fix nemotron capture for non attention layers (#21436)
This commit is contained in:
@@ -2424,6 +2424,8 @@ class ModelRunner(ModelRunnerKVCacheMixin):
|
|||||||
|
|
||||||
if attn_layer is not None:
|
if attn_layer is not None:
|
||||||
self.attention_layers.append(attn_layer)
|
self.attention_layers.append(attn_layer)
|
||||||
|
elif hasattr(layer, "mixer"):
|
||||||
|
self.attention_layers.append(None)
|
||||||
|
|
||||||
moe_block = None
|
moe_block = None
|
||||||
moe_fusion = None
|
moe_fusion = None
|
||||||
|
|||||||
Reference in New Issue
Block a user