Enable DeepGEMM PDL on by default (#23979)
Co-authored-by: b8zhong <b8zhong@users.noreply.github.com> Co-authored-by: Brayden Zhong <brayden@radixark.ai>
This commit is contained in:
co-authored by
b8zhong
Brayden Zhong
parent
cd98d97037
commit
5bf90ad988
@@ -483,6 +483,7 @@ class Envs:
|
||||
SGLANG_DG_USE_NVRTC = EnvBool(False)
|
||||
SGLANG_USE_DEEPGEMM_BMM = EnvBool(False)
|
||||
SGLANG_DEEPGEMM_SANITY_CHECK = EnvBool(False)
|
||||
SGLANG_DEEPGEMM_PDL = EnvBool(True)
|
||||
SGLANG_PP_PARALLEL_DEEPGEMM_WARMUP = EnvBool(False)
|
||||
|
||||
# DeepSeek MHA Optimization
|
||||
|
||||
@@ -20,6 +20,9 @@ if ENABLE_JIT_DEEPGEMM:
|
||||
import deep_gemm
|
||||
from deep_gemm.utils.layout import get_mn_major_tma_aligned_tensor # noqa: F401
|
||||
|
||||
if envs.SGLANG_DEEPGEMM_PDL.get() and hasattr(deep_gemm, "set_pdl"):
|
||||
deep_gemm.set_pdl(True)
|
||||
|
||||
_SANITY_CHECK = envs.SGLANG_DEEPGEMM_SANITY_CHECK.get()
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user