venv/lib/python3.12/site-packages/transformers/kernels/yoso/common_cuda.h

#define MAX_THREADS_PER_BLOCK 1024
#define OPTIMAL_THREADS_PER_BLOCK 256
#define WARP_SIZE 32
#define MAX_NUM_BLOCK_X 2147483647
#define MAX_NUM_BLOCK_Y 65535
#define MAX_NUM_BLOCK_Z 65535
#define MAX_SHARED_MEM_PER_BLOCK 48000
#define FULL_MASK 0xffffffff