[Speculative Decoding] Support mtp expert-parallel and support different modality deploy (#7018)

* support mtp ep and support different modality

* fix default arg

fastdeploy/worker/gpu_model_runner.py

@@ -2019,6 +2019,13 @@ class GPUModelRunner(ModelRunnerBase):
 
         return prefill_done_idxs
 
+    def _execute_empty_mtp_input(self, forward_meta) -> None:
+        """
+        run ep inference forward with empty input.
+        """
+        for _ in range(self.fd_config.speculative_config.num_model_steps):
+            self.proposer.model.empty_input_forward(forward_meta)
+
     def execute_model(
         self,
         model_forward_batch: Optional[List[Request]] = None,
@@ -2046,6 +2053,12 @@ class GPUModelRunner(ModelRunnerBase):
         model_inputs, p_done_idxs, _ = self._preprocess(model_forward_batch, num_running_requests)
         model_output = self._execute(model_inputs)
         if model_output is None or self.share_inputs["seq_lens_this_time_cpu"].numpy().sum().item() <= 0:
+            if (
+                self.fd_config.speculative_config.method == SpecMethod.MTP
+                and hasattr(self.proposer.model, "empty_input_forward")
+                and self.parallel_config.use_ep
+            ):
+                self._execute_empty_mtp_input(self.forward_meta)
             return
         model_output_data, sampler_output, post_process_event = self._postprocess(
             model_output, p_done_idxs, model_forward_batch, num_running_requests