apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-11-02 20:54:03 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

fix_fp8_deepgemm_moe_tp_bug (#3658)

Browse Source
This commit is contained in:
Liumengyuan
2025-08-28 17:19:02 +08:00
committed by GitHub
parent e93d4cfcdd
commit 2a73a6df03
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/model_executor/layers/moe/fused_moe_deepgemm_backend.py
View File

@@ -49,13 +49,13 @@ class DeepGemmFusedMoeMethod(MoEMethodBase):
]
self.up_gate_proj_scale_shape = [
layer.num_local_experts,
layer.moe_intermediate_size * 2 // self.quant_config.weight_block_size[0],
layer.hidden_size // self.quant_config.weight_block_size[1],
ceil_div(layer.moe_intermediate_size * 2, self.quant_config.weight_block_size[0]),
ceil_div(layer.hidden_size, self.quant_config.weight_block_size[1]),
]
self.down_proj_scale_shape = [
layer.num_local_experts,
layer.hidden_size // self.quant_config.weight_block_size[0],
layer.moe_intermediate_size // self.quant_config.weight_block_size[1],
ceil_div(layer.hidden_size, self.quant_config.weight_block_size[0]),
ceil_div(layer.moe_intermediate_size, self.quant_config.weight_block_size[1]),
]
if self.quant_config.is_checkpoint_bf16:
layer.up_gate_proj_weight = layer.create_parameter(