[Deepseek] Pass hidden_states_fp4 to shared_experts (NVIDIA#3819)

Signed-off-by: Hao Lu <[email protected]@users.noreply.github.com>

Author: hlu1

tensorrt_llm/_torch/models/modeling_deepseekv3.py

@@ -466,7 +466,8 @@ def forward(
             assert not self.use_dp
 
         def _compute_shared_output():
-            shared_output = self.shared_experts(hidden_states)
+            shared_output = self.shared_experts(hidden_states_fp4
+                                                or hidden_states)
             if self.shared_output_scale is not None:
                 shared_output *= self.shared_output_scale
             return shared_output