[DeepSeek V3] Reland: run routed experts on main stream in dual-stream MoE (#29463)

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

python/sglang/srt/models/deepseek_v2.py

@@ -926,11 +926,13 @@ class DeepseekV2MoE(nn.Module):
         *,
         use_flashinfer_trtllm_bypass: bool = False,
     ) -> torch.Tensor:
+        # Note(kpham-sgl): launch the shared expert BEFORE the routed call.
+        # The routed deep_gemm pre-permute calls `dispose_tensor` which
+        # `set_()`s `hidden_states` to empty (host-side); any later kernel
+        # launch consuming `hidden_states` then captures `data_ptr() == 0`
+        # into the decode CUDA graph and replays from null.
         current_stream = torch.cuda.current_stream()
         self.alt_stream.wait_stream(current_stream)
-        shared_output = self._forward_shared_experts(
-            hidden_states, gemm_output_zero_allocator
-        )
         server_args = get_global_server_args()
         dispatch_info = (
             ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id)
@@ -938,6 +940,9 @@ class DeepseekV2MoE(nn.Module):
             else None
         )
         with torch.cuda.stream(self.alt_stream):
+            shared_output = self._forward_shared_experts(
+                hidden_states, gemm_output_zero_allocator
+            )
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states, gemm_output_zero_allocator)
         if use_flashinfer_trtllm_bypass:
@@ -969,9 +974,7 @@ class DeepseekV2MoE(nn.Module):
                 hidden_states, topk_output
             )
         elif use_flashinfer_trtllm_bypass:
-                final_hidden_states = self.experts.forward_impl(
-                    hidden_states, topk_output
-                )
+            final_hidden_states = self.experts.forward_impl(hidden_states, topk_output)
         else:
             final_hidden_states = self.experts(hidden_states, topk_output)
         if (