Fix down projection weight shape in fused MOE layer (#4041)

2025-10-19 23:16:15 +08:00 · 2025-09-10 12:49:03 +08:00
parent dfc94371ee
commit 7272afe3dc
1 changed files with 1 additions and 1 deletions
--- a/fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
+++ b/fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
@@ -695,7 +695,7 @@ class BlockWiseFP8MoEMethod(QuantMethodBase):
                layer,
                down_proj_weight_name,
                layer.create_parameter(
-                    shape=self.up_gate_proj_weight_shape,
+                    shape=self.down_proj_weight_shape,
                    dtype=self.weight_dtype,
                    default_initializer=paddle.nn.initializer.Constant(0),
                ),