[BugFix] Fix the abnormal memory usage caused by shape errors in the triton moe backend (#4026)

* fix device_id to in * fix triton_moe bug
2025-10-05 00:33:03 +08:00 · 2025-09-10 11:05:54 +08:00
parent dbab579299
commit c3b2a60fb8
4 changed files with 12 additions and 10 deletions
--- a/fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
+++ b/fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
@@ -671,7 +671,7 @@ 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),
                ),