[qwen3] support qwen3 ep moe (sgl-project#5917)

Co-authored-by: sleepcoo <[email protected]>

Author: 2 people

python/sglang/srt/models/qwen2_moe.py

@@ -36,6 +36,7 @@
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.ep_moe.layer import EPMoE
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
@@ -45,6 +46,7 @@
     VocabParallelEmbedding,
 )
 from sglang.srt.managers.expert_distribution import ExpertDistributionRecorder
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, make_layers
@@ -108,12 +110,13 @@ def __init__(
                 f"the number of experts {config.num_experts}."
             )
 
-        self.experts = FusedMoE(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        self.experts = MoEImpl(
             num_experts=config.num_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
@@ -427,7 +430,9 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        expert_params_mapping = MoEImpl.make_expert_params_mapping(
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",

python/sglang/srt/models/qwen3_moe.py

@@ -40,6 +40,7 @@
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.ep_moe.layer import EPMoE
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
@@ -48,6 +49,7 @@
     ParallelLMHead,
     VocabParallelEmbedding,
 )
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2_moe import Qwen2MoeMLP as Qwen3MoeMLP
@@ -73,12 +75,13 @@ def __init__(
                 f"the number of experts {config.num_experts}."
             )
 
-        self.experts = FusedMoE(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        self.experts = MoEImpl(
             num_experts=config.num_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
@@ -356,7 +359,9 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
+
+        expert_params_mapping = MoEImpl.make_expert_params_mapping(
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",