transformers-benchmarks

transformers-benchmarks copied to clipboard

…16-bit weight, 'should be 'with 16-bit'

fanjinglian

Exceptionally high memory bandwidth

2

我在 RTX 4090 工作站上运行 banchmark 程序，取得了异常高的总线带宽数据： ``` bash Pytorch version : 1.14.0a0+44dac51 CUDA version : 12.0 GPU : NVIDIA GeForce RTX 4090 Matrix Multiplication: n=128 n=512 n=2048 n=8192 torch.float32 1.048...

MingruiZhuang

AttributeError: 'NoneType' object has no attribute 'samples_per_sec'

2

  I try to run your "transformers.ipynb" with python. But it had an error. How can I solve it?

michelleqyhqyh

About the theoretical value of the GPU

2

请问沐神： 1. 在notebook中指向的wiki里，3090ti的理论值40是从表中的core boosted value（[39.997](https://en.wikipedia.org/wiki/GeForce_30_series#:~:text=33.546-,(39.997),-0.524%0A(0.625))）得到的吗？ 2. 我在自己的很多块3090上，用CUDA11.7和nvidia-driver 525跑出来的TFLOPS都只有24，距离base（29.3）和boost（35.6）的理论值都有一定的差距。请问notebook中用3090ti跑的TFLOPS是经过超频的吗？要想达到接近理论值的FLOPS需要做怎样的设置呢？

fingertap

the transformers library has changed its internal structure, and BertLayer is no longer accessible through the old import paths.

Saibo-creator

Hi thank you for the great benchmark as well as the videos explaining papers! Currently the table/explanations says that actual matrix multiplication value is much smaller than theoretical one. But...

fzyzcjy