Improve code styles (#4021)

python/sglang/srt/managers/data_parallel_controller.py

@@ -198,8 +198,6 @@ class DataParallelController:
         self.max_total_num_tokens = scheduler_info[0]["max_total_num_tokens"]
         self.max_req_input_len = scheduler_info[0]["max_req_input_len"]
 
-        print(f"{scheduler_info=}")
-
     def round_robin_scheduler(self, req):
         self.workers[self.round_robin_counter].send_pyobj(req)
         self.round_robin_counter = (self.round_robin_counter + 1) % len(self.workers)
@@ -222,7 +220,6 @@ class DataParallelController:
                         TokenizedEmbeddingReqInput,
                     ),
                 ):
-                    logger.info("dispatching")
                     self.dispatching(recv_req)
                 else:
                     # Send other control messages to first worker of tp group

python/sglang/srt/managers/io_struct.py

@@ -158,7 +158,7 @@ class GenerateReqInput:
                 # Expand parallel_sample_num
                 num = self.batch_size * self.parallel_sample_num
 
-            if self.image_data is None:
+            if not self.image_data:
                 self.image_data = [None] * num
             elif not isinstance(self.image_data, list):
                 self.image_data = [self.image_data] * num

python/sglang/srt/managers/schedule_batch.py

@@ -282,6 +282,8 @@ class Req:
         # If we want to abort the request in the middle of the event loop, set this to true
         # Note: We should never set finished_reason in the middle, the req will get filtered and never respond
         self.to_abort = False
+        # This carries the error message for `.to_abort` and will be attached to the finished_reason at the end of the event loop
+        self.to_abort_message: str = "Unknown error"
         self.stream = stream
         self.eos_token_ids = eos_token_ids
 
@@ -359,8 +361,6 @@ class Req:
         # The tokens is prefilled but need to be considered as decode tokens
         # and should be updated for the decode logprobs
         self.last_update_decode_tokens = 0
-        # The relative logprob_start_len in an extend batch
-        self.extend_logprob_start_len = 0
 
         # Embedding (return values)
         self.embedding = None
@@ -377,9 +377,6 @@ class Req:
         self.spec_verify_ct = 0
         self.lora_path = lora_path
 
-        # This carries the error message for `.to_abort` and will be attached to the finished_reason at the end of the event loop
-        self.to_abort_message: str = "Unknown error"
-
     @property
     def seqlen(self):
         return len(self.origin_input_ids) + len(self.output_ids)

python/sglang/srt/managers/scheduler.py

@@ -358,7 +358,6 @@ class Scheduler:
         self.cum_spec_accept_count = 0
         self.last_decode_stats_tic = time.time()
         self.return_health_check_ct = 0
-        self.stream_interval = server_args.stream_interval
         self.current_stream = torch.get_device_module(self.device).current_stream()
         if self.device == "cpu":
             self.current_stream.synchronize = lambda: None  # No-op for CPU
@@ -444,11 +443,6 @@ class Scheduler:
                 },
             )
 
-        # The largest prefill length of a single request
-        self._largest_prefill_len: int = 0
-        # The largest context length (prefill + generation) of a single request
-        self._largest_prefill_decode_len: int = 0
-
         # Init request dispatcher
         self._request_dispatcher = TypeBasedDispatcher(
             [
@@ -2309,8 +2303,6 @@ def run_scheduler_process(
     if get_bool_env_var("SGLANG_SET_CPU_AFFINITY"):
         set_gpu_proc_affinity(server_args.tp_size, server_args.nnodes, gpu_id)
 
-    parent_process = psutil.Process().parent()
-
     # Create a scheduler and run the event loop
     try:
         scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, dp_rank)