• W
    Use cuda virtual memory management and merge blocks (#36189) · a1ec1d5a
    wanghuancoder 提交于
    * Use cuda virtual memory management and merge blocks, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * window dll, test=develop
    
    * fix cuda error of CUDA_ERROR_NOT_INITIALIZED, test=develop
    
    * use autogrowthv2 for system allocator, test=develop
    
    * remove ~CUDAVirtualMemAllocator(), test=develop
    
    * refine, test=develop
    
    * fix cuda error of CUDA_ERROR_NOT_INITIALIZED, test=develop
    
    * fix cuda error of CUDA_ERROR_NOT_INITIALIZED, test=develop
    
    * fix bug, test=develop
    
    * revert system allocator, test =develop
    
    * revert multiprocessing, test=develop
    
    * fix AutoGrowthBestFitAllocatorV2 mutxt, test=develop
    
    * catch cudaErrorInitializationError when create allocator, test=develop
    
    * fix cuMemSetAccess use, test=develop
    
    * refine cuda api use, test=develop
    
    * refine, test=develop
    
    * for test, test=develop
    
    * for test, test=develop
    
    * switch to v2, test=develop
    
    * refine virtual allocator, test=develop
    
    * Record cuMemCreate and cuMemRelease, test=develop
    
    * refine, test=develop
    
    * avoid out of bounds, test=develop
    
    * rename allocator, test=develop
    
    * refine, test=develop
    
    * use PADDLE_ENFORCE_CUDA_SUCCESS, test=develop
    
    * for test,test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    
    * refine, test=develop
    a1ec1d5a
allocator_facade.cc 18.2 KB