wangshuai09

Thanks for your reply! Here is the training loss of weight_cross_entropy where pos_weight = beta/(1-beta) passing to F.binary_cross_entropy_with_logits as weight param， ---------------------------------------------------------------------- ``` Fri Apr 16 17:54:21 2021 Epoch: 0...

If there are no video card and NPU, the `torch_command` will not change. https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cf2772fab0af5573da775e7437e6acdca424f26e/modules/launch_utils.py#L318 It will use cpu as backend and here is screen shoot running with a downloaded embedding,...

Sorry for my misunderstanding. It will print error on screen using `elif eval "npu-smi info`. Your advise is so great.

This problem is that torch_npu don't support multi-card communication within a single process. And this will be fixed in the official torch-npu&cann release at the end of April. FYI https://github.com/huggingface/accelerate/issues/2368....

FYI https://github.com/lm-sys/FastChat/issues/3237

`Issue 6: slow!!` has been initially fixed by memory reuse, almost speedup 10x. Also, there are other space for optimization. Current inference speed: 

@AllernChen you can follow [ggml-cann develop](https://github.com/hipudding/llama.cpp/issues/4) and we recommend use the latest CANN version. I'm using `8.0.RC2.alpha001` to develop.

Rope operator has been fixed.

i just set the "segmentation" = [[1,1,1,1,1,1]] ,and during the training ,the mask loss is 0.0

Hi @nttstar, thank you for this great job. Because the onnxruntime is officially support Ascend with CANN EP, it does not modify too much code to adapt this project. May...