| 6 | import subprocess, ctypes, torch, os, sys |
| 7 | |
| 8 | class CUtensorMap(ctypes.Structure): |
| 9 | _fields_ = [("data", ctypes.c_ubyte * 128)] |
| 10 | |
| 11 | BM, BN, BK = 128, 256, 64 # BM=128 with 2 math warpgroups |
| 12 | libcuda = ctypes.CDLL("libcuda.so.1") |
no outgoing calls