Skip to content

Commit 752c70a

Browse files
committed
fix naming to backbone
Signed-off-by: jenchen13 <[email protected]>
1 parent 110b78e commit 752c70a

File tree

1 file changed

+9
-11
lines changed

1 file changed

+9
-11
lines changed

modelopt/torch/export/plugins/mcore_nemotron.py

Lines changed: 9 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -26,8 +26,6 @@
2626
NameRemapping,
2727
QKVMerging,
2828
QKVSlicing,
29-
GatedMLPSlicing,
30-
GatedMLPMerging,
3129
)
3230

3331
# Example on adding a new CausalLM.
@@ -69,11 +67,11 @@
6967
"linear_fc2": NameRemapping("backbone.layers.{}.mixer.down_proj.", ROW_TP),
7068
# MoE
7169
"router": NameRemapping("model.layers.{}.mlp.gate.", REPLICATE),
72-
"local_experts.linear_fc1": NameRemapping("model.layers.{}.mlp.experts.{}.up_proj", COL_ETP),
73-
"local_experts.linear_fc2": NameRemapping("model.layers.{}.mlp.experts.{}.down_proj.", ROW_ETP),
74-
"shared_experts.linear_fc1": NameRemapping("model.layers.{}.mlp.shared_experts.up_proj.", COL_TP),
70+
"local_experts.linear_fc1": NameRemapping("backbone.layers.{}.mlp.experts.{}.up_proj", COL_ETP),
71+
"local_experts.linear_fc2": NameRemapping("backbone.layers.{}.mlp.experts.{}.down_proj.", ROW_ETP),
72+
"shared_experts.linear_fc1": NameRemapping("backbone.layers.{}.mlp.shared_experts.up_proj.", COL_TP),
7573
"shared_experts.linear_fc2": NameRemapping(
76-
"model.layers.{}.mlp.shared_experts.down_proj.", ROW_TP
74+
"backbone.layers.{}.mlp.shared_experts.down_proj.", ROW_TP
7775
),
7876

7977
}
@@ -101,10 +99,10 @@
10199
"linear_fc1": NameRemapping("backbone.layers.{}.mixer.up_proj."),
102100
"linear_fc2": NameRemapping("backbone.layers.{}.mixer.down_proj."),
103101
# MoE
104-
"router": NameRemapping("model.layers.{}.mlp.gate."),
105-
"local_experts.linear_fc1": NameRemapping("model.layers.{}.mlp.experts.{}.up_proj."),
106-
"local_experts.linear_fc2": NameRemapping("model.layers.{}.mlp.experts.{}.down_proj."),
107-
"shared_experts.linear_fc1": NameRemapping("model.layers.{}.mlp.shared_experts.up_proj."),
108-
"shared_experts.linear_fc2": NameRemapping("model.layers.{}.mlp.shared_experts.down_proj."),
102+
"router": NameRemapping("backbone.layers.{}.mlp.gate."),
103+
"local_experts.linear_fc1": NameRemapping("backbone.layers.{}.mlp.experts.{}.up_proj."),
104+
"local_experts.linear_fc2": NameRemapping("backbone.layers.{}.mlp.experts.{}.down_proj."),
105+
"shared_experts.linear_fc1": NameRemapping("backbone.layers.{}.mlp.shared_experts.up_proj."),
106+
"shared_experts.linear_fc2": NameRemapping("backbone.layers.{}.mlp.shared_experts.down_proj."),
109107

110108
}

0 commit comments

Comments
 (0)