fix e score bias

Signed-off-by: jenchen13 <[email protected]>

Author: jenchen13

modelopt/torch/export/plugins/mcore_nemotron.py

@@ -66,7 +66,10 @@
     "linear_fc1": NameRemapping("backbone.layers.{}.mixer.up_proj.", COL_TP),
     "linear_fc2": NameRemapping("backbone.layers.{}.mixer.down_proj.", ROW_TP),
     # MoE
-    "router": NameRemapping("model.layers.{}.mlp.gate.", REPLICATE),
+    "router": NameRemapping(
+        "backbone.layers.{}.mlp.gate.", {"mapping": {"expert_bias": "e_score_correction_bias"}}
+    ),
+
     "local_experts.linear_fc1": NameRemapping(
         "backbone.layers.{}.mixer.experts.{}.up_proj", COL_ETP
     ),
@@ -104,7 +107,9 @@
     "linear_fc1": NameRemapping("backbone.layers.{}.mixer.up_proj."),
     "linear_fc2": NameRemapping("backbone.layers.{}.mixer.down_proj."),
     # MoE
-    "router": NameRemapping("backbone.layers.{}.mlp.gate."),
+    "router": NameRemapping(
+        "backbone.layers.{}.mlp.gate.", {"mapping": {"expert_bias": "e_score_correction_bias"}}
+    ),
     "local_experts.linear_fc1": NameRemapping("backbone.layers.{}.mixer.experts.{}.up_proj."),
     "local_experts.linear_fc2": NameRemapping("backbone.layers.{}.mixer.experts.{}.down_proj."),
     "shared_experts.linear_fc1": NameRemapping("backbone.layers.{}.mixer.shared_experts.up_proj."),