DP Enhancement (#8280)

python/sglang/srt/models/deepseek_v2.py

@@ -550,9 +550,8 @@ class DeepseekV2MoE(nn.Module):
     def forward_deepep(
         self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
     ) -> torch.Tensor:
-        forward_mode = forward_batch.forward_mode
         shared_output = None
-        if is_non_idle_and_non_empty(forward_mode, hidden_states):
+        if hidden_states.shape[0] > 0:
             # router_logits: (num_tokens, n_experts)
             router_logits = self.gate(hidden_states)
             shared_output = self._forward_shared_experts(hidden_states)

python/sglang/srt/models/qwen2_moe.py

@@ -43,10 +43,6 @@ from sglang.srt.layers.communicator import (
     ScatterMode,
 )
 from sglang.srt.layers.dp_attention import (
-    attn_tp_all_gather,
-    attn_tp_reduce_scatter,
-    dp_gather_partial,
-    dp_scatter,
     get_attention_tp_rank,
     get_attention_tp_size,
     get_local_attention_dp_size,

python/sglang/srt/models/qwen3_moe.py

@@ -38,10 +38,6 @@ from sglang.srt.eplb.expert_location_dispatch import ExpertLocationDispatchInfo
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
 from sglang.srt.layers.dp_attention import (
-    attn_tp_all_gather,
-    attn_tp_reduce_scatter,
-    dp_gather_partial,
-    dp_scatter,
     get_attention_tp_rank,
     get_attention_tp_size,
     get_local_attention_dp_size,
@@ -193,8 +189,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
     def forward_deepep(
         self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
     ) -> torch.Tensor:
-        forward_mode = forward_batch.forward_mode
-        if is_non_idle_and_non_empty(forward_mode, hidden_states):
+        if hidden_states.shape[0] > 0:
             # router_logits: (num_tokens, n_experts)
             router_logits, _ = self.gate(hidden_states)
             topk_weights, topk_idx, _ = self.topk(