chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Ascend] qwen optimization (#12078)

Browse Source 
Co-authored-by: Even Zhou <even.y.zhou@outlook.com>
This commit is contained in:
Liwansi
2025-11-25 19:44:24 +08:00
committed by GitHub
parent 0b3f002daf
commit 432ecf841e
16 changed files with 562 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/server_args.py
View File

@@ -404,7 +404,7 @@ class ServerArgs:
# Expert parallelism
ep_size: int = 1
moe_a2a_backend: Literal["none", "deepep", "mooncake"] = "none"
moe_a2a_backend: Literal["none", "deepep", "mooncake", "ascend_fuseep"] = "none"
moe_runner_backend: str = "auto"
flashinfer_mxfp4_moe_precision: Literal["default", "bf16"] = "default"
enable_flashinfer_allreduce_fusion: bool = False
@@ -1516,6 +1516,12 @@ class ServerArgs:
f"Mooncake MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
)
if self.moe_a2a_backend == "ascend_fuseep":
self.ep_size = self.tp_size
logger.warning(
f"Ascend fused EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
)
def _handle_eplb_and_dispatch(self):
if self.enable_eplb and (self.expert_distribution_recorder_mode is None):
self.expert_distribution_recorder_mode = "stat"
@@ -1523,7 +1529,7 @@ class ServerArgs:
"EPLB is enabled. The expert_distribution_recorder_mode is automatically set."
)
if (self.enable_eplb or (self.init_expert_location is not None)) and (
if (self.enable_eplb or (self.init_expert_location != "trivial")) and (
self.ep_dispatch_algorithm is None
):
self.ep_dispatch_algorithm = "static"
@@ -2969,7 +2975,7 @@ class ServerArgs:
parser.add_argument(
"--moe-a2a-backend",
type=str,
choices=["none", "deepep", "mooncake"],
choices=["none", "deepep", "mooncake", "ascend_fuseep"],
default=ServerArgs.moe_a2a_backend,
help="Choose the backend for MoE A2A.",
)