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

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

vllm/model_executor/layers/fused_moe/layer.py

@@ -2007,6 +2007,7 @@ class FusedMoE(CustomOp):
     @classmethod
     def make_expert_params_mapping(
         cls,
+        model: torch.nn.Module,
         ckpt_gate_proj_name: str,
         ckpt_down_proj_name: str,
         ckpt_up_proj_name: str,
@@ -2025,13 +2026,19 @@ class FusedMoE(CustomOp):
             )
         )
 
+        base_layer = (
+            "base_layer."
+            if any(".base_layer." in name for name, _ in model.named_parameters())
+            else ""
+        )
+
         return [
             # (param_name, weight_name, expert_id, shard_id)
             (
-                "experts.w13_"
+                f"experts.{base_layer}w13_"
                 if weight_name in [ckpt_gate_proj_name, ckpt_up_proj_name]
-                else "experts.w2_",
-                f"experts.{physical_to_logical_map[expert_id]}.{weight_name}.",
+                else f"experts.{base_layer}w2_",
+                f"experts.{physical_to_logical_map[expert_id]}.{weight_name}.{base_layer}",
                 expert_id,
                 shard_id,
             )

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",