[Bugfix] LoRA: extend expert base_layer loading to Qwen3.5 and Step3.x (#37114)

Signed-off-by: Hollow Man <hollowman@opensuse.org>
This commit is contained in:
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟
2026-04-21 14:17:03 +00:00
committed by GitHub
parent ec5ef0ac73
commit 908a713488
6 changed files with 34 additions and 16 deletions
+6 -3
View File
@@ -463,11 +463,14 @@ class Step3TextForCausalLM(nn.Module, SupportsPP):
]
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
expert_params_mapping = [
(".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
(".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
(".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
(f".moe.experts.{base_layer}w13_weight", ".moe.gate_proj.weight", "w1"),
(f".moe.experts.{base_layer}w13_weight", ".moe.up_proj.weight", "w3"),
(f".moe.experts.{base_layer}w2_weight", ".moe.down_proj.weight", "w2"),
]
disable_moe_stacked_params = [data[1] for data in expert_params_mapping]