deyituo
deyituo
@ZFTurbo how do you collect the songs?
en... I wonder the diff using much smaller dataset with less speakers?
Add real/imag L1 loss can solve it.
dataset is essential..
have a dev group ?
@qthequartermasterman v0 engine cannot use some optimizations(enable_prefix_caching， enable_chunked_prefill， use_cudagraph, etc.), and is much slower than v1 pormpt tokens. When will the feature can be preview or PR?
@qthequartermasterman will this modifications support these features? https://github.com/vllm-project/vllm/compare/main...qthequartermasterman:vllm:inputs_embeds_in_v1_outline
@qthequartermasterman has a PR now？
微调在同样设置情况下，cv1和cv2的mel、f0微调曲线有较大差异  ``` # set random seed, so that you may reproduce your result. __set_seed1: !apply:random.seed [1986] __set_seed2: !apply:numpy.random.seed [1986] __set_seed3: !apply:torch.manual_seed [1986] __set_seed4: !apply:torch.cuda.manual_seed_all [1986] # fixed params sample_rate:...