[deepseek] Enable DP attention + TBO + shared experts fusion (#27510)

python/sglang/srt/batch_overlap/two_batch_overlap.py

@@ -680,6 +680,11 @@ class TboForwardBatchPreparer:
             ):
                 output_dict[key] = None
                 continue
+            elif key == "rids" and len(old_value) != num_seqs:
+                output_dict[key] = old_value[
+                    start_seq_index : min(end_seq_index, len(old_value))
+                ]
+                continue
             assert (
                 len(old_value) == num_seqs
             ), f"{key=} {old_value=} {num_seqs=} {batch=}"

python/sglang/srt/models/deepseek_v2.py

@@ -1339,9 +1339,7 @@ class DeepseekV2MoE(nn.Module):
             return None
 
     def op_gate(self, state):
-        if is_non_idle_and_non_empty(
-            state.forward_batch.forward_mode, state.hidden_states_mlp_input
-        ):
+        if state.hidden_states_mlp_input.shape[0] > 0:
             # router_logits: (num_tokens, n_experts)
             state.router_logits = self.gate(state.hidden_states_mlp_input)
         else:

python/sglang/srt/server_args.py

@@ -7858,11 +7858,6 @@ class ServerArgs:
                 "When enabling two batch overlap, moe_a2a_backend cannot be 'none'."
             )
 
-        if self.enable_two_batch_overlap and self.enforce_shared_experts_fusion:
-            raise ValueError(
-                "--enable-two-batch-overlap and --enforce-shared-experts-fusion cannot be used together."
-            )
-
         # Check communications compression
         if self.enable_quant_communications and self.tp_size == 1:
             raise ValueError(