Xiang Lyu

Results 202 comments of Xiang Lyu

> I also met with this issue but cannot solve it. is there any body could help? It is very easy to modify this problem. Just debug the code at...

@zhezhaoa 你是指Co-occurrence Type为Word → Character (1)的Context Word Vectors嘛?还是说任意一个Context Word Vectors都可以呢?

> Hello, > > Thanks for asking. > > First, I was wondering why are you using `torchrun` instead of `python -m torch.distributed.launch ...` ? Are there any reasons for...

150ms请使用阿里云,而且是预训练音色不需要提取token和embedding情况下。开源版本是python且zero shot推理,跟150ms肯定是有差距的,但是模型方案是一模一样的,有能力的可以先进行工程话,开源后续可能会更新一个llm用vllm推理

> 用3秒克隆代码保存的pth文件做推理测试,相同的一句话“你好,请稍等,我正在思考您的问题…”,从fastapi里的server启动服务,推理这句话。300M推理速度是2秒,0.5B是8秒。 是否是在相同的prompt下进行对比,cosyvoice和coysvoice2初始化时,都指定load_jit等为false

> torch.compile()可以加速一下吗,有没有方案 我们现在提供的是与我们阿里云一致的引擎导出方案(jit和trt),其他的没有做过尝试

1. use mode. add_zero_shot_spk, check readme for usage 2. no, for new language, you need to prepare corresponding data and do sft training

no, we only support sentence level stream input

可以把合成结果发一下,评估一下是否可以修复