feat: add support for deepgemm sglang

This commit is contained in:
2025-12-27 11:07:56 +00:00
parent d763aaf304
commit d4ffb6a9bf
2 changed files with 11 additions and 4 deletions
+4 -3
View File
@@ -56,6 +56,7 @@ set_compile_mode = _wrap_op('set_compile_mode')
get_compile_mode = _wrap_op('get_compile_mode')
set_tc_util = _wrap_op('set_tc_util')
get_tc_util = _wrap_op('get_tc_util')
preload_kernels = _wrap_op('preload_kernels')
fp8_gemm_nt = _wrap_op('fp8_gemm_nt')
fp8_gemm_nn = _wrap_op('fp8_gemm_nn')
@@ -109,7 +110,7 @@ from .utils import *
def _verify_ops_loaded():
expected_ops = [
'init', 'set_num_sms', 'get_num_sms', 'set_tc_util', 'get_tc_util',
'init', 'preload_kernels', 'set_num_sms', 'get_num_sms', 'set_tc_util', 'get_tc_util',
'fp8_gemm_nt', 'fp8_gemm_nn', 'fp8_gemm_tn', 'fp8_gemm_tt',
'm_grouped_fp8_gemm_nt_contiguous', 'm_grouped_fp8_gemm_nn_contiguous',
'm_grouped_fp8_gemm_nt_masked', 'k_grouped_fp8_gemm_nt_contiguous',
@@ -141,10 +142,10 @@ if __debug__:
# Preload cached kernels if enabled via environment variable
if os.environ.get('DG_PRELOAD_KERNELS', '0') == '1':
try:
deep_gemm_cpp.preload_kernels()
preload_kernels()
except Exception as e:
# Preloading failure should not block initialization
import warnings
warnings.warn(f"Failed to preload DeepGEMM kernels: {e}")
__version__ = '2.1.2'
__version__ = '2.2.1'