EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Temporay work-around for rocm 7.0.0 alpha with enabling data-parallel issue (#10434)

Browse Source 
Co-authored-by: wunhuang <wunhuang@amd.com>
Co-authored-by: Sai Enduri <saimanas.enduri@amd.com>
This commit is contained in:
kk
2025-09-16 10:08:04 +08:00
committed by GitHub
parent a220c14f81
commit ec272dda9c
1 changed files with 10 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/layers/dp_attention.py
View File

@@ -17,6 +17,7 @@ from sglang.srt.distributed import (
get_tp_group, get_tp_group,
tensor_model_parallel_all_reduce, tensor_model_parallel_all_reduce,
) )
from sglang.srt.utils import get_bool_env_var, is_hip
if TYPE_CHECKING: if TYPE_CHECKING:
from sglang.srt.configs.model_config import ModelConfig from sglang.srt.configs.model_config import ModelConfig
@@ -36,6 +37,9 @@ _LOCAL_ATTN_DP_SIZE: Optional[int] = None
_LOCAL_ATTN_DP_RANK: Optional[int] = None _LOCAL_ATTN_DP_RANK: Optional[int] = None
_ENABLE_DP_ATTENTION_FLAG: bool = False _ENABLE_DP_ATTENTION_FLAG: bool = False
_is_hip = is_hip()
_USE_ROCM700A_WA = _is_hip and get_bool_env_var("SGLANG_USE_ROCM700A")
class DpPaddingMode(IntEnum): class DpPaddingMode(IntEnum):
@@ -67,7 +71,12 @@ class DpPaddingMode(IntEnum):
@classmethod @classmethod
def get_default_mode_in_cuda_graph(cls) -> DpPaddingMode: def get_default_mode_in_cuda_graph(cls) -> DpPaddingMode:
return cls.MAX_LEN # TODO(kkhuang-amd): noqa, temporary work-around for rocm 7.0.0 alpha
# it can be safely removed later, once RCCL fixed
if _USE_ROCM700A_WA:
return cls.SUM_LEN
else:
return cls.MAX_LEN
class _DpGatheredBufferWrapper: class _DpGatheredBufferWrapper: