str(self.device):
+            return self.mel_extractor(audio, center=center)
+        return run_cuda_graph(
+            self.mel_extractor,
+            "rmvpe-mel-center-%s" % int(bool(center)),
+            lambda input_audio: self.mel_extractor(input_audio, center=center),
+            audio,
+        )

    def decode(self, hidden, thred=0.03):
        cents_pred = self.to_local_average_cents(hidden, thred=thred)
@@ -564,11 +586,7 @@ class RMVPE:
    def infer_from_audio(self, audio, thred=0.03):
        # torch.cuda.synchronize()
        # t0 = ttime()
-        if not torch.is_tensor(audio):
-            audio = torch.from_numpy(audio)
-        mel = self.mel_extractor(
-            audio.float().to(self.device).unsqueeze(0), center=True
-        )
+        mel = self.extract_mel(audio, center=True)
        # print(123123123,mel.device.type)
        # torch.cuda.synchronize()
        # t1 = ttime()
--- a/infer/rtrvc.py
+++ b/infer/rtrvc.py
@@ -10,6 +10,7 @@ from torchaudio.transforms import Resample

 from infer.hubert import extract_hubert_features, load_hubert_model
 from i18n.i18n import I18nAuto
+from tools.cuda_graph import run_cuda_graph


 i18n = I18nAuto()
@@ -93,6 +94,7 @@ class RVC:
            self.cache_pitchf = torch.zeros(
                1024, device=self.device, dtype=torch.float32
            )
+            self.infer_count = 0

            self.resample_kernel = {}

@@ -179,11 +181,9 @@ class RVC:
        if len(f0) < p_len:
            f0 = np.pad(f0, (0, p_len - len(f0)))
        f0 = f0[:p_len]
-        try:
-            uv = f0 == 0
+        uv = f0 == 0
+        if np.any(~uv):
            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
-        except Exception:
-            traceback.print_exc()
        f0 *= pow(2, f0_up_key / 12)
        return self.get_f0_post(f0)

@@ -198,11 +198,9 @@ class RVC:
                device=self.device,
            )
        f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
-        try:
-            uv = f0 == 0
+        uv = f0 == 0
+        if np.any(~uv):
            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
-        except Exception:
-            traceback.print_exc()
        f0 *= pow(2, f0_up_key / 12)
        return self.get_f0_post(f0)

@@ -218,11 +216,9 @@ class RVC:
            decoder_mode="local_argmax",
            threshold=0.006,
        ).squeeze().detach().cpu().numpy()
-        try:
-            uv = f0 == 0
+        uv = f0 == 0
+        if np.any(~uv):
            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
-        except Exception:
-            traceback.print_exc()
        f0 *= pow(2, f0_up_key / 12)
        return self.get_f0_post(f0)

@@ -234,6 +230,8 @@ class RVC:
        return_length,
        f0method,
    ) :
+        report_status = self.infer_count < 3 or self.infer_count % 100 == 0
+        self.infer_count += 1
        t1 = ttime()
        with torch.no_grad():
            if self.config.is_half:
@@ -271,7 +269,8 @@ class RVC:
                        i18n("索引无效：必须使用added_xxxx.index，不能使用trained_xxxx.index")
                    )
            else:
-                printt(i18n("索引检索失败或未启用"))
+                if report_status:
+                    printt(i18n("索引检索失败或未启用"))
        except Exception:
            traceback.print_exc()
            printt(i18n("索引检索失败"))
@@ -298,26 +297,49 @@ class RVC:
        t4 = ttime()
        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
        feats = feats[:, :p_len, :]
-        p_len = torch.LongTensor([p_len]).to(self.device)
+        p_len_tensor = torch.LongTensor([p_len]).to(self.device)
        sid = torch.LongTensor([0]).to(self.device)
-        skip_head = torch.LongTensor([skip_head])
-        return_length2 = torch.LongTensor([return_length2])
-        return_length = torch.LongTensor([return_length])
+        skip_head_value = int(skip_head)
+        return_length_value = int(return_length)
+        return_length2_value = int(return_length2)
        with torch.no_grad():
            if self.if_f0 == 1:
-                infered_audio, _, _ = self.net_g.infer(
+                infered_audio = run_cuda_graph(
+                    self.net_g,
+                    "rvc-realtime-f0-%s-%s-%s"
+                    % (skip_head_value, return_length_value, return_length2_value),
+                    lambda phone, lengths, coarse, continuous, speaker: self.net_g.infer(
+                        phone,
+                        lengths,
+                        coarse,
+                        continuous,
+                        speaker,
+                        skip_head_value,
+                        return_length_value,
+                        return_length2_value,
+                    )[0],
                    feats,
-                    p_len,
+                    p_len_tensor,
                    cache_pitch,
                    cache_pitchf,
                    sid,
-                    skip_head,
-                    return_length,
-                    return_length2,
                )
            else:
-                infered_audio, _, _ = self.net_g.infer(
-                    feats, p_len, sid, skip_head, return_length, return_length2
+                infered_audio = run_cuda_graph(
+                    self.net_g,
+                    "rvc-realtime-no-f0-%s-%s-%s"
+                    % (skip_head_value, return_length_value, return_length2_value),
+                    lambda phone, lengths, speaker: self.net_g.infer(
+                        phone,
+                        lengths,
+                        speaker,
+                        skip_head_value,
+                        return_length_value,
+                        return_length2_value,
+                    )[0],
+                    feats,
+                    p_len_tensor,
+                    sid,
                )
        infered_audio = infered_audio.squeeze(1).float()
        upp_res = int(np.floor(factor * self.tgt_sr // 100))
@@ -332,11 +354,12 @@ class RVC:
                infered_audio[:, : return_length * upp_res]
            )
        t5 = ttime()
-        printt(
-            i18n("耗时：特征=%.3f秒，索引=%.3f秒，音高=%.3f秒，模型=%.3f秒"),
-            t2 - t1,
-            t3 - t2,
-            t4 - t3,
-            t5 - t4,
-        )
+        if report_status:
+            printt(
+                i18n("耗时：特征=%.3f秒，索引=%.3f秒，音高=%.3f秒，模型=%.3f秒"),
+                t2 - t1,
+                t3 - t2,
+                t4 - t3,
+                t5 - t4,
+            )
        return infered_audio.squeeze()
--- a/infer/vc/modules.py
+++ b/infer/vc/modules.py
@@ -19,6 +19,7 @@ from infer.vc.pipeline import Pipeline
 from infer.vc.utils import *
 from i18n.i18n import I18nAuto
 from tools.progress import batch_status, should_report
+from tools.cuda_graph import clear_cuda_graph_cache


 i18n = I18nAuto()
@@ -68,6 +69,8 @@ class VC:
                self.hubert_model is not None
            ):  # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
                logger.info(i18n("清理模型缓存"))
+                clear_cuda_graph_cache(self.net_g)
+                clear_cuda_graph_cache(self.hubert_model)
                del (self.net_g, self.n_spk, self.hubert_model, self.tgt_sr)  # ,cpt
                self.hubert_model = self.net_g = self.n_spk = self.hubert_model = (
                    self.tgt_sr
@@ -112,6 +115,9 @@ class VC:
        person = f'{os.getenv("weight_root")}/{sid}'
        logger.info("%s: %s", i18n("正在加载模型"), person)

+        if self.net_g is not None:
+            clear_cuda_graph_cache(self.net_g)
+
        self.cpt = torch.load(person, map_location="cpu")
        self.tgt_sr = self.cpt["config"][-1]
        self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0]  # n_spk
--- a/infer/vc/pipeline.py
+++ b/infer/vc/pipeline.py
@@ -15,6 +15,7 @@ import torch.nn.functional as F
 from scipy import signal

 from infer.hubert import extract_hubert_features
+from tools.cuda_graph import cuda_graph_enabled, run_cuda_graph

 bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)

@@ -217,11 +218,34 @@ class Pipeline(object):
        p_len = torch.tensor([p_len], device=self.device).long()
        with torch.no_grad():
            hasp = pitch is not None and pitchf is not None
-            arg = (feats, p_len, pitch, pitchf, sid) if hasp else (feats, p_len, sid)
-            audio1 = (net_g.infer(*arg)[0][0, 0]).data.cpu().float().numpy()
-            del hasp, arg
+            if hasp:
+                synthesized = run_cuda_graph(
+                    net_g,
+                    "rvc-synth-f0",
+                    lambda phone, lengths, coarse, continuous, speaker: net_g.infer(
+                        phone, lengths, coarse, continuous, speaker
+                    )[0],
+                    feats,
+                    p_len,
+                    pitch,
+                    pitchf,
+                    sid,
+                )
+            else:
+                synthesized = run_cuda_graph(
+                    net_g,
+                    "rvc-synth-no-f0",
+                    lambda phone, lengths, speaker: net_g.infer(
+                        phone, lengths, speaker
+                    )[0],
+                    feats,
+                    p_len,
+                    sid,
+                )
+            audio1 = synthesized[0, 0].data.cpu().float().numpy()
+            del hasp, synthesized
        del feats, p_len, padding_mask
-        if torch.cuda.is_available():
+        if torch.cuda.is_available() and not cuda_graph_enabled(self.device):
            torch.cuda.empty_cache()
        t2 = ttime()
        times[0] += t1 - t0
@@ -381,6 +405,6 @@ class Pipeline(object):
            max_int16 /= audio_max
        audio_opt = (audio_opt * max_int16).astype(np.int16)
        del pitch, pitchf, sid
-        if torch.cuda.is_available():
+        if torch.cuda.is_available() and not cuda_graph_enabled(self.device):
            torch.cuda.empty_cache()
        return audio_opt
--- a/realtime_gui.py
+++ b/realtime_gui.py
@@ -36,9 +36,10 @@ if __name__ == "__main__":
    import torch.nn.functional as F
    import torchaudio.transforms as tat

+    from configs.config import Config
    from infer import rtrvc as rvc_for_realtime
-    from i18n.i18n import I18nAuto
-    from configs.config import Config
+    from i18n.i18n import I18nAuto
+    from tools.cuda_graph import cuda_graph_enabled, run_cuda_graph

    i18n = I18nAuto()

@@ -64,9 +65,10 @@ if __name__ == "__main__":
            self.sg_output_device = ""

    class GUI:
-        def __init__(self) :
-            self.gui_config = GUIConfig()
-            self.config = Config()
+        def __init__(self) :
+            self.gui_config = GUIConfig()
+            self.config = Config()
+            printt("RVC_CUDA_GRAPH=%s", os.environ.get("RVC_CUDA_GRAPH", "0"))
            self.function = "vc"
            self.delay_time = 0
            self.hostapis = None
@@ -629,9 +631,16 @@ if __name__ == "__main__":
                dtype=torch.float32,
            )
            self.rms_buffer = np.zeros(4 * self.zc, dtype="float32")
-            self.sola_buffer = torch.zeros(
-                self.sola_buffer_frame, device=self.config.device, dtype=torch.float32
-            )
+            self.sola_buffer = torch.zeros(
+                self.sola_buffer_frame, device=self.config.device, dtype=torch.float32
+            )
+            self.sola_den_kernel = torch.ones(
+                1,
+                1,
+                self.sola_buffer_frame,
+                device=self.config.device,
+                dtype=torch.float32,
+            )
            self.nr_buffer = self.sola_buffer.clone()
            self.output_buffer = self.input_wav.clone()
            self.skip_head = self.extra_frame // self.zc
@@ -666,12 +675,85 @@ if __name__ == "__main__":
                ).to(self.config.device)
            else:
                self.resampler2 = None
-            self.tg = TorchGate(
-                sr=self.gui_config.samplerate, n_fft=4 * self.zc, prop_decrease=0.9
-            ).to(self.config.device)
-            self.start_stream()
-
-        def start_stream(self):
+            self.tg = TorchGate(
+                sr=self.gui_config.samplerate, n_fft=4 * self.zc, prop_decrease=0.9
+            ).to(self.config.device)
+            self.prewarm_cuda_graph()
+            self.start_stream()
+
+        def prewarm_cuda_graph(self):
+            if not cuda_graph_enabled(self.config.device):
+                return
+            try:
+                printt(i18n("正在预热CUDA Graph"))
+                samples = self.input_wav_res.shape[0]
+                phase = torch.arange(
+                    samples, device=self.config.device, dtype=torch.float32
+                )
+                probe = 0.05 * torch.sin(2 * np.pi * 220.0 * phase / 16000.0)
+                self.input_wav_res.copy_(probe)
+
+                if self.gui_config.I_noise_reduce:
+                    short = self.input_wav[
+                        -self.sola_buffer_frame - self.block_frame :
+                    ].unsqueeze(0)
+                    run_cuda_graph(
+                        self.tg,
+                        "realtime-input-noise-reduction",
+                        lambda short_audio, full_audio: self.tg(
+                            short_audio, full_audio
+                        ),
+                        short,
+                        self.input_wav.unsqueeze(0),
+                    )
+
+                resample_input = self.input_wav[-self.block_frame - 2 * self.zc :]
+                run_cuda_graph(
+                    self.resampler,
+                    "realtime-input-resample",
+                    lambda audio: self.resampler(audio),
+                    resample_input,
+                )
+
+                inferred = self.rvc.infer(
+                    self.input_wav_res,
+                    self.block_frame_16k,
+                    self.skip_head,
+                    self.return_length,
+                    self.gui_config.f0method,
+                )
+                if self.resampler2 is not None:
+                    inferred = run_cuda_graph(
+                        self.resampler2,
+                        "realtime-output-resample",
+                        lambda audio: self.resampler2(audio),
+                        inferred,
+                    )
+                if self.gui_config.O_noise_reduce:
+                    run_cuda_graph(
+                        self.tg,
+                        "realtime-output-noise-reduction",
+                        lambda short_audio, full_audio: self.tg(
+                            short_audio, full_audio
+                        ),
+                        inferred.unsqueeze(0),
+                        self.output_buffer.unsqueeze(0),
+                    )
+                torch.cuda.synchronize(self.config.device)
+                printt(i18n("CUDA Graph预热完成"))
+            except Exception:
+                printt(traceback.format_exc())
+            finally:
+                self.input_wav.zero_()
+                self.input_wav_denoise.zero_()
+                self.input_wav_res.zero_()
+                self.output_buffer.zero_()
+                self.sola_buffer.zero_()
+                self.nr_buffer.zero_()
+                self.rvc.cache_pitch.zero_()
+                self.rvc.cache_pitchf.zero_()
+
+        def start_stream(self):
            global flag_vc
            if not flag_vc:
                flag_vc = True
@@ -739,9 +821,13 @@ if __name__ == "__main__":
                    self.block_frame :
                ].clone()
                input_wav = self.input_wav[-self.sola_buffer_frame - self.block_frame :]
-                input_wav = self.tg(
-                    input_wav.unsqueeze(0), self.input_wav.unsqueeze(0)
-                ).squeeze(0)
+                input_wav = run_cuda_graph(
+                    self.tg,
+                    "realtime-input-noise-reduction",
+                    lambda short, full: self.tg(short, full),
+                    input_wav.unsqueeze(0),
+                    self.input_wav.unsqueeze(0),
+                ).squeeze(0)
                input_wav[: self.sola_buffer_frame] *= self.fade_in_window
                input_wav[: self.sola_buffer_frame] += (
                    self.nr_buffer * self.fade_out_window
@@ -750,15 +836,23 @@ if __name__ == "__main__":
                    : self.block_frame
                ]
                self.nr_buffer[:] = input_wav[self.block_frame :]
-                self.input_wav_res[-self.block_frame_16k - 160 :] = self.resampler(
-                    self.input_wav_denoise[-self.block_frame - 2 * self.zc :]
-                )[160:]
-            else:
-                self.input_wav_res[-160 * (indata.shape[0] // self.zc + 1) :] = (
-                    self.resampler(self.input_wav[-indata.shape[0] - 2 * self.zc :])[
-                        160:
-                    ]
-                )
+                resample_input = self.input_wav_denoise[
+                    -self.block_frame - 2 * self.zc :
+                ]
+                self.input_wav_res[-self.block_frame_16k - 160 :] = run_cuda_graph(
+                    self.resampler,
+                    "realtime-input-resample",
+                    lambda audio: self.resampler(audio),
+                    resample_input,
+                )[160:]
+            else:
+                resample_input = self.input_wav[-indata.shape[0] - 2 * self.zc :]
+                self.input_wav_res[-160 * (indata.shape[0] // self.zc + 1) :] = run_cuda_graph(
+                    self.resampler,
+                    "realtime-input-resample",
+                    lambda audio: self.resampler(audio),
+                    resample_input,
+                )[160:]
            # infer
            if self.function == "vc":
                infer_wav = self.rvc.infer(
@@ -767,9 +861,14 @@ if __name__ == "__main__":
                    self.skip_head,
                    self.return_length,
                    self.gui_config.f0method,
-                )
-                if self.resampler2 is not None:
-                    infer_wav = self.resampler2(infer_wav)
+                )
+                if self.resampler2 is not None:
+                    infer_wav = run_cuda_graph(
+                        self.resampler2,
+                        "realtime-output-resample",
+                        lambda audio: self.resampler2(audio),
+                        infer_wav,
+                    )
            elif self.gui_config.I_noise_reduce:
                infer_wav = self.input_wav_denoise[self.extra_frame :].clone()
            else:
@@ -780,9 +879,13 @@ if __name__ == "__main__":
                    self.block_frame :
                ].clone()
                self.output_buffer[-self.block_frame :] = infer_wav[-self.block_frame :]
-                infer_wav = self.tg(
-                    infer_wav.unsqueeze(0), self.output_buffer.unsqueeze(0)
-                ).squeeze(0)
+                infer_wav = run_cuda_graph(
+                    self.tg,
+                    "realtime-output-noise-reduction",
+                    lambda short, full: self.tg(short, full),
+                    infer_wav.unsqueeze(0),
+                    self.output_buffer.unsqueeze(0),
+                ).squeeze(0)
            # volume envelop mixing
            if self.gui_config.rms_mix_rate < 1 and self.function == "vc":
                if self.gui_config.I_noise_reduce:
@@ -813,20 +916,20 @@ if __name__ == "__main__":
                    mode="linear",
                    align_corners=True,
                )[0, 0, :-1]
-                rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-3)
-                infer_wav *= torch.pow(
-                    rms1 / rms2, torch.tensor(1 - self.gui_config.rms_mix_rate)
-                )
+                rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-3)
+                infer_wav *= torch.pow(
+                    rms1 / rms2, 1.0 - self.gui_config.rms_mix_rate
+                )
            # SOLA algorithm from https://github.com/yxlllc/DDSP-SVC
            conv_input = infer_wav[
                None, None, : self.sola_buffer_frame + self.sola_search_frame
            ]
            cor_nom = F.conv1d(conv_input, self.sola_buffer[None, None, :])
            cor_den = torch.sqrt(
-                F.conv1d(
-                    conv_input**2,
-                    torch.ones(1, 1, self.sola_buffer_frame, device=self.config.device),
-                )
+                F.conv1d(
+                    conv_input**2,
+                    self.sola_den_kernel,
+                )
                + 1e-8
            )
            if sys.platform == "darwin":
--- a/tools/cuda_graph.py
+++ b/tools/cuda_graph.py
@@ -0,0 +1,227 @@
+import logging
+import os
+import threading
+import time
+from collections import OrderedDict
+
+import torch
+
+
+logger = logging.getLogger(__name__)
+
+ENV_NAME = "RVC_CUDA_GRAPH"
+MAX_CACHE_ENV = "RVC_CUDA_GRAPH_MAX_CACHE"
+_probe_lock = threading.Lock()
+_probe_result = None
+
+
+def _device_type(device):
+    if isinstance(device, torch.device):
+        return device.type
+    return str(device).split(":", 1)[0].lower()
+
+
+def _cuda_device(device):
+    parsed = device if isinstance(device, torch.device) else torch.device(device)
+    if parsed.index is None:
+        parsed = torch.device("cuda", torch.cuda.current_device())
+    return parsed
+
+
+def _clone_output(value):
+    if torch.is_tensor(value):
+        return value.clone()
+    if isinstance(value, tuple):
+        return tuple(_clone_output(item) for item in value)
+    if isinstance(value, list):
+        return [_clone_output(item) for item in value]
+    if isinstance(value, dict):
+        return {key: _clone_output(item) for key, item in value.items()}
+    return value
+
+
+def detect_cuda_graph_support(device):
+    if _device_type(device) != "cuda" or not torch.cuda.is_available():
+        return False
+    if not hasattr(torch.cuda, "CUDAGraph") or not hasattr(torch.cuda, "graph"):
+        return False
+    cuda_device = _cuda_device(device)
+    try:
+        with torch.cuda.device(cuda_device):
+            current = torch.cuda.current_stream(cuda_device)
+            warmup = torch.cuda.Stream(device=cuda_device)
+            warmup.wait_stream(current)
+            with torch.cuda.stream(warmup):
+                probe = torch.arange(32, device=cuda_device, dtype=torch.float32)
+                for _ in range(3):
+                    expected = probe.square().add_(1)
+            current.wait_stream(warmup)
+            torch.cuda.synchronize(cuda_device)
+            graph = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(graph):
+                captured = probe.square() + 1
+            probe.copy_(torch.arange(32, device=cuda_device, dtype=torch.float32))
+            graph.replay()
+            torch.cuda.synchronize(cuda_device)
+            valid = torch.equal(
+                captured.cpu(), torch.arange(32, dtype=torch.float32).square() + 1
+            )
+            del captured, expected, graph, probe
+            return bool(valid)
+    except Exception:
+        logger.exception("CUDA Graph support probe failed on %s", cuda_device)
+        return False
+
+
+def configure_cuda_graph(device):
+    global _probe_result
+    explicit = os.environ.get(ENV_NAME)
+    if explicit in {"0", "1"}:
+        if explicit == "0":
+            return False
+        if _device_type(device) != "cuda":
+            os.environ[ENV_NAME] = "0"
+            return False
+    with _probe_lock:
+        if _probe_result is None:
+            _probe_result = detect_cuda_graph_support(device)
+        os.environ[ENV_NAME] = "1" if _probe_result else "0"
+    return bool(_probe_result)
+
+
+def cuda_graph_enabled(device):
+    return (
+        os.environ.get(ENV_NAME) == "1"
+        and _device_type(device) == "cuda"
+        and torch.cuda.is_available()
+    )
+
+
+def _tensor_signature(tensor):
+    return (
+        tuple(tensor.shape),
+        tuple(tensor.stride()),
+        str(tensor.dtype),
+        str(tensor.device),
+        bool(tensor.requires_grad),
+    )
+
+
+class _CapturedCall:
+    def __init__(self, function, inputs):
+        started = time.perf_counter()
+        self.lock = threading.RLock()
+        self.inputs = tuple(torch.empty_like(value) for value in inputs)
+        for static, value in zip(self.inputs, inputs):
+            static.copy_(value)
+        device = self.inputs[0].device
+        current = torch.cuda.current_stream(device)
+        warmup = torch.cuda.Stream(device=device)
+        warmup.wait_stream(current)
+        with torch.cuda.stream(warmup), torch.no_grad():
+            for _ in range(3):
+                output = function(*self.inputs)
+        current.wait_stream(warmup)
+        torch.cuda.synchronize(device)
+        self.graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(self.graph), torch.no_grad():
+            self.output = function(*self.inputs)
+        self.capture_ms = (time.perf_counter() - started) * 1000.0
+        self.done_event = None
+        del output
+
+    def replay(self, inputs):
+        with self.lock:
+            stream = torch.cuda.current_stream(self.inputs[0].device)
+            if self.done_event is not None:
+                stream.wait_event(self.done_event)
+            for static, value in zip(self.inputs, inputs):
+                static.copy_(value, non_blocking=True)
+            self.graph.replay()
+            output = _clone_output(self.output)
+            self.done_event = torch.cuda.Event(blocking=False)
+            self.done_event.record(stream)
+            return output
+
+
+class _GraphCache:
+    def __init__(self):
+        self.entries = OrderedDict()
+        self.failures = set()
+        self.lock = threading.RLock()
+        self.capture_count = 0
+        self.replay_count = 0
+        self.fallback_count = 0
+        self.eviction_count = 0
+        self.capture_ms = 0.0
+
+    def run(self, key, function, inputs):
+        signature = key + tuple(_tensor_signature(value) for value in inputs)
+        with self.lock:
+            if signature in self.failures:
+                self.fallback_count += 1
+                return function(*inputs)
+            entry = self.entries.get(signature)
+            if entry is None:
+                try:
+                    entry = _CapturedCall(function, inputs)
+                    self.entries[signature] = entry
+                    self.capture_count += 1
+                    self.capture_ms += entry.capture_ms
+                    max_entries = max(1, int(os.environ.get(MAX_CACHE_ENV, "8")))
+                    while len(self.entries) > max_entries:
+                        self.entries.popitem(last=False)
+                        self.eviction_count += 1
+                except Exception:
+                    self.failures.add(signature)
+                    self.fallback_count += 1
+                    logger.exception("CUDA Graph capture failed for %s; using eager", key)
+                    return function(*inputs)
+            else:
+                self.entries.move_to_end(signature)
+        output = entry.replay(inputs)
+        with self.lock:
+            self.replay_count += 1
+        return output
+
+
+def run_cuda_graph(owner, namespace, function, *inputs):
+    if not inputs or not cuda_graph_enabled(inputs[0].device):
+        return function(*inputs)
+    cache = getattr(owner, "_rvc_cuda_graph_cache", None)
+    if cache is None:
+        cache = _GraphCache()
+        setattr(owner, "_rvc_cuda_graph_cache", cache)
+    return cache.run((str(namespace),), function, tuple(inputs))
+
+
+def clear_cuda_graph_cache(owner):
+    cache = getattr(owner, "_rvc_cuda_graph_cache", None)
+    if cache is not None:
+        cache.entries.clear()
+        cache.failures.clear()
+        delattr(owner, "_rvc_cuda_graph_cache")
+
+
+def get_cuda_graph_stats(owner):
+    cache = getattr(owner, "_rvc_cuda_graph_cache", None)
+    if cache is None:
+        return {
+            "entries": 0,
+            "failures": 0,
+            "captures": 0,
+            "replays": 0,
+            "fallbacks": 0,
+            "evictions": 0,
+            "capture_ms": 0.0,
+        }
+    with cache.lock:
+        return {
+            "entries": len(cache.entries),
+            "failures": len(cache.failures),
+            "captures": cache.capture_count,
+            "replays": cache.replay_count,
+            "fallbacks": cache.fallback_count,
+            "evictions": cache.eviction_count,
+            "capture_ms": cache.capture_ms,
+        }
--- a/tools/uvr5/bsroformer.py
+++ b/tools/uvr5/bsroformer.py
@@ -9,6 +9,8 @@ import soundfile as sf
 import torch
 import torch.nn as nn
 import yaml
+
+from tools.cuda_graph import run_cuda_graph
 from tqdm import tqdm
 from tools.file_io import read_text
 from i18n.i18n import I18nAuto
@@ -178,7 +180,12 @@ class Roformer_Loader:
                    if len(batch_data) >= batch_size or (i >= mix.shape[1]):
                        arr = torch.stack(batch_data, dim=0)
                        # print(23333333,arr.dtype)
-                        x = model(arr)
+                        x = run_cuda_graph(
+                            model,
+                            "uvr-bsroformer",
+                            lambda audio: model(audio),
+                            arr,
+                        )

                        window = window_middle
                        if i - step == 0:  # First audio chunk, no fadein
--- a/tools/uvr5/lib/utils.py
+++ b/tools/uvr5/lib/utils.py
@@ -1,5 +1,6 @@
 import numpy as np
 import torch
+from tools.cuda_graph import run_cuda_graph
 from tqdm import tqdm


@@ -34,7 +35,12 @@ def inference(X_spec, device, model, aggressiveness, data):
                    X_mag_window = X_mag_window.half()
                X_mag_window = X_mag_window.to(device)

-                pred = model.predict(X_mag_window, aggressiveness)
+                pred = run_cuda_graph(
+                    model,
+                    "uvr-vr-%s" % repr(aggressiveness),
+                    lambda window: model.predict(window, aggressiveness),
+                    X_mag_window,
+                )

                pred = pred.detach().cpu().numpy()
                preds.append(pred[0])
--- a/webui.py
+++ b/webui.py
@@ -25,6 +25,7 @@ for name in os.listdir(tmp):
    except Exception as error:
        print(str(error))

+from configs.config import Config, GPU_INDEX, GPU_INFOS, GPU_MEMORY, IS_GPU
 from infer.vc.modules import VC
 from tools.uvr5.webui import uvr
 from tools.file_io import read_text
@@ -35,7 +36,6 @@ from train.process_ckpt import (
    show_info,
 )
 from i18n.i18n import I18nAuto
-from configs.config import Config, GPU_INDEX, GPU_INFOS, GPU_MEMORY, IS_GPU
 import torch, platform
 import numpy as np
 import gradio as gr
@@ -143,6 +143,7 @@ print(
    i18n("当前设备：%s | 推理精度：%s") % (config.device, config.dtype),
    flush=True,
 )
+logger.info("RVC_CUDA_GRAPH=%s", os.environ.get("RVC_CUDA_GRAPH", "0"))
 # GPU filtering and precision rules are shared with inference/extraction/training.
 gpu_infos = list(GPU_INFOS)
 gpu_indices = sorted(GPU_INDEX)
@@ -394,6 +395,10 @@ def train_task_stopped(state):

 def start_train_process(state, cmd):
    kwargs = {"shell": True, "cwd": now_dir}
+    if "train/train.py" in cmd.replace("\\", "/"):
+        training_env = os.environ.copy()
+        training_env["RVC_CUDA_GRAPH"] = "0"
+        kwargs["env"] = training_env
    if platform.system() == "Windows":
        kwargs["creationflags"] = subprocess.CREATE_NEW_PROCESS_GROUP
    else: