Fix NPU code predictor device mismatch in concurrent mode (#3453)

Co-authored-by: houzechen <h00875519@china.huawei.com>

vllm_omni/platforms/npu/worker/npu_model_runner.py

@@ -460,7 +460,6 @@ class OmniNPUModelRunner(OmniGPUModelRunner, NPUModelRunner):
                 **talker_kwargs,
             )
         # update the inputs_embeds and code_predictor_codes
-        code_predictor_codes_cpu = code_predictor_codes.detach().to("cpu").contiguous()
         out_key = getattr(self.model, "talker_mtp_output_key", ("codes", "audio"))
         if not isinstance(out_key, tuple) or len(out_key) != 2:
             raise TypeError(f"talker_mtp_output_key must be a 2-tuple, got {type(out_key).__name__}: {out_key!r}")
@@ -468,5 +467,5 @@ class OmniNPUModelRunner(OmniGPUModelRunner, NPUModelRunner):
             req_index = self.input_batch.req_ids.index(req_id)
             start_offset = int(self.query_start_loc.cpu[req_index])
             inputs_embeds[start_offset : start_offset + 1] = req_embeds[idx : idx + 1]
-            update_dict = {out_key[0]: {out_key[1]: code_predictor_codes_cpu[idx : idx + 1]}}
+            update_dict = {out_key[0]: {out_key[1]: code_predictor_codes[idx : idx + 1]}}
             self._merge_additional_information_update(req_id, update_dict)