PyTorch 线程亲和性测试：CUDA 上下文绑定的惊人代价

pytorch 主线程推理：

import time import queue import threading import torchaudio from cosyvoice.cli.cosyvoice import AutoModel class CosyVoiceQueueTTS: def __init__(self, model_dir, fp16=False, target_sr=16000): self.cosyvoice = AutoModel(model_dir=model_dir, fp16=fp16) self.sample_rate = self.cosyvoice.sample_rate # 原始采样率 (24000) self.target_sr = target_sr # 目标采样率 (16000) # 创建重采样器（如果目标采样率不同） if self.target_sr != self.sample_rate: self.resampler = torchaudio.transforms.Resample( orig_freq=self.sample_rate, new_freq=self.target_sr ) else: self.resampler = None self.running = False