Fix test_fused_dropout_act_bias failure on H100

[Wong4j]

PR types

Bug fixes

PR changes

Others

Describe

test_fused_dropout_act_bias UT can pass on A100 but failed on H100. Adding __launch_bounds__(THREADS_PER_CTA) to FusedDropoutActBiasGrad kernel can solve the bug.
I have already reported this bug internally. I guess the nvcc compiler doesn't allocate resources properly for Hopper. This bug may also occur in H800. Thus, I file this PR.

Compute-sanitizer error message:

========= COMPUTE-SANITIZER
========= Program hit cudaErrorLaunchOutOfResources (error 701) due to "too many resources requested for launch" on CUDA API call to cudaLaunchKernel.
=========     Saved host backtrace up to driver entry point at error
=========     Host Frame: [0x454676]
=========                in /lib/x86_64-linux-gnu/libcuda.so.1
=========     Host Frame:cudaLaunchKernel [0x6d4a8]
=========                in /test-hopper/./test.out
=========     Host Frame:cudaError cudaLaunchKernel<char>(char const*, dim3, dim3, void**, unsigned long, CUstream_st*) [0xb777]
=========                in /test-hopper/./test.out
=========     Host Frame:__device_stub__Z23FusedDropoutActBiasGradIdEv15GeluGradFunctorIT_EPKS1_PKhS4_S4_S1_llPS1_S7_(GeluGradFunctor<double>&, double const*, unsigned char const*, double const*, double const*, double, long, long, double*, double*) [0xb5ad]  
=========                in /test-hopper/./test.out
=========     Host Frame:void __wrapper__device_stub_FusedDropoutActBiasGrad<double>(GeluGradFunctor<double>&, double const*&, unsigned char const*&, double const*&, double const*&, double const&, long const&, long const&, double*&, double*&) [0xb64c]        
=========                in /test-hopper/./test.out
=========     Host Frame:void FusedDropoutActBiasGrad<double>(GeluGradFunctor<double>, double const*, unsigned char const*, double const*, double const*, double, long, long, double*, double*) [0xb849]

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zkh2016]

LGTM