DeepseekV2MoE: defer shared experts when routed kernel is non-mutating (#25279)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

python/sglang/srt/models/deepseek_v2.py

@@ -826,10 +826,9 @@ class DeepseekV2MoE(nn.Module):
             if server_args.enable_eplb
             else None
         )
+        defer_shared = not self.experts.moe_runner_config.inplace
         if hidden_states.shape[0] > 0:
-            if (
-                not self._fuse_shared_experts_inside_sbo
-            ):  # TODO: check if it supports mtp
+            if not defer_shared and not self._fuse_shared_experts_inside_sbo:
                 shared_output = self._forward_shared_experts(
                     hidden_states, gemm_output_zero_allocator
                 )
@@ -894,6 +893,15 @@ class DeepseekV2MoE(nn.Module):
             # fused in biased_grouped_topk so we can skip here
             final_hidden_states *= self.routed_scaling_factor
 
+        if (
+            defer_shared
+            and hidden_states.shape[0] > 0
+            and not self._fuse_shared_experts_inside_sbo
+        ):
+            shared_output = self._forward_shared_experts(
+                hidden_states, gemm_output_zero_allocator
+            )
+
         final_hidden_states = maybe_fuse_routed_scale_and_shared_add(
             self.experts,
             final_hidden_states,