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

Fix the moe padding conditional logic (#4081)

Browse Source
This commit is contained in:
HAI
2025-03-05 10:56:51 -08:00
committed by GitHub
parent d3d4d76758
commit 71ab0dabe0
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
View File

@@ -18,6 +18,7 @@ from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.layers.quantization.int8_kernel import per_token_group_quant_int8 from sglang.srt.layers.quantization.int8_kernel import per_token_group_quant_int8
from sglang.srt.utils import ( from sglang.srt.utils import (
direct_register_custom_op, direct_register_custom_op,
get_bool_env_var,
get_device_name, get_device_name,
is_cuda_available, is_cuda_available,
is_hip, is_hip,
@@ -941,7 +942,11 @@ def fused_experts_impl(
no_combine: bool = False, no_combine: bool = False,
): ):
padded_size = padding_size padded_size = padding_size
if not use_fp8_w8a8 or not use_int8_w8a8 or block_shape is not None: if (
not (use_fp8_w8a8 or use_int8_w8a8)
or block_shape is not None
or (is_hip_ and get_bool_env_var("CK_MOE"))
):
padded_size = 0 padded_size = 0
# Check constraints. # Check constraints.