remove seq_lens_this_time (#4821)

fastdeploy/model_executor/pre_and_post_process.py

@@ -91,12 +91,7 @@ else:
 
 from fastdeploy.output.pooler import PoolerOutput, PoolingSequenceGroupOutput
 from fastdeploy.output.stream_transfer_data import DecoderState, StreamTransferData
-from fastdeploy.worker.output import (
-    LogprobsTensors,
-    ModelOutputData,
-    ModelRunnerOutput,
-    SamplerOutput,
-)
+from fastdeploy.worker.output import LogprobsTensors, ModelOutputData, SamplerOutput
 
 DISABLE_RECOVER = envs.FD_DISABLED_RECOVER == "1"
 
@@ -208,7 +203,7 @@ def pre_process(
     """
     token_num = paddle.sum(seq_lens_this_time)
 
-    if current_platform.is_cuda() and not speculative_decoding:
+    if (current_platform.is_cuda() or current_platform.is_iluvatar()) and not speculative_decoding:
         # Note(ZKK): This case's code is very simple!
         ids_remove_padding, batch_id_per_token, cu_seqlens_q, cu_seqlens_k = get_padding_offset(
             input_ids, token_num, seq_lens_this_time
@@ -322,7 +317,7 @@ def post_process_normal(
     async_output_queue: queue.Queue = None,
     think_end_id: int = -1,
     line_break_id: int = -1,
-) -> ModelRunnerOutput:
+):
     """Post-processing steps after completing a single token generation."""
     if think_end_id > 0:
         limit_thinking_content_length(

fastdeploy/worker/gpu_model_runner.py

@@ -2050,7 +2050,7 @@ class GPUModelRunner(ModelRunnerBase):
         self,
         model_forward_batch: Optional[List[Request]] = None,
         num_running_requests: int = None,
-    ) -> Optional[ModelRunnerOutput]:
+    ) -> None:
         """
         The Entrance of model execute.
         Args:
@@ -2140,10 +2140,6 @@ class GPUModelRunner(ModelRunnerBase):
                 async_output_queue=self.async_output_queue,
             )
 
-            self.seq_lens_this_time_buffer[:num_running_requests].copy_(
-                self.share_inputs["seq_lens_this_time"][:num_running_requests], False
-            )
-
             return None
         else:
             hidden_states = rebuild_padding(
@@ -2305,9 +2301,6 @@ class GPUModelRunner(ModelRunnerBase):
                     self.speculative_config.num_speculative_tokens,
                 )
 
-            self.seq_lens_this_time_buffer[:num_running_requests].copy_(
-                self.share_inputs["seq_lens_this_time"][:num_running_requests], False
-            )
             return None
 
     def _pool(self, hidden_states: paddle.Tensor, num_running_requests: int) -> Optional[ModelRunnerOutput]: