[XPU] fix(xpu_model_runner): reset seq_lens_encoder to 0 for decode role in PD splitwise mode (#6048)

* fix(xpu_model_runner): reset seq_lens_encoder to 0 for decode role in PD splitwise mode

- Set seq_lens_encoder to 0 when splitwise_role is 'decode' during prefill processing
- This ensures proper continuation of decoding after P generate first token in PD disaggregated architecture
- Fixes potential sequence length inconsistency in PD splitwise deployment scenarios

* format

fastdeploy/worker/xpu_model_runner.py

@@ -565,6 +565,10 @@ class XPUModelRunner(ModelRunnerBase):
                     len(request.output_token_ids) if prefill_end_index >= len(input_ids) else 0
                 )
                 self.share_inputs["pre_ids"][idx : idx + 1] = -1
+                if (
+                    self.fd_config.scheduler_config.splitwise_role == "decode"
+                ):  # In PD, we continue to decode after P generate first token
+                    self.share_inputs["seq_lens_encoder"][idx : idx + 1] = 0
                 has_prefill_task = True
             elif request.task_type.value == RequestType.DECODE.value:  # decode task
                 logger.debug(f"Handle decode request {request} at idx {idx}")