spicysama
We've supported the inference for mps devices
Theoretically any cpu device will work.
You can see here : https://speech.fish.audio/#macos
[audio (22).zip](https://github.com/user-attachments/files/16318732/audio.22.zip)  可以试试这个参数
说明一下你的硬件情况
it should be ok now.
I guess your os is linux now? This is a problem caused by torch 2.4.0 implementation. Downgrade torch to version ~ 2.3.0
Well, that's ok though. Before torch~=2.3.0 it's changable for input tensor shape without re-compiling. Thanks for your exploration. We'll consider if we can change the code.
```bash %cd ~/autodl-tmp/workdir/fish-speech !python -m tools.webui \ --llama-config-name dual_ar_2_codebook_medium \ --decoder-checkpoint-path "output/vits_decoder/step_000005000.ckpt" \ --decoder-config-name vits_decoder_finetune \ --compile ``` 用这个