Yibo Jin comments

Results 13 comments of


                                            Yibo Jin

yolov4 稀疏化训练mAP与darknet常规训练获得的差距大

遇到类似问题，请问有找到办法吗？

CodeLlama-7B int4-awq get error of "The value updated is not the same shape as the original. "

> `Assertion failed: mpiSize == tp * pp ` Did you run with mpi? Hi, how did you solve this error? I met the same problem

4卡L20,A100对YI-34B进行lora训练，报CUDA错误RuntimeError: CUDA error: device-side assert triggered

> base 模型请使用 default 模板这里使用的是chat模型，default也试验过，有同样的问题

为什么BELLE经过GPTQ量化（8bit/4bit）后，模型的推理速度变慢了很多呢

> > 请问量化后速度变慢的原因有找到么，我这边对比量化前后的速度，同等环境下量化后的速度反而更慢一些了 > > 同样发现了，不知道原因可能是量化后插入的节点没法调用cuda的kernel导致的吧

[Bug] Failed to convert fcos-resnet50 to onnx

Hi, could anyone help?

How to load a LoRA for inference?

Hi~ Is that okay to load a LoRA and do quantization works?How should I set for that

ValueError: The model did not return a loss from the inputs

Hi, I met the same problem. Have you get rid of this?

Can not install with 2080ti

> @wanghongtai92: You will have to write custom CUDA kernels which, I assume is a tall order. Hi, I wonder would that be successful for cards with SM_75 like Tesla...

Nan or Infs when using llama-13B-chat

Hi~I have seen the same error recently. Is there any conclusion now? @jamesdborin

Nan or Infs when using llama-13B-chat

It seems that torch.half() is not enough to complete the inference process and a higher precision numerical format needs to be used, like torch.bf16 or sth else