fix: input and input duplication feed into expert and shared_expert respectively.

vllm/envs.py

@@ -1268,7 +1268,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
                  ("true", "1")),
     
     "VLLM_DISABLE_SHARED_EXPERTS_STREAM": lambda: bool(
-        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "0"))
+        int(os.getenv("VLLM_DISABLE_SHARED_EXPERTS_STREAM", "1"))
     ),
 }
 

vllm/model_executor/layers/fused_moe/layer.py

@@ -1483,6 +1483,7 @@ class FusedMoE(torch.nn.Module):
 
     def forward(self, hidden_states: torch.Tensor,
                 router_logits: torch.Tensor,
+                hidden_states_copy: Optional[torch.Tensor] = None,
                 shared_output: Optional[torch.Tensor] = None,
                 i_q: Optional[torch.Tensor] = None,
                 i_s: Optional[torch.Tensor] = None, **_
@@ -1500,9 +1501,9 @@ class FusedMoE(torch.nn.Module):
         else:
             if current_platform.is_tpu():
                 assert i_q is None and i_s is None, "moe.quant fused not support TPU now" 
-                return self.forward_impl(hidden_states, router_logits)
+                return self.forward_impl(hidden_states, hidden_states_copy, router_logits)
             else:
-                return torch.ops.vllm.moe_forward_shared(hidden_states, router_logits, self.layer_name)
+                return torch.ops.vllm.moe_forward_shared(hidden_states, router_logits, hidden_states_copy, self.layer_name)
 
     def forward_impl_chunked(self, full_hidden_states: torch.Tensor,
                              full_router_logits: torch.Tensor):
@@ -1623,11 +1624,14 @@ class FusedMoE(torch.nn.Module):
 
     def forward_impl(self, hidden_states: torch.Tensor,
                      router_logits: torch.Tensor,
+                     hidden_states_copy: Optional[torch.Tensor] = None, 
                      shared_output: Optional[torch.Tensor] = None,
                      i_q: Optional[torch.Tensor] = None,
                      i_s: Optional[torch.Tensor] = None, **_) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
         assert self.quant_method is not None
 
+        if self.shared_experts_stream is not None and hidden_states_copy is not None:
+            hidden_states_copy.record_stream(self.shared_experts_stream)
         use_shared_experts_stream, hidden_states_clone = self._maybe_setup_shared_experts_stream(hidden_states,
                 self.shared_experts is not None and self.shared_experts_stream is not None,
                 self.moe_parallel_config.use_pplx_kernels or self.moe_parallel_config.use_deepep_ll_kernels)    
@@ -1707,7 +1711,8 @@ class FusedMoE(torch.nn.Module):
                 with torch.cuda.stream(self.shared_experts_stream):
                     # Note that hidden_states clone() is necessary here to avoid
                     # conflict with the main stream
-                    shared_output = self.shared_experts(hidden_states)
+                    assert hidden_states_copy is not None
+                    shared_output = self.shared_experts(hidden_states_copy)
                 torch.cuda.current_stream().wait_stream(self.shared_experts_stream)
 
                 final_hidden_states = (
@@ -1822,11 +1827,12 @@ direct_register_custom_op(
 def moe_forward_shared(
             hidden_states: torch.Tensor, 
             router_logits: torch.Tensor, 
+            hidden_states_copy: torch.Tensor,
             layer_name: str) -> tuple[torch.Tensor, torch.Tensor]:
     forward_context: ForwardContext = get_forward_context()
     self = forward_context.no_compile_layers[layer_name]
     assert self.quant_method is not None
-    out = self.forward_impl(hidden_states, router_logits)
+    out = self.forward_impl(hidden_states, router_logits, hidden_states_copy)
     return out
         
 
@@ -1834,6 +1840,7 @@ def moe_forward_shared(
 def moe_forward_shared_fake(
             hidden_states: torch.Tensor,
             router_logits: torch.Tensor,
+            hidden_states_copy: torch.Tensor, 
             layer_name: str) -> tuple[torch.Tensor, torch.Tensor]:
     shared_out = torch.empty_like(hidden_states)
     fused_out = torch.empty_like(hidden_states)

vllm/model_executor/layers/fused_moe/shared_fused_moe.py

@@ -42,12 +42,14 @@ class SharedFusedMoE(FusedMoE):
 
     def forward(self, 
             hidden_states: torch.Tensor,
-            router_logits: torch.Tensor
+            router_logits: torch.Tensor,
+            hidden_states_copy: torch.Tensor
     ) -> tuple[torch.Tensor, torch.Tensor]:
 
         shared_out, fused_out = super().forward(
-            hidden_states=hidden_states,
-            router_logits=router_logits,
+            hidden_states = hidden_states,
+            router_logits = router_logits,
+            hidden_states_copy = hidden_states_copy
         )
         # # ensure early TP reduction of shared expert outputs when required
         # if (

vllm/model_executor/models/deepseek_v2.py

@@ -244,7 +244,11 @@ class DeepseekV2MoE(nn.Module):
             num_tokens, hidden_dim = hidden_states.shape
             hidden_states = hidden_states.view(-1, hidden_dim)
             router_logits, _ = self.gate(hidden_states)
-            shared_output, final_hidden_states = self.experts(hidden_states=hidden_states, router_logits=router_logits)
+            hidden_states_copy = hidden_states.clone()
+            shared_output, final_hidden_states = self.experts(
+                    hidden_states=hidden_states,
+                    router_logits=router_logits,
+                    hidden_states_copy = hidden_states_copy)
             
             if self.shared_experts is None:
                 assert shared_output is None
@@ -261,15 +265,6 @@ class DeepseekV2MoE(nn.Module):
                 assert shared_output is not None
                 final_hidden_states += shared_output
 
-            # if self.is_sequence_parallel:
-            #     final_hidden_states = tensor_model_parallel_all_gather(
-            #         final_hidden_states, 0
-            #     )
-            #     final_hidden_states = final_hidden_states[:num_tokens]
-            # elif self.tp_size > 1:
-            #     final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
-            #         final_hidden_states
-            #     )
             if self.tp_size > 1:
                 if envs.VLLM_ENABLE_TBO:
                     final_hidden_states = self.tbo_all_reduce(final_hidden_states)