Tianqi Ren
Tianqi Ren
这个可能需要后端检查UA然后决定返回nginx文件索引
之前没有写过nginx主题，可能得等我先学学（
I have tested PyTorch 2.5 under Ubuntu WSL2. It seems that `ats-m150` works for `dg2-g10` devices, such as A770. `dg2-g11` devices like A380 are not supported., and recompilation with `TORCH_XPU_ARCH_LIST=dg2-g11`...
> @iori2333 : Can you, please, comment which behavior you observe? what means "not supported"? 1. Training using torch.float32 works fine, but initializing costs a lot of time. This is...
I made a simple implementation [here](https://github.com/iori2333/mmengine). I got it tested with my Arc A770 using unmodified [mmsegmentation](https://github.com/open-mmlab/mmsegmentation) and it worked fine.
我使用glm-4-9b-chat在int8、fp16量化时也遇到了相同的问题
Maybe you can try [PyTorch 2.5](https://pytorch.org/docs/stable/notes/get_start_xpu.html) that supports Ultra 7 155H, and nightly build binaries for Windows are provided as well.