[Bugfix]: Fix Quark gpt-oss weight loading broken by FusedMoe refactor (#45067)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
co-authored by
Andreas Karatzas
parent
86111c00c7
commit
7920ccb97c
@@ -635,6 +635,13 @@ class GptOssModel(nn.Module, EagleModelMixin):
|
||||
"an unexpected condition. Please open an issue if encountered."
|
||||
)
|
||||
|
||||
# The MoE refactor (#41184) moved expert params under
|
||||
# `mlp.experts.routed_experts.*`; remap the legacy checkpoint
|
||||
# name so keys like w2_bias resolve against params_dict.
|
||||
fused_name = fused_name.replace(
|
||||
".mlp.experts.", ".mlp.experts.routed_experts."
|
||||
)
|
||||
|
||||
moe_quant_method = _get_moe_weight_dtype(layer_id=layer_id)
|
||||
|
||||
if (
|
||||
|
||||
Reference in New Issue
Block a user