Move args from global_config to environ (#11332)

python/sglang/srt/managers/schedule_batch.py

@@ -37,7 +37,6 @@ import copy
 import dataclasses
 import logging
 import re
-import threading
 import time
 from enum import Enum, auto
 from http import HTTPStatus
@@ -47,7 +46,6 @@ from typing import TYPE_CHECKING, Any, List, Optional, Set, Tuple, Union
 import numpy as np
 import torch
 
-from sglang.global_config import global_config
 from sglang.srt.constrained.base_grammar_backend import BaseGrammarObject
 from sglang.srt.disaggregation.base import BaseKVSender
 from sglang.srt.disaggregation.decode_schedule_batch_mixin import (
@@ -55,6 +53,7 @@ from sglang.srt.disaggregation.decode_schedule_batch_mixin import (
 )
 from sglang.srt.disaggregation.utils import DisaggregationMode
 from sglang.srt.distributed.parallel_state import get_tensor_model_parallel_rank
+from sglang.srt.environ import envs
 from sglang.srt.mem_cache.allocator import (
     BaseTokenToKVPoolAllocator,
     SWATokenToKVPoolAllocator,
@@ -1481,7 +1480,8 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         total_max_new_tokens = sum(r.sampling_params.max_new_tokens for r in self.reqs)
 
         new_estimate_ratio = (
-            total_decoded_tokens + global_config.retract_decode_steps * len(self.reqs)
+            total_decoded_tokens
+            + envs.SGLANG_RETRACT_DECODE_STEPS.get() * len(self.reqs)
         ) / total_max_new_tokens
         new_estimate_ratio = min(1.0, new_estimate_ratio)
 
@@ -1520,7 +1520,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 self.tree_cache.dec_lock_ref(req.last_node)
 
             # NOTE(lsyin): we should use the newly evictable memory instantly.
-            num_tokens = remaing_req_count * global_config.retract_decode_steps
+            num_tokens = remaing_req_count * envs.SGLANG_RETRACT_DECODE_STEPS.get()
             self._evict_tree_cache_if_needed(num_tokens)
 
         req.reset_for_retract()