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

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

vllm/model_executor/models/afmoe.py

@@ -475,6 +475,7 @@ class AfmoeModel(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/bailing_moe.py

@@ -476,6 +476,7 @@ class BailingMoeModel(nn.Module):
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
         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/deepseek_eagle.py

@@ -106,6 +106,7 @@ class DeepseekV2Model(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/deepseek_mtp.py

@@ -245,6 +245,7 @@ class DeepSeekMTP(nn.Module, SupportsPP, DeepseekV2MixtureOfExperts):
         ]
 
         expert_params_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/deepseek_v2.py

@@ -1486,6 +1486,7 @@ class DeepseekV2ForCausalLM(
         # 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",
@@ -1519,6 +1520,7 @@ class DeepseekV2ForCausalLM(
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         expert_params_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/dots1.py

@@ -424,6 +424,7 @@ class Dots1Model(nn.Module):
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
         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/ernie45_moe.py

@@ -497,6 +497,7 @@ class Ernie4_5_MoeModel(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/ernie45_vl_moe.py

@@ -675,6 +675,7 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP):
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
         expert_params_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/glm4_moe.py

@@ -496,6 +496,7 @@ class Glm4MoeModel(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/glm4_moe_mtp.py

@@ -248,6 +248,7 @@ class Glm4MoeMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
         # 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/gpt_oss.py

@@ -729,6 +729,7 @@ class GptOssForCausalLM(nn.Module, SupportsPP, SupportsEagle3, SupportsLoRA):
         # Params for weights, weight scales, 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/granitemoe.py

@@ -353,6 +353,7 @@ class GraniteMoeModel(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="w1",
             ckpt_down_proj_name="w2",
             ckpt_up_proj_name="w3",

vllm/model_executor/models/grok1.py

@@ -369,6 +369,7 @@ class Grok1Model(nn.Module):
         # Grok1 uses "num_experts" in its config
         num_experts = getattr(self.config, "num_experts", 8)
         return FusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="linear",  # Grok1 specific
             ckpt_down_proj_name="linear_1",  # Grok1 specific
             ckpt_up_proj_name="linear_v",  # Grok1 specific

vllm/model_executor/models/hunyuan_v1.py

@@ -706,6 +706,7 @@ class HunYuanModel(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/jamba.py

@@ -378,6 +378,7 @@ class JambaModel(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/kimi_linear.py

@@ -560,6 +560,7 @@ class KimiLinearForCausalLM(
             # 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="w1",
                 ckpt_down_proj_name="w2",
                 ckpt_up_proj_name="w3",

vllm/model_executor/models/kimi_vl.py

@@ -462,6 +462,7 @@ class KimiVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
             # 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/lfm2_moe.py

@@ -486,6 +486,7 @@ class Lfm2MoeModel(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/llama4.py

@@ -539,6 +539,7 @@ class Llama4Model(LlamaModel):
         # Expert parameter mapping for the case where the expert weights are
         # not fused into a single weight tensor.
         expert_params_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",
@@ -548,6 +549,7 @@ class Llama4Model(LlamaModel):
         # Expert parameter mapping for the case where the expert weights are
         # fused into a single weight tensor.
         expert_params_mapping_fused = SharedFusedMoE.make_expert_params_mapping(
+            self,
             ckpt_gate_proj_name="gate_up_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="gate_up_proj",

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,