BeerTai

Hi, thank you very much for your code. Will you still upload a friendly version of the code and documentation? Thanks~

How to terminate the experiment

9

How can I terminate the following state. ``` Experiments could take a while There are 32 other tasks pending before your latest task on 12 Worker(s). ```

您好，感谢您的代码，在crf.py中的_forward_alg函数里面有一行代码不太明白： `log_alpha = (log_sum_exp_batch(self.transitions + log_alpha, axis=-1) + feats[:, t]).unsqueeze(1)` 在_forward_alg函数中的log_alpha处，为什么在函数log_sum_exp_batch的外面加feats[:, t]，感觉在pytorch官方非batch版本中，是先加feats再进行log_sum_exp的啊，batch版本有什么不同吗？谢谢您的回复~

用DeepSpeed训练时，会保存一些.pt文件

2

您好，我在用DeepSpeed进行多卡训练时，会保存一些zero_pp_rank_XX_mp_rank_XX_model_states.pt文件，文件夹有20多个G，导致存储空间不足。这应该如何处理？ 而用accelerate进行多卡训练时，会报错： ··· File "/root/miniconda3/envs/llm/lib/python3.8/site-packages/peft/tuners/lora.py", line 565, in forward result = F.linear(x, transpose(self.weight, self.fan_in_fan_out), bias=self.bias) RuntimeError: CUDA error: CUBLAS_STATUS_NOT_INITIALIZED when calling "cublasCreate(handle)" ··· 训练脚本是： ··· accelerate launch src/train_sft.py \ --do_train...

Results in paper

1

Hi，how were the results in the paper obtained? The maximum value of the sum of each batch or the maximum value of each dataset?

[BFCL] Adding UniGPT model handler

1

This PR adds model handler for UniGPT2.0 and UniGPT4.0. UniGPT series models are authored by UniSound. I have sent the relevant information of api to [[email protected]](mailto:[email protected])