[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
+5 -2
View File
@@ -306,9 +306,12 @@ class Qwen3_5Model(Qwen3NextModel):
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
is_fused_expert = False
base_layer = (
"base_layer." if any(".base_layer." in name for name in params_dict) else ""
)
fused_expert_params_mapping = [
("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
("experts.w2_weight", "experts.down_proj", 0, "w2"),
(f"experts.{base_layer}w13_weight", "experts.gate_up_proj", 0, "w1"),
(f"experts.{base_layer}w2_weight", "experts.down_proj", 0, "w2"),
]
num_experts = (
self.config.num_experts if hasattr(self.config, "num_experts") else 0