Optimize Qwen3-moe model by using flashinfer fused allreduce (#9973)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>

python/sglang/srt/layers/communicator.py

@@ -42,9 +42,15 @@ from sglang.srt.layers.moe import (
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.utils import is_cuda, is_flashinfer_available, is_sm100_supported
+from sglang.srt.utils import (
+    is_cuda,
+    is_flashinfer_available,
+    is_sm90_supported,
+    is_sm100_supported,
+)
 
 _is_flashinfer_available = is_flashinfer_available()
+_is_sm90_supported = is_cuda() and is_sm90_supported()
 _is_sm100_supported = is_cuda() and is_sm100_supported()
 
 FUSE_ALLREDUCE_MAX_BATCH_SIZE = 2048
@@ -484,11 +490,11 @@ class CommunicateWithAllReduceAndLayerNormFn:
             # According to the discussion in https://github.com/flashinfer-ai/flashinfer/issues/1223#issuecomment-3047256465
             # We set the max token num to 128 for allreduce fusion with min-latency case(use_oneshot=True).
             if (
-                _is_sm100_supported
+                (_is_sm100_supported or _is_sm90_supported)
                 and _is_flashinfer_available
                 and hasattr(layernorm, "forward_with_allreduce_fusion")
                 and global_server_args_dict["enable_flashinfer_allreduce_fusion"]
-                and hidden_states.shape[0] <= 2048
+                and hidden_states.shape[0] <= 4096
             ):
                 hidden_states, residual = layernorm.forward_with_allreduce_fusion(
                     hidden_states, residual

python/sglang/srt/models/qwen2_moe.py

@@ -105,11 +105,14 @@ class Qwen2MoeMLP(nn.Module):
     def forward(
         self,
         x,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ):
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
-        x, _ = self.down_proj(x, skip_all_reduce=use_reduce_scatter)
+        x, _ = self.down_proj(
+            x, skip_all_reduce=should_allreduce_fusion or use_reduce_scatter
+        )
         return x
 
 

python/sglang/srt/models/qwen3_moe.py

@@ -42,7 +42,10 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe import get_moe_a2a_backend
+from sglang.srt.layers.moe import (
+    get_moe_a2a_backend,
+    should_use_flashinfer_cutlass_moe_fp4_allgather,
+)
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
@@ -57,10 +60,17 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTe
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2_moe import Qwen2MoeMLP as Qwen3MoeMLP
 from sglang.srt.models.qwen2_moe import Qwen2MoeModel
-from sglang.srt.utils import add_prefix, is_cuda, is_non_idle_and_non_empty
+from sglang.srt.utils import (
+    add_prefix,
+    is_cuda,
+    is_flashinfer_available,
+    is_non_idle_and_non_empty,
+)
 
 Qwen3MoeConfig = None
 
+_is_flashinfer_available = is_flashinfer_available()
+
 logger = logging.getLogger(__name__)
 _is_cuda = is_cuda()
 
@@ -119,11 +129,14 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         forward_batch: Optional[ForwardBatch] = None,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
 
         if not get_moe_a2a_backend().is_deepep():
-            return self.forward_normal(hidden_states, use_reduce_scatter)
+            return self.forward_normal(
+                hidden_states, should_allreduce_fusion, use_reduce_scatter
+            )
         else:
             return self.forward_deepep(hidden_states, forward_batch)
 
@@ -137,6 +150,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
     def forward_normal(
         self,
         hidden_states: torch.Tensor,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
@@ -146,7 +160,12 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
         router_logits, _ = self.gate(hidden_states)
         topk_output = self.topk(hidden_states, router_logits)
         final_hidden_states = self.experts(hidden_states, topk_output)
-        if self.tp_size > 1 and not use_reduce_scatter:
+        if (
+            self.tp_size > 1
+            and not should_allreduce_fusion
+            and not use_reduce_scatter
+            and not should_use_flashinfer_cutlass_moe_fp4_allgather()
+        ):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(num_tokens, hidden_dim)
@@ -500,6 +519,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
             input_layernorm=self.input_layernorm,
             post_attention_layernorm=self.post_attention_layernorm,
             allow_reduce_scatter=True,
+            is_last_layer=(self.layer_id == self.config.num_hidden_layers - 1),
         )
 
     def forward(
@@ -525,13 +545,24 @@ class Qwen3MoeDecoderLayer(nn.Module):
             hidden_states, residual, forward_batch
         )
 
+        should_allreduce_fusion = (
+            self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer(
+                forward_batch
+            )
+        )
+
         # For DP with padding, reduce scatter can be used instead of all-reduce.
         use_reduce_scatter = self.layer_communicator.should_use_reduce_scatter(
             forward_batch
         )
 
-        hidden_states = self.mlp(hidden_states, forward_batch, use_reduce_scatter)
+        hidden_states = self.mlp(
+            hidden_states, forward_batch, should_allreduce_fusion, use_reduce_scatter
+        )
 
+        if should_allreduce_fusion:
+            hidden_states._sglang_needs_allreduce_fusion = True
+        else:
             hidden_states, residual = self.layer_communicator.postprocess_layer(
                 hidden_states, residual, forward_batch
             )