[BugFix] LoRA: Support loading base_layer of experts (#31104)

Signed-off-by: Hollow Man <hollowman@opensuse.org>

vllm/model_executor/models/longcat_flash.py

@@ -626,6 +626,7 @@ class LongcatFlashForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/mimo_v2_flash.py

@@ -512,6 +512,7 @@ class MiMoV2Model(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/minimax_m2.py

@@ -392,6 +392,7 @@ class MiniMaxM2Model(nn.Module):
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",
             ckpt_up_proj_name="w3",

vllm/model_executor/models/mixtral.py

@@ -366,6 +366,7 @@ class MixtralModel(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",
             ckpt_up_proj_name="w3",

vllm/model_executor/models/mllama4.py

@@ -1084,6 +1084,7 @@ class Llama4ForConditionalGeneration(
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/nemotron_h.py

@@ -636,6 +636,7 @@ class NemotronHModel(nn.Module):
                 #   what the activation is applied to
                 # - FusedMoe.w3 (aka up_proj) should be ignored since we're
                 #   using non-gated MoE
+                self,
                 ckpt_gate_proj_name="up_proj",
                 ckpt_down_proj_name="down_proj",
                 ckpt_up_proj_name="",

vllm/model_executor/models/olmoe.py

@@ -338,6 +338,7 @@ class OlmoeModel(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/openpangu.py

@@ -1161,6 +1161,7 @@ class OpenPanguModel(nn.Module):
         has_experts = hasattr(self.config, "n_routed_experts")
         if has_experts:
             expert_merge_mapping = SharedFusedMoE.make_expert_params_mapping(
+                self,
                 ckpt_gate_proj_name="gate_proj",
                 ckpt_down_proj_name="down_proj",
                 ckpt_up_proj_name="up_proj",

vllm/model_executor/models/openpangu_mtp.py

@@ -149,6 +149,7 @@ class OpenPanguMTP(nn.Module, SupportsPP):
         ]
 
         expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/phimoe.py

@@ -516,6 +516,7 @@ class PhiMoEModel(nn.Module):
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",
             ckpt_up_proj_name="w3",

vllm/model_executor/models/qwen2_moe.py

@@ -423,6 +423,7 @@ class Qwen2MoeModel(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return SharedFusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/qwen3_moe.py

@@ -470,6 +470,7 @@ class Qwen3MoeModel(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/qwen3_next.py

@@ -1031,6 +1031,7 @@ class Qwen3NextModel(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         return SharedFusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/qwen3_next_mtp.py

@@ -147,6 +147,7 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

vllm/model_executor/models/transformers/moe.py

@@ -165,6 +165,7 @@ class MoEMixin(MixtureOfExperts):
         for gate_proj, down_proj, up_proj in ckpt_names:
             expert_mapping.extend(
                 FusedMoE.make_expert_params_mapping(
+                    self,
                     ckpt_gate_proj_name=gate_proj,
                     ckpt_down_proj_name=down_proj,
                     ckpt_up_proj_name=up_proj,