fuse allreduce and residual_rmsnorm (#8731)

python/sglang/srt/layers/communicator.py

@@ -441,7 +441,7 @@ class CommunicateWithAllReduceAndLayerNormFn:
                 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] <= 128
+                and hidden_states.shape[0] <= 2048
             ):
                 hidden_states, residual = layernorm.forward_with_allreduce_fusion(
                     hidden_states, residual

python/sglang/srt/layers/flashinfer_comm_fusion.py

@@ -92,7 +92,7 @@ _workspace_manager = FlashInferWorkspaceManager()
 
 
 def ensure_workspace_initialized(
-    max_token_num: int = 128, hidden_dim: int = 4096, use_fp32_lamport: bool = False
+    max_token_num: int = 2048, hidden_dim: int = 4096, use_fp32_lamport: bool = False
 ):
     """Ensure workspace is initialized"""
     if not is_flashinfer_available() or _flashinfer_comm is None:
@@ -124,8 +124,8 @@ def flashinfer_allreduce_residual_rmsnorm(
     residual: torch.Tensor,
     weight: torch.Tensor,
     eps: float = 1e-6,
-    max_token_num: int = 128,
-    use_oneshot: bool = True,
+    max_token_num: int = 2048,
+    use_oneshot: Optional[bool] = None,
     trigger_completion_at_end: bool = False,
     fp32_acc: bool = False,
 ) -> Tuple[torch.Tensor, torch.Tensor]:

python/sglang/srt/layers/linear.py

@@ -1294,6 +1294,7 @@ class RowParallelLinear(LinearBase):
         with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
             output_parallel = self.quant_method.apply(self, input_parallel, bias=bias_)
             sm.tag(output_parallel)
+
         if self.reduce_results and self.tp_size > 1 and not skip_all_reduce:
             output = tensor_model_parallel_all_reduce(output_parallel)
         else:

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -847,10 +847,14 @@ class FusedMoE(torch.nn.Module):
             )
             sm.tag(final_hidden_states)
 
+        final_hidden_states = final_hidden_states[
+            ..., :origin_hidden_states_dim
+        ].contiguous()
+
         if self.reduce_results and (self.moe_tp_size > 1 or self.moe_ep_size > 1):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
-        return final_hidden_states[..., :origin_hidden_states_dim].contiguous()
+        return final_hidden_states
 
     @classmethod
     def make_expert_params_mapping(

python/sglang/srt/models/deepseek_v2.py

@@ -212,7 +212,7 @@ class DeepseekV2MLP(nn.Module):
         self,
         x,
         forward_batch=None,
-        can_fuse_mlp_allreduce: bool = False,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ):
         if (self.tp_size == 1) and x.shape[0] == 0:
@@ -221,7 +221,7 @@ class DeepseekV2MLP(nn.Module):
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
         x, _ = self.down_proj(
-            x, skip_all_reduce=can_fuse_mlp_allreduce or use_reduce_scatter
+            x, skip_all_reduce=should_allreduce_fusion or use_reduce_scatter
         )
         return x
 
@@ -448,7 +448,7 @@ class DeepseekV2MoE(nn.Module):
         self,
         hidden_states: torch.Tensor,
         forward_batch: Optional[ForwardBatch] = None,
-        can_fuse_mlp_allreduce: bool = False,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
         if not self._enable_deepep_moe:
@@ -459,11 +459,11 @@ class DeepseekV2MoE(nn.Module):
                 and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
             ):
                 return self.forward_normal_dual_stream(
-                    hidden_states, can_fuse_mlp_allreduce, use_reduce_scatter
+                    hidden_states, should_allreduce_fusion, use_reduce_scatter
                 )
             else:
                 return self.forward_normal(
-                    hidden_states, can_fuse_mlp_allreduce, use_reduce_scatter
+                    hidden_states, should_allreduce_fusion, use_reduce_scatter
                 )
         else:
             return self.forward_deepep(hidden_states, forward_batch)
@@ -471,7 +471,7 @@ class DeepseekV2MoE(nn.Module):
     def forward_normal_dual_stream(
         self,
         hidden_states: torch.Tensor,
-        can_fuse_mlp_allreduce: bool = False,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
 
@@ -500,20 +500,20 @@ class DeepseekV2MoE(nn.Module):
         torch.add(final_hidden_states, shared_output, out=final_hidden_states_out)
         final_hidden_states = final_hidden_states_out
         sm.tag(final_hidden_states)
-        if self.tp_size > 1 and not can_fuse_mlp_allreduce and not use_reduce_scatter:
+        if self.tp_size > 1 and not should_allreduce_fusion and not use_reduce_scatter:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
 
     def forward_normal(
         self,
         hidden_states: torch.Tensor,
-        can_fuse_mlp_allreduce: bool = False,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
         if hasattr(self, "shared_experts") and use_intel_amx_backend(
             self.shared_experts.gate_up_proj
         ):
-            return self.forward_cpu(hidden_states, can_fuse_mlp_allreduce)
+            return self.forward_cpu(hidden_states, should_allreduce_fusion)
 
         shared_output = self._forward_shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
@@ -537,12 +537,14 @@ class DeepseekV2MoE(nn.Module):
             torch.add(final_hidden_states, shared_output, out=final_hidden_states_out)
             final_hidden_states = final_hidden_states_out
             sm.tag(final_hidden_states)
-        if self.tp_size > 1 and not can_fuse_mlp_allreduce and not use_reduce_scatter:
+        if self.tp_size > 1 and not should_allreduce_fusion and not use_reduce_scatter:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
 
     def forward_cpu(
-        self, hidden_states: torch.Tensor, can_fuse_mlp_allreduce: bool = False
+        self,
+        hidden_states: torch.Tensor,
+        should_allreduce_fusion: bool = False,
     ) -> torch.Tensor:
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states)
@@ -593,7 +595,7 @@ class DeepseekV2MoE(nn.Module):
             None,  # a2_scale
             True,  # is_vnni
         )
-        if self.tp_size > 1 and not can_fuse_mlp_allreduce:
+        if self.tp_size > 1 and not should_allreduce_fusion:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
 
@@ -1842,6 +1844,8 @@ class DeepseekV2DecoderLayer(nn.Module):
             allow_reduce_scatter=True,
         )
 
+        self._fuse_allreduce_lookup_table = self._build_fuse_allreduce_lookup_table()
+
     def _is_layer_sparse(self, layer_id: int, is_nextn: bool) -> bool:
         return is_nextn or (
             self.config.n_routed_experts is not None
@@ -1850,27 +1854,18 @@ class DeepseekV2DecoderLayer(nn.Module):
         )
 
     def _should_fuse_mlp_allreduce_with_next_layer(self, forward_batch) -> bool:
-        """Check if MLP allreduce can be fused with next layer's add_rmsnorm"""
-
-        if (
-            self.layer_id == self.config.num_hidden_layers - 1
-            or get_tensor_model_parallel_world_size() <= 1
-        ):
-            return False
-
-        if not global_server_args_dict.get("enable_flashinfer_allreduce_fusion", False):
-            return False
+        """Check if MLP allreduce can be fused with next layer's residual_rmsnorm"""
 
-        if not _is_sm100_supported or not _is_flashinfer_available:
-            return False
+        batch_size = (
+            forward_batch.input_ids.shape[0]
+            if hasattr(forward_batch, "input_ids")
+            else 0
+        )
 
-        if hasattr(forward_batch, "input_ids") and (
-            forward_batch.input_ids.shape[0] == 0
-            or forward_batch.input_ids.shape[0] > 128
-        ):
+        if batch_size > 128:
             return False
 
-        return True
+        return self._fuse_allreduce_lookup_table.get(batch_size, False)
 
     def forward(
         self,
@@ -1896,7 +1891,7 @@ class DeepseekV2DecoderLayer(nn.Module):
             hidden_states, residual, forward_batch
         )
 
-        can_fuse_mlp_allreduce = (
+        should_allreduce_fusion = (
             self._should_fuse_mlp_allreduce_with_next_layer(forward_batch)
             and not (self.enable_dp_attention and self.speculative_algorithm.is_eagle())
             and not self.is_nextn
@@ -1907,13 +1902,13 @@ class DeepseekV2DecoderLayer(nn.Module):
             forward_batch
         )
         hidden_states = self.mlp(
-            hidden_states, forward_batch, can_fuse_mlp_allreduce, use_reduce_scatter
+            hidden_states, forward_batch, should_allreduce_fusion, use_reduce_scatter
         )
 
-        if can_fuse_mlp_allreduce:
+        if should_allreduce_fusion:
             hidden_states._sglang_needs_allreduce_fusion = True
 
-        if not can_fuse_mlp_allreduce:
+        if not should_allreduce_fusion:
             hidden_states, residual = self.layer_communicator.postprocess_layer(
                 hidden_states, residual, forward_batch
             )
@@ -1990,6 +1985,26 @@ class DeepseekV2DecoderLayer(nn.Module):
         )
         return output
 
+    def _build_fuse_allreduce_lookup_table(self):
+        static_conditions_met = (
+            self.layer_id != self.config.num_hidden_layers - 1
+            and get_tensor_model_parallel_world_size() > 1
+            and global_server_args_dict.get("enable_flashinfer_allreduce_fusion", False)
+            and _is_sm100_supported
+            and _is_flashinfer_available
+        )
+
+        if not static_conditions_met:
+            return {}
+
+        lookup_table = {}
+        for batch_size in range(129):  # 0 to 128
+            is_last_layer = self.layer_id == self.config.num_hidden_layers - 1
+            should_fuse = batch_size > 0 and batch_size <= 128 and not is_last_layer
+            lookup_table[batch_size] = should_fuse
+
+        return lookup_table
+
 
 class DeepseekV2Model(nn.Module):
     fall_back_to_pt_during_load = False

python/sglang/srt/models/glm4_moe.py

@@ -154,13 +154,13 @@ class Glm4MoeMLP(nn.Module):
             )
         self.act_fn = SiluAndMul()
 
-    def forward(self, x, forward_batch=None, can_fuse_mlp_allreduce=False):
+    def forward(self, x, forward_batch=None, should_allreduce_fusion=False):
         if (self.tp_size == 1) and x.shape[0] == 0:
             return x
 
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
-        x, _ = self.down_proj(x, skip_all_reduce=can_fuse_mlp_allreduce)
+        x, _ = self.down_proj(x, skip_all_reduce=should_allreduce_fusion)
         return x
 
 
@@ -529,7 +529,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
     def forward_normal_dual_stream(
         self,
         hidden_states: torch.Tensor,
-        can_fuse_mlp_allreduce: bool = False,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
 
@@ -553,7 +553,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
         if self.ep_size > 1:
             if (
                 self.tp_size > 1
-                and not can_fuse_mlp_allreduce
+                and not should_allreduce_fusion
                 and not use_reduce_scatter
             ):
                 final_hidden_states = tensor_model_parallel_all_reduce(
@@ -564,7 +564,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
             final_hidden_states += shared_output
             if (
                 self.tp_size > 1
-                and not can_fuse_mlp_allreduce
+                and not should_allreduce_fusion
                 and not use_reduce_scatter
             ):
                 final_hidden_states = tensor_model_parallel_all_reduce(
@@ -575,13 +575,13 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
     def forward_normal(
         self,
         hidden_states: torch.Tensor,
-        can_fuse_mlp_allreduce: bool = False,
+        should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
         if hasattr(self, "shared_experts") and use_intel_amx_backend(
             self.shared_experts.gate_up_proj
         ):
-            return self.forward_cpu(hidden_states, can_fuse_mlp_allreduce)
+            return self.forward_cpu(hidden_states, should_allreduce_fusion)
 
         shared_output = self._forward_shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
@@ -596,7 +596,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
             # fused in biased_grouped_topk so we can skip here
             final_hidden_states *= self.routed_scaling_factor
         if self.ep_size > 1:
-            if self.tp_size > 1 and not can_fuse_mlp_allreduce:
+            if self.tp_size > 1 and not should_allreduce_fusion:
                 final_hidden_states = tensor_model_parallel_all_reduce(
                     final_hidden_states
                 )
@@ -605,7 +605,7 @@ class Glm4MoeSparseMoeBlock(DeepseekV2MoE):
         else:
             if shared_output is not None:
                 final_hidden_states += shared_output
-            if self.tp_size > 1 and not can_fuse_mlp_allreduce:
+            if self.tp_size > 1 and not should_allreduce_fusion:
                 final_hidden_states = tensor_model_parallel_all_reduce(
                     final_hidden_states
                 )

python/sglang/srt/models/gpt_oss.py

@@ -56,7 +56,7 @@ from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8_utils import dequant_mxfp4
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
-from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
+from sglang.srt.layers.utils import PPMissingLayer, get_layer_id, is_sm100_supported
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
@@ -64,7 +64,10 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils import add_prefix, is_cuda, is_flashinfer_available, make_layers
+
+_is_flashinfer_available = is_flashinfer_available()
+_is_sm100_supported = is_cuda() and is_sm100_supported()
 
 
 class GptOssConfig(PretrainedConfig):
@@ -151,10 +154,13 @@ class GptOssSparseMoeBlock(nn.Module):
         )
 
     def forward(
-        self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: Optional[ForwardBatch] = None,
+        should_allreduce_fusion: bool = False,
     ) -> torch.Tensor:
         if not global_server_args_dict["moe_a2a_backend"].is_deepep():
-            return self.forward_normal(hidden_states)
+            return self.forward_normal(hidden_states, should_allreduce_fusion)
         else:
             raise Exception("forward_deepep branch not implemented yet")
 
@@ -165,7 +171,11 @@ class GptOssSparseMoeBlock(nn.Module):
             if name not in ["correction_bias"]
         ]
 
-    def forward_normal(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    def forward_normal(
+        self,
+        hidden_states: torch.Tensor,
+        should_allreduce_fusion: bool = False,
+    ) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
 
@@ -179,7 +189,7 @@ class GptOssSparseMoeBlock(nn.Module):
             kwargs["topk_output"] = (self.top_k, router_logits)
         final_hidden_states = self.experts(**kwargs)
 
-        if self.tp_size > 1:
+        if self.tp_size > 1 and not should_allreduce_fusion:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         ans = final_hidden_states.view(num_tokens, hidden_dim)
@@ -370,6 +380,7 @@ class GptOssDecoderLayer(nn.Module):
 
         # GptOss all layers are sparse and have no nextn now
         self.is_layer_sparse = True
+        self.is_nextn = False
         is_previous_layer_sparse = True
 
         self.layer_scatter_modes = LayerScatterModes.init_new(
@@ -402,6 +413,42 @@ class GptOssDecoderLayer(nn.Module):
             post_attention_layernorm=self.post_attention_layernorm,
         )
 
+        self._fuse_allreduce_lookup_table = self._build_fuse_allreduce_lookup_table()
+
+    def _should_fuse_mlp_allreduce_with_next_layer(self, forward_batch) -> bool:
+        """Check if MLP allreduce can be fused with next layer's residual_rmsnorm"""
+
+        batch_size = (
+            forward_batch.input_ids.shape[0]
+            if hasattr(forward_batch, "input_ids")
+            else 0
+        )
+
+        if batch_size > 128:
+            return False
+
+        return self._fuse_allreduce_lookup_table.get(batch_size, False)
+
+    def _build_fuse_allreduce_lookup_table(self):
+        static_conditions_met = (
+            self.layer_id != self.config.num_hidden_layers - 1
+            and get_tensor_model_parallel_world_size() > 1
+            and global_server_args_dict.get("enable_flashinfer_allreduce_fusion", False)
+            and _is_sm100_supported
+            and _is_flashinfer_available
+        )
+
+        if not static_conditions_met:
+            return {}
+
+        lookup_table = {}
+        for batch_size in range(129):  # 0 to 128
+            is_last_layer = self.layer_id == self.config.num_hidden_layers - 1
+            should_fuse = batch_size > 0 and batch_size <= 128 and not is_last_layer
+            lookup_table[batch_size] = should_fuse
+
+        return lookup_table
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -424,8 +471,17 @@ class GptOssDecoderLayer(nn.Module):
             hidden_states, residual, forward_batch
         )
 
-        hidden_states = self.mlp(hidden_states, forward_batch)
+        should_allreduce_fusion = (
+            self._should_fuse_mlp_allreduce_with_next_layer(forward_batch)
+            and not self.is_nextn
+        )
+
+        hidden_states = self.mlp(hidden_states, forward_batch, should_allreduce_fusion)
+
+        if should_allreduce_fusion:
+            hidden_states._sglang_needs_allreduce_fusion = True
 
+        if not should_allreduce_fusion:
             hidden_states, residual = self.layer_communicator.postprocess_layer(
                 hidden_states, residual, forward_batch
             )

python/sglang/srt/server_args.py

@@ -1435,7 +1435,7 @@ class ServerArgs:
         parser.add_argument(
             "--enable-flashinfer-allreduce-fusion",
             action="store_true",
-            help="Enable FlashInfer allreduce fusion for Add_RMSNorm.",
+            help="Enable FlashInfer allreduce fusion with Residual RMSNorm.",
         )
         parser.add_argument(
             "--deepep-mode",