config: delete the dead get_server_args() bindings across the repo (#33888)

python/sglang/srt/layers/attention/linear/inkling_sconv_backend.py

@@ -115,7 +115,6 @@ class InklingShortConvAttnBackend(ShortConvAttnBackend):
         """Sized from the CONFIGURED capture shapes, once, never reallocated:
         growing a buffer after a graph captured it moves the address that graph
         reads, and prefill captures before the decode runner reports its bounds."""
-        server_args = get_server_args()
         cuda_graph_config = get_exec().graph.cuda_graph_config
         decode_bs: list[int] = []
         prefill_tokens: list[int] = []

python/sglang/srt/layers/moe/topk.py

@@ -1374,7 +1374,7 @@ def _eplb_remap_enabled() -> bool:
     from sglang.srt.runtime_context import get_server_args
 
     try:
-        server_args = get_server_args()
+        get_server_args()  # probes that a config is published
     except ValueError:
         # Global server args are not initialized outside the server runtime
         # (e.g. in unit tests that call select_experts directly). In that case

python/sglang/srt/managers/schedule_batch.py

@@ -3244,7 +3244,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def maybe_evict_swa(self):
         if self.tree_cache.supports_swa():
             sliding_window_size = self.tree_cache.sliding_window_size
-            server_args = get_server_args()
 
             release_leaf_lock = (
                 envs.SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW.get()

python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py

@@ -21,7 +21,7 @@ from sglang.srt.environ import envs
 from sglang.srt.mem_cache.base_swa_memory_pool import BaseSWAKVPool
 from sglang.srt.mem_cache.deepseek_v4_compress_state import CompressStatePool
 from sglang.srt.mem_cache.memory_pool import KVCache
-from sglang.srt.runtime_context import get_exec, get_server_args, get_spec
+from sglang.srt.runtime_context import get_exec, get_spec
 from sglang.srt.utils import ceil_div, is_hip
 
 logger = logging.getLogger(__name__)
@@ -575,7 +575,6 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
             self.swa_kv_pool = None
             self.c4_kv_pool = None
             self.c128_kv_pool = None
-            server_args = get_server_args()
             spec_extra = (
                 (get_spec().speculative_num_draft_tokens - 1)
                 if get_spec().speculative_algorithm is not None
@@ -658,7 +657,6 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
         self.full_to_swa_index_mapping = full_to_swa_index_mapping
 
     def get_ring_size(self, compress_ratio: int) -> int:
-        server_args = get_server_args()
         is_speculative = get_spec().speculative_algorithm is not None
         return get_compress_state_ring_size(compress_ratio, is_speculative)
 

python/sglang/srt/mem_cache/storage/flexkv/flexkv_radix_cache.py

@@ -391,9 +391,6 @@ class FlexKVRadixCache(RadixCache):
             return
 
         # Compute the committed prefix mirroring LMCRadixCache's logic.
-        from sglang.srt.runtime_context import get_server_args
-
-        global_server_args = get_server_args()
         topk = get_spec().speculative_eagle_topk
         enable_kv_committed_len = topk is None or topk == 1
         if enable_kv_committed_len:

python/sglang/srt/mem_cache/storage/lmcache/lmc_radix_cache.py

@@ -16,7 +16,7 @@ from sglang.srt.mem_cache.base_prefix_cache import (
     MatchResult,
 )
 from sglang.srt.mem_cache.radix_cache import RadixCache, RadixKey, TreeNode
-from sglang.srt.runtime_context import get_memory, get_server_args, get_spec
+from sglang.srt.runtime_context import get_memory, get_spec
 from sglang.srt.utils import create_device_stream, device_stream_context
 
 try:
@@ -447,7 +447,6 @@ class LMCRadixCache(RadixCache):
                 self.lmcache_connector.end_session(req.rid)
             return
 
-        global_server_args = get_server_args()
         topk = get_spec().speculative_eagle_topk
         enable_kv_committed_len = topk is None or topk == 1
         if enable_kv_committed_len:

python/sglang/srt/model_loader/loader.py

@@ -586,7 +586,6 @@ class DefaultModelLoader(BaseModelLoader):
                 hf_weights_files,
             )
         elif use_safetensors:
-            server_args = get_server_args()
             weight_loader_disable_mmap = get_model().weight_loader_disable_mmap
             weight_loader_prefetch = get_model().weight_loader_prefetch_checkpoints
             prefetch_num_threads = get_model().weight_loader_prefetch_num_threads

python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py

@@ -69,7 +69,7 @@ from sglang.srt.models.deepseek_common.utils import (
     _use_aiter_bpreshuffle_gfx95,
     _use_aiter_gfx95,
 )
-from sglang.srt.runtime_context import get_exec, get_parallel, get_server_args
+from sglang.srt.runtime_context import get_exec, get_parallel
 from sglang.srt.state_capturer.indexer_topk import (
     maybe_capture_indexer_topk,
 )
@@ -1280,7 +1280,6 @@ class DeepseekMLAForwardMixin:
         """
         Check if we should skip rope and use fused rope+cache path for TileLang DSA on gfx95.
         """
-        server_args = get_server_args()
         return (
             _use_aiter_gfx95
             and self.current_attention_backend in ("dsa", "nsa")

python/sglang/srt/models/deepseek_v2.py

@@ -954,7 +954,6 @@ class DeepseekV2MoE(nn.Module):
         has_shared_output = (
             hidden_states.shape[0] > 0 and self.num_fused_shared_experts == 0
         )
-        server_args = get_server_args()
         dispatch_info = (
             ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id)
             if get_exec().moe.enable_eplb and not self.is_nextn
@@ -1055,7 +1054,6 @@ class DeepseekV2MoE(nn.Module):
             self.shared_experts.gate_up_proj
         ):
             return self.forward_cpu(hidden_states)
-        server_args = get_server_args()
         dispatch_info = (
             ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id)
             if get_exec().moe.enable_eplb and not self.is_nextn
@@ -2972,7 +2970,6 @@ class DeepseekV2ForCausalLM(nn.Module, DeepseekV2WeightLoaderMixin):
         self, architecture: str = "DeepseekV3ForCausalLM"
     ):
         self.num_fused_shared_experts = 0
-        server_args = get_server_args()
 
         if get_exec().moe.disable_shared_experts_fusion:
             return

python/sglang/srt/models/minimax_m3_vl.py

@@ -43,7 +43,7 @@ from sglang.srt.models.minimax_vl_common import (
     merge_vit_qkv_weights,
 )
 from sglang.srt.models.utils import WeightsMapper
-from sglang.srt.runtime_context import get_exec, get_mm, get_parallel, get_server_args
+from sglang.srt.runtime_context import get_exec, get_mm, get_parallel
 from sglang.srt.utils import add_prefix, get_device_sm, is_cuda, log_info_on_rank0
 from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
@@ -134,7 +134,6 @@ class MiniMaxM3SparseForConditionalGeneration(nn.Module):
 
     def _determine_num_fused_shared_experts(self) -> None:
         text_config = self.config.text_config
-        server_args = get_server_args()
         if get_exec().moe.disable_shared_experts_fusion:
             return
 

python/sglang/srt/sampling/sampling_batch_info.py

@@ -12,7 +12,7 @@ from sglang.srt.constrained.base_grammar_backend import (
     GrammarMask,
     GrammarRow,
 )
-from sglang.srt.runtime_context import get_exec, get_server_args
+from sglang.srt.runtime_context import get_exec
 from sglang.srt.sampling.custom_logit_processor import CustomLogitProcessor
 from sglang.srt.sampling.penaltylib.repetition_penalty import apply_scaling_penalties
 from sglang.srt.sampling.sampling_params import TOP_K_ALL
@@ -85,7 +85,6 @@ class SamplingBatchInfo:
 
     @classmethod
     def from_schedule_batch(cls, batch: ScheduleBatch, vocab_size: int):
-        global_server_args = get_server_args()
         enable_deterministic = get_exec().deterministic.enable_deterministic_inference
 
         reqs = batch.reqs

python/sglang/srt/state_capturer/indexer_topk.py

@@ -21,8 +21,6 @@ class IndexerTopkCapturer(BaseTopkCapturer):
         max_running_requests: int,
         device: str,
     ):
-        from sglang.srt.runtime_context import get_server_args
-
         self.num_indexer_layers = num_indexer_layers
         self.index_topk = index_topk
 
@@ -31,7 +29,6 @@ class IndexerTopkCapturer(BaseTopkCapturer):
 
         # DP-attention capture is per-rank-local: each rank writes [:local_batch, ...]
         # to its own device_cache, so the buffer only needs to fit one rank's batch.
-        server_args = get_server_args()
         max_batch_size = max(get_schedule().chunked_prefill_size, max_running_requests)
 
         super().__init__(

python/sglang/srt/state_capturer/routed_experts.py

@@ -16,7 +16,6 @@ from sglang.srt.runtime_context import (
     get_exec,
     get_parallel,
     get_schedule,
-    get_server_args,
 )
 from sglang.srt.state_capturer.base import BaseTopkCapturer
 
@@ -40,7 +39,6 @@ class RoutedExpertsCapturer(BaseTopkCapturer):
         max_running_requests: int,
         device: str,
     ) -> Optional["RoutedExpertsCapturer"]:
-        server_args = get_server_args()
         if not get_exec().features.enable_return_routed_experts:
             return None
         if not get_exec().moe.disable_shared_experts_fusion and hasattr(