faster tranformer cause "cross_check False " about fastertransformer

[wykdg]

addition
python 3.6 with conda

[IrishCoffee]

You are testing fp16 or fp32?

For fp16, if the diff number is something like 0.02, the answer is right.

[wykdg]

fp32，and the max diff is 3.
I compile the code without any modify.

and run the transformer_fp32.py,it returns cross check fail

[IrishCoffee]

What commands do you use to run the transformer_fp32.py?

Did you specify the -DSM when compiling the code? Did you generate a new gemm_config.in?

[wykdg]

gemm_config.in was generate by gemmfp32
DSM i have try 60 61 70 75
I run the check with commond "transformer_fp32.py 2 2 2 2 2" .And some other parmamters, but all of them are wrong ,except 1 1 1 1 1 .

[IrishCoffee]

Please try with some common configurations. For example, python transformer_fp32.py 1 12 128 12 64 to test the case batch size = 1, seq_len = 128, 12 layers with bert base.

[IrishCoffee]

2 2 2 2 2 are meaningless parameters. Especially the settings of the head_num and size_per_head.

[wykdg]

here is the result of "python transformer_fp32.py 1 12 128 12 64 "
cross_check False
max diff 4.412327
min diff 6.79539e-06

maybe there is something wrong with abi ?

[IrishCoffee]

What GPU do you use? What DSM values did you set?

What's the content of the gemm_config.in you generated?

[wykdg]

GPU is tesla M40,Driver Version: 418.67
DSM, i had try 60 61 70
gemm_config.in is
"13
13
13
6
2
"

[IrishCoffee]

The compute capability of M40 is 52. You can have a try but I am not sure whether the code can be executed correctly on M40.