@@ -377,7 +377,6 @@ def run(self) -> None:
377
377
if isinstance (ext , CMakeExtension ):
378
378
print (f"Building CMake extension { ext .name } " )
379
379
with tempfile .TemporaryDirectory () as build_dir :
380
- build_dir = '/code/pr-graph-api/v5/TransformerEngine/build'
381
380
build_dir = Path (build_dir )
382
381
package_path = Path (self .get_ext_fullpath (ext .name ))
383
382
install_dir = package_path .resolve ().parent
@@ -497,8 +496,8 @@ def setup_pytorch_extension() -> setuptools.Extension:
497
496
else :
498
497
if version >= (11 , 2 ):
499
498
nvcc_flags .extend (["--threads" , "4" ])
500
- # if version >= (11, 0):
501
- # nvcc_flags.extend(["-gencode", "arch=compute_80,code=sm_80"])
499
+ if version >= (11 , 0 ):
500
+ nvcc_flags .extend (["-gencode" , "arch=compute_80,code=sm_80" ])
502
501
if version >= (11 , 8 ):
503
502
nvcc_flags .extend (["-gencode" , "arch=compute_90,code=sm_90" ])
504
503
@@ -548,8 +547,8 @@ def setup_paddle_extension() -> setuptools.Extension:
548
547
cxx_flags = ["-O3" ]
549
548
nvcc_flags = [
550
549
"-O3" ,
551
- # "-gencode",
552
- # "arch=compute_70,code=sm_70",
550
+ "-gencode" ,
551
+ "arch=compute_70,code=sm_70" ,
553
552
"-U__CUDA_NO_HALF_OPERATORS__" ,
554
553
"-U__CUDA_NO_HALF_CONVERSIONS__" ,
555
554
"-U__CUDA_NO_BFLOAT16_OPERATORS__" ,
@@ -569,8 +568,8 @@ def setup_paddle_extension() -> setuptools.Extension:
569
568
else :
570
569
if version >= (11 , 2 ):
571
570
nvcc_flags .extend (["--threads" , "4" ])
572
- # if version >= (11, 0):
573
- # nvcc_flags.extend(["-gencode", "arch=compute_80,code=sm_80"])
571
+ if version >= (11 , 0 ):
572
+ nvcc_flags .extend (["-gencode" , "arch=compute_80,code=sm_80" ])
574
573
if version >= (11 , 8 ):
575
574
nvcc_flags .extend (["-gencode" , "arch=compute_90,code=sm_90" ])
576
575
0 commit comments