diff --git a/vllm/lora/model_manager.py b/vllm/lora/model_manager.py index 92c99b3fa4..8063f485bf 100644 --- a/vllm/lora/model_manager.py +++ b/vllm/lora/model_manager.py @@ -1052,7 +1052,7 @@ class LoRAModelManager: """Narrow a flat expert-major sub-module list to this rank's experts. ``new_module_names`` is produced by - ``FusedMoE.make_expert_params_mapping`` and is ordered + ``fused_moe_make_expert_params_mapping`` and is ordered ``[e=0,w1, e=0,w2, e=0,w3, e=1,w1, ...]`` (non-gated MoE has 2 entries per expert instead of 3). When the module is a 2D ``FusedMoEWithLoRA`` with EP enabled, we slice the list to the diff --git a/vllm/model_executor/models/cohere2_moe.py b/vllm/model_executor/models/cohere2_moe.py index 16e299ea27..7a5253b195 100644 --- a/vllm/model_executor/models/cohere2_moe.py +++ b/vllm/model_executor/models/cohere2_moe.py @@ -16,7 +16,10 @@ from vllm.distributed import ( ) from vllm.model_executor.layers.activation import SiluAndMul from vllm.model_executor.layers.attention import Attention -from vllm.model_executor.layers.fused_moe import FusedMoE +from vllm.model_executor.layers.fused_moe import ( + FusedMoE, + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.linear import ( MergedColumnParallelLinear, QKVParallelLinear, @@ -450,7 +453,7 @@ class Cohere2MoeModel(nn.Module): ("gate_up_proj", "up_proj", 1), ] - expert_params_mapping = FusedMoE.make_expert_params_mapping( + expert_params_mapping = fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/hy_v3.py b/vllm/model_executor/models/hy_v3.py index 6142f2c086..7653cddd6c 100644 --- a/vllm/model_executor/models/hy_v3.py +++ b/vllm/model_executor/models/hy_v3.py @@ -43,7 +43,11 @@ from vllm.distributed import ( from vllm.logger import init_logger from vllm.model_executor.layers.activation import SiluAndMul from vllm.model_executor.layers.attention import Attention -from vllm.model_executor.layers.fused_moe import FusedMoE, GateLinear +from vllm.model_executor.layers.fused_moe import ( + FusedMoE, + GateLinear, + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import ( MergedColumnParallelLinear, @@ -474,7 +478,7 @@ class HYV3Model(nn.Module): def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) - return FusedMoE.make_expert_params_mapping( + return fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/hy_v3_mtp.py b/vllm/model_executor/models/hy_v3_mtp.py index 37ace1dada..77f323aae4 100644 --- a/vllm/model_executor/models/hy_v3_mtp.py +++ b/vllm/model_executor/models/hy_v3_mtp.py @@ -32,7 +32,9 @@ from torch import nn from transformers import PretrainedConfig from vllm.config import CacheConfig, ModelConfig, VllmConfig -from vllm.model_executor.layers.fused_moe import FusedMoE +from vllm.model_executor.layers.fused_moe import ( + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig @@ -294,7 +296,7 @@ class HYV3MTP(nn.Module): ] if _is_moe(self.config): - expert_params_mapping = FusedMoE.make_expert_params_mapping( + expert_params_mapping = fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/models/deepseek_v4/xpu/model.py b/vllm/models/deepseek_v4/xpu/model.py index 356d9f5024..9f20d0ed3e 100644 --- a/vllm/models/deepseek_v4/xpu/model.py +++ b/vllm/models/deepseek_v4/xpu/model.py @@ -18,7 +18,11 @@ from vllm.distributed import ( ) from vllm.forward_context import get_forward_context from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp -from vllm.model_executor.layers.fused_moe import FusedMoE, GateLinear +from vllm.model_executor.layers.fused_moe import ( + FusedMoE, + GateLinear, + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.fused_moe.router.fused_topk_bias_router import ( fused_topk_bias, ) @@ -1220,7 +1224,7 @@ class DeepseekV4Model(nn.Module): return make_deepseek_v4_expert_params_mapping(self.config.n_routed_experts) # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) - return FusedMoE.make_expert_params_mapping( + return fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="w1", ckpt_down_proj_name="w2", diff --git a/vllm/models/deepseek_v4/xpu/mtp.py b/vllm/models/deepseek_v4/xpu/mtp.py index 9539a2d999..8dbe40bb6a 100644 --- a/vllm/models/deepseek_v4/xpu/mtp.py +++ b/vllm/models/deepseek_v4/xpu/mtp.py @@ -25,7 +25,9 @@ from vllm.distributed import ( get_tensor_model_parallel_world_size, ) from vllm.logger import init_logger -from vllm.model_executor.layers.fused_moe import FusedMoE +from vllm.model_executor.layers.fused_moe import ( + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import ReplicatedLinear from vllm.model_executor.layers.logits_processor import LogitsProcessor @@ -332,7 +334,7 @@ class DeepSeekV4MTP(nn.Module): self.config.n_routed_experts ) else: - expert_mapping = FusedMoE.make_expert_params_mapping( + expert_mapping = fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="w1", ckpt_down_proj_name="w2",