* fix generate_kernels.py in CUDA 12.0

* fix attrs bug

* first version, notest

* return final rst, notest

* use infinity() instead of max

* ut structure

* start up of ut

* generate lse

* update

* add depense

* reconstruct cmake

* move file

* add memory efficient attention and fix blasimpl

* update

* update cmake

* add namespace

* update cmake

* use .cu

* update for pad3d

* bug fix

* bug fix

* update

* bug fix

* update enforce

* add test case

* merge the lse pad

* fix kernel_fn of backward

* fix PADDLE_ENFORCE_EQ and phi_api

* fix PADDLE_ENFORCE

* fix PADDLE_ENFORCE

* rerun coverage

* fix memory efficient attention test

* rerun ci

* add cuda version condition

* add cuda version condition

* delete WIP test

* replace PADDLE_ENFORCE

* edit the namespace of datatype in multiple.cc

* rerun

* rerun

---------
Co-authored-by: Nliuyuang <liuyuang@baidu.com>

* cmake-lint

* update

* update ampere sm

* update ampere sm

* update ampere sm

* Adjust GPU Arches for Whl releases

* Adjusted CUDA arches

* fixed minor issue

* adjusted gpu arches

[Background]
Expansion in code size can be irreversible in the long run, leading to huge release packages which
not only hampers user experience but also exceeds a hard limit of pypi.

In such, NV_FATBIN section takes up 86% of the compiled dylib size, owing to the vast number of GPU
arches supported.

This PR aims to prune this NV_FATBIN.

[Solution]
In the new release strategy, two types of whl packages will be involved:

Cubin PIP package:
PIP package maintains a smaller window for GPU arches support, containing
sm_60, sm_70, sm_75, sm_80 cubins, covering Pascal - Ampere arches

JIT release package:
This is a backup for Cubin PIP package, containing compute_35, compute_50, compute_60,
compute_70, compute_75, compute_80, with best performance and GPU arches coverage.

However, it takes around 10 min to install due to the JIT compilation.

[How to use]
The new release strategy is disabled by default.
To compile for Cubin PIP package, add this to cmake: -DCUBIN_RELEASE_PIP
To compile for JIT release package, add this to cmake: -DJIT_RELEASE_WHL

* Support sccache to speed up compilation on Windows

* Support sccache to speed up compilation on Windows

* add gpu implementation of shuffle batch
test=develop

* add thrust cuda patches
test=develop

* fix macro guard

* fix shuffle batch compile on windows/hip

* fix hip compilation error

* refine CMakeLists.txt

* fix windows compile error

* try to fix windows CI compilation error

* fix windows compilation again

* fix shuffle_batch op test on Windows

* fix ninja compilation bug on windows

* polish windows ci

* polish windows ci

* update compilation with C++14, test=develop

* fix compilation error in eigen, test=develop

* Modify relu native implementation

* fix GPU performance

* remove -DSUPPORTS_CUDA_FP16 in cuda.cmake

* comile with cuda9

* add some unittest

* notest;test=coverage

* add unittest for trt plugin swish && split

* update ernie unittest

* fix some error message

* remove repeated judgement of CUDA version in mbEltwiseLayerNormOpConverter

* fix comile errror when CUDA_ARCH_NAME < Pascal"

* fix comile error

* update unittest timeout

* compile with cuda9

* update error msg

* fix code style

* add some comments

* add define IF_CUDA_ARCH_SUPPORT_FP16

* rename IF_CUDA_ARCH_SUPPORT_FP16 to CUDA_ARCH_FP16_SUPPORTED

specify cuda arch when dectected fail

* Fixed compile warning about incorrect compile options,fix paddle_build.bat

* fix paddle_build.bat to more safe

fix bug in CUDA_NVCC_FALS and CMAKE_CUDA_FLAGS

* add -DPADDLE_CUDA_BINVER. test=develop, test=win_gpu

* nvcc will use add_compile_options, avoid using it if you don't want to pass arguments to nvcc. test=develop

* test=develop, test=win_gpu

* support CUDA using cmake built-in way (#24395)

* support CUDA using cmake built-in way. test=develop

* test=develop

* cmake_minimum_required 3.10

* test=develop

This reverts commit 068d3690.

* support CUDA using cmake built-in way. test=develop

* test=develop