Deprecate global_server_args_dict (#11331)

python/sglang/srt/layers/logits_processor.py

@@ -38,17 +38,15 @@ from sglang.srt.layers.dp_attention import (
     get_dp_device,
     get_dp_dtype,
     get_dp_hidden_size,
-    get_global_dp_buffer,
     get_local_attention_dp_size,
-    set_dp_buffer_len,
 )
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import (
     CaptureHiddenMode,
     ForwardBatch,
     ForwardMode,
 )
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import dump_to_file, is_npu, use_intel_amx_backend
 
 logger = logging.getLogger(__name__)
@@ -230,8 +228,8 @@ class LogitsProcessor(nn.Module):
         super().__init__()
         self.config = config
         self.logit_scale = logit_scale
-        self.use_attn_tp_group = global_server_args_dict["enable_dp_lm_head"]
-        self.use_fp32_lm_head = global_server_args_dict["enable_fp32_lm_head"]
+        self.use_attn_tp_group = get_global_server_args().enable_dp_lm_head
+        self.use_fp32_lm_head = get_global_server_args().enable_fp32_lm_head
         if self.use_attn_tp_group:
             self.attn_tp_size = get_attention_tp_size()
             self.do_tensor_parallel_all_gather = (
@@ -254,8 +252,8 @@ class LogitsProcessor(nn.Module):
         ):
             self.final_logit_softcapping = None
 
-        self.debug_tensor_dump_output_folder = global_server_args_dict.get(
-            "debug_tensor_dump_output_folder", None
+        self.debug_tensor_dump_output_folder = (
+            get_global_server_args().debug_tensor_dump_output_folder
         )
 
     def compute_logprobs_for_multi_item_scoring(
@@ -372,9 +370,7 @@ class LogitsProcessor(nn.Module):
             logits_metadata = LogitsMetadata.from_forward_batch(logits_metadata)
 
         # Check if multi-item scoring is enabled via server args (only for prefill-only requests)
-        multi_item_delimiter = global_server_args_dict.get(
-            "multi_item_scoring_delimiter"
-        )
+        multi_item_delimiter = get_global_server_args().multi_item_scoring_delimiter
         if multi_item_delimiter is not None and logits_metadata.is_prefill_only:
             return self.compute_logprobs_for_multi_item_scoring(
                 input_ids, hidden_states, lm_head, logits_metadata, multi_item_delimiter