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

avoid fused_moe_triton padding circular import (#2624)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2024-12-28 14:01:35 +08:00
committed by GitHub
parent 8a2681e26a
commit 9254a33ad4
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/layers/quantization/fp8.py
View File

@@ -28,7 +28,6 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
from vllm.model_executor.parameter import ModelWeightParameter, PerTensorScaleParameter from vllm.model_executor.parameter import ModelWeightParameter, PerTensorScaleParameter
from sglang.srt.layers.linear import LinearMethodBase, UnquantizedLinearMethod from sglang.srt.layers.linear import LinearMethodBase, UnquantizedLinearMethod
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import padding_size
from sglang.srt.layers.quantization.base_config import ( from sglang.srt.layers.quantization.base_config import (
QuantizationConfig, QuantizationConfig,
QuantizeMethodBase, QuantizeMethodBase,
@@ -548,6 +547,10 @@ class Fp8MoEMethod:
layer.w2_input_scale = None layer.w2_input_scale = None
def process_weights_after_loading(self, layer: Module) -> None: def process_weights_after_loading(self, layer: Module) -> None:
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
padding_size, # Avoid circular import
)
# Block quant doesn't need to process weights after loading # Block quant doesn't need to process weights after loading
if self.block_quant: if self.block_quant:
return return