[refactor] Retire the legacy config accessor and the remaining process singletons (#30493)

python/sglang/srt/arg_groups/overrides.py

@@ -380,7 +380,7 @@ def _deepseek_family_overrides(server_args: Any, hf_config: Any) -> dict:
                 )
 
             # Deferred import to avoid a circular import at module-load
-            # time (dsa.utils imports get_global_server_args).
+            # time (dsa.utils imports the runtime-context accessors).
             from sglang.srt.layers.attention.dsa.utils import (
                 aiter_can_use_preshuffle_paged_mqa,
             )

python/sglang/srt/batch_overlap/two_batch_overlap.py

@@ -40,7 +40,6 @@ from sglang.srt.model_executor.forward_batch_info import (
 )
 from sglang.srt.model_executor.forward_context import get_attn_backend
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.spec_info import SpecInput
 from sglang.srt.utils import BumpAllocator, empty_context, get_bool_env_var, is_hip
 
@@ -184,7 +183,7 @@ def _update_device_and_sum_field_from_cpu_field(
         cpu_value
         if isinstance(cpu_value, torch.Tensor)
         else torch.tensor(cpu_value, dtype=old_device_value.dtype)
-    ).to(device=get_global_server_args().device, non_blocking=True)
+    ).to(device=get_server_args().device, non_blocking=True)
     setattr(batch, device_field, new_device_value)
 
     if sum_field is not None:
@@ -336,7 +335,7 @@ def compute_split_indices_for_cuda_graph_replay(
 class TboCudaGraphRunnerPlugin:
     def __init__(self):
         self._tbo_children_num_token_non_padded = torch.zeros(
-            (2,), dtype=torch.int32, device=get_global_server_args().device
+            (2,), dtype=torch.int32, device=get_server_args().device
         )
 
     def capture_one_batch_size(self, batch: ForwardBatch, num_tokens: int):
@@ -760,7 +759,7 @@ class TboForwardBatchPreparer:
 
         # TODO improve, e.g. unify w/ `init_raw`
         if (
-            get_global_server_args().moe_dense_tp_size == 1
+            get_server_args().moe_dense_tp_size == 1
             and batch.global_dp_buffer_len is not None
         ):
             sum_len = end_token_index - start_token_index
@@ -835,7 +834,7 @@ class TboForwardBatchPreparer:
         value_a = min(tbo_split_token_index, num_token_non_padded)
         value_b = max(0, num_token_non_padded - tbo_split_token_index)
         return torch.tensor([value_a, value_b], dtype=torch.int32).to(
-            device=get_global_server_args().device, non_blocking=True
+            device=get_server_args().device, non_blocking=True
         )
 
     @classmethod

python/sglang/srt/debug_utils/dumper.py

@@ -1783,9 +1783,9 @@ class _SGLangPlugin(_FrameworkPlugin):
             return None
 
         try:
-            from sglang.srt.server_args import get_global_server_args
+            from sglang.srt.runtime_context import get_server_args
 
-            args = get_global_server_args()
+            args = get_server_args()
             if args is None:
                 return None
 

python/sglang/srt/distributed/device_communicators/pymscclpp.py

@@ -14,7 +14,7 @@ from sglang.srt.compilation.compile_phase import (
 from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph import (
     is_in_tc_piecewise_cuda_graph,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 logger = logging.getLogger(__name__)
 
@@ -25,7 +25,7 @@ class PyMscclppCommunicator:
 
     def _is_symm_mem_enabled(self) -> bool:
         try:
-            return get_global_server_args().enable_symm_mem
+            return get_server_args().enable_symm_mem
         except ValueError:
             return False
 

python/sglang/srt/distributed/device_communicators/pynccl_allocator.py

@@ -15,7 +15,7 @@ from torch.cuda.memory import (
 
 from sglang.srt.distributed.parallel_state import GroupCoordinator
 from sglang.srt.environ import envs
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils.common import torch_release
 
 after_2_8_0 = torch_release >= (2, 8)
@@ -159,7 +159,7 @@ _register_func = None
 
 def is_symmetric_memory_enabled():
     try:
-        return get_global_server_args().enable_symm_mem
+        return get_server_args().enable_symm_mem
     except ValueError:
         return False
 

python/sglang/srt/distributed/device_communicators/triton_symm_mem_ag.py

@@ -423,9 +423,9 @@ def recommended_max_tokens(include_prefill: bool, floor: int = 0) -> int:
     NCCL. Covers the spec-decode batch plus, if ``include_prefill``, a prefill
     chunk. Returns ``floor`` if server args are unavailable."""
     try:
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
-        sa = get_global_server_args()
+        sa = get_server_args()
 
         def g(name: str) -> int:
             v = getattr(sa, name, 0)

python/sglang/srt/distributed/utils.py

@@ -19,19 +19,13 @@ from torch.distributed import TCPStore
 
 logger = logging.getLogger(__name__)
 
-# Global TCPStore that is created during distributed initialization
-# This is the single shared store that all components should use
-_global_tcp_store: Optional[TCPStore] = None
-
 
 def set_global_tcp_store(store: TCPStore) -> None:
-    """Set the global TCPStore instance.
+    """Install the shared TCPStore created during distributed initialization;
+    the handle lives on ``ctx.resources``."""
+    from sglang.srt.runtime_context import get_resources
 
-    This should be called during distributed initialization to make
+    get_resources().tcp_store = store
-    the store available to all components that need it.
-    """
-    global _global_tcp_store
-    _global_tcp_store = store
     logger.info("Global TCPStore has been set")
 
 
@@ -45,15 +39,15 @@ def get_global_tcp_store() -> Optional[TCPStore]:
     Returns:
         The global TCPStore instance, or None if not initialized yet.
     """
-    global _global_tcp_store
+    from sglang.srt.runtime_context import get_resources
 
-    if _global_tcp_store is None:
+    store = get_resources().tcp_store
+    if store is None:
         logger.warning(
             "Global TCPStore not found. Make sure init_distributed_environment "
             "was called with a tcp:// init method."
         )
-
+    return store
-    return _global_tcp_store
 
 
 def ensure_divisibility(numerator, denominator):

python/sglang/srt/entrypoints/engine.py

@@ -166,7 +166,7 @@ def init_tokenizer_manager(
         if getattr(server_args, attr) != "auto":
             continue
         if suggested is not None:
-            setattr(server_args, attr, suggested)
+            server_args.override(source="template-detection", **{attr: suggested})
             logger.info(
                 f"Auto-detected --{attr.replace('_', '-')} as '{suggested}' from chat template"
             )
@@ -175,7 +175,7 @@ def init_tokenizer_manager(
                 f"--{attr.replace('_', '-')}=auto specified but could not detect "
                 f"{label} from chat template. Disabling {label}."
             )
-            setattr(server_args, attr, None)
+            server_args.override(source="template-detection", **{attr: None})
 
     return tokenizer_manager, template_manager
 

python/sglang/srt/eplb/expert_location_dispatch.py

@@ -18,7 +18,7 @@ from typing import Literal, Optional
 import torch
 
 from sglang.srt.eplb.expert_location import get_global_expert_location_metadata
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import is_hip
 
 _is_hip = is_hip()
@@ -37,7 +37,7 @@ class ExpertLocationDispatchInfo:
 
     @classmethod
     def init_new(cls, layer_id: int):
-        ep_dispatch_algorithm = get_global_server_args().ep_dispatch_algorithm
+        ep_dispatch_algorithm = get_server_args().ep_dispatch_algorithm
         expert_location_metadata = get_global_expert_location_metadata()
         assert expert_location_metadata is not None
 

python/sglang/srt/eplb/expert_location_updater.py

@@ -25,7 +25,7 @@ from sglang.srt.eplb.expert_location import (
     ExpertLocationMetadata,
     get_global_expert_location_metadata,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import get_bool_env_var, get_int_env_var, is_hip
 
 logger = logging.getLogger(__name__)
@@ -108,7 +108,7 @@ def _update_expert_weights_with_canary(
         canary_tensor = (
             _get_canary_value(old_expert_location_metadata, layer_id)
             .clone()
-            .to(device=get_global_server_args().device, non_blocking=True)
+            .to(device=get_server_args().device, non_blocking=True)
         )
         routed_experts_weights_of_layer[layer_id].append(canary_tensor)
 

python/sglang/srt/hardware_backend/mlx/model_runner.py

@@ -49,7 +49,7 @@ from sglang.srt.hardware_backend.mlx.kv_cache import (
     uses_sliding_window_attention,
 )
 from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 logger = logging.getLogger(__name__)
 
@@ -281,7 +281,7 @@ class MlxModelRunner:
         ):
             return None
 
-        chunk_size = get_global_server_args().mamba_cache_chunk_size
+        chunk_size = get_server_args().mamba_cache_chunk_size
         track_len = prefix_len + (new_token_count // chunk_size) * chunk_size
         branching_len = getattr(req, "mamba_branching_seqlen", None)
         if (

python/sglang/srt/hardware_backend/musa/attention/flashattention_backend.py

@@ -23,7 +23,7 @@ from sglang.srt.layers.utils.cp_utils import (
     cp_allgather_and_save_kv_cache,
 )
 from sglang.srt.mem_cache.memory_pool import KVWriteLoc
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
@@ -515,7 +515,7 @@ class MusaFlashAttentionBackend(FlashAttentionBackend):
                 and not forward_batch.forward_mode.is_draft_extend_v2()
             ):
                 if forward_batch.attn_attend_prefix_cache:
-                    assert not get_global_server_args().disable_chunked_prefix_cache
+                    assert not get_server_args().disable_chunked_prefix_cache
                     assert forward_batch.prefix_chunk_idx is not None
                     assert forward_batch.prefix_chunk_cu_seq_lens is not None
                     assert forward_batch.prefix_chunk_max_seq_lens is not None

python/sglang/srt/hardware_backend/npu/attention/ascend_dsv4_backend.py

@@ -1362,9 +1362,9 @@ class DeepseekV4AscendAttnBackend(
             or forward_batch.forward_mode.is_draft_extend_v2()
         ):
             B = forward_batch.batch_size
-            from sglang.srt.server_args import get_global_server_args
+            from sglang.srt.runtime_context import get_server_args
 
-            n_draft = get_global_server_args().speculative_num_draft_tokens or 1
+            n_draft = get_server_args().speculative_num_draft_tokens or 1
             actual_q = torch.arange(
                 n_draft, B * n_draft + 1, n_draft, dtype=torch.int32, device=device
             )
@@ -1409,9 +1409,9 @@ class DeepseekV4AscendAttnBackend(
             forward_batch.forward_mode.is_target_verify()
             or forward_batch.forward_mode.is_draft_extend_v2()
         ):
-            from sglang.srt.server_args import get_global_server_args
+            from sglang.srt.runtime_context import get_server_args
 
-            max_seqlen_q = get_global_server_args().speculative_num_draft_tokens or 1
+            max_seqlen_q = get_server_args().speculative_num_draft_tokens or 1
         else:
             max_seqlen_q = 1
         return self._kernel_metadata_from_parts(

python/sglang/srt/hardware_backend/npu/graph_runner/vit_npu_graph_runner.py

@@ -27,7 +27,7 @@ from sglang.srt.distributed.device_communicators.pynccl_allocator import (
 )
 from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.multimodal.vit_cuda_graph_runner import ViTCudaGraphRunner
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 
 class ViTNpuGraphRunner(ViTCudaGraphRunner):
@@ -70,7 +70,7 @@ class ViTNpuGraphRunner(ViTCudaGraphRunner):
         graph = torch_npu.npu.NPUGraph()
         vit = self.vit
 
-        override_backend = get_global_server_args().mm_attention_backend
+        override_backend = get_server_args().mm_attention_backend
         with torch_npu.npu.graph(graph, pool=ViTNpuGraphRunner._graph_memory_pool):
             y = None
             deepstack_outs: List[torch.Tensor] = []

python/sglang/srt/layers/activation.py

@@ -33,8 +33,7 @@ from sglang.srt.model_executor.cuda_graph_config import (
     Phase,
     check_cuda_graph_backend,
 )
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
@@ -90,7 +89,7 @@ logger = logging.getLogger(__name__)
 class SiluAndMul(MultiPlatformOp):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        if get_global_server_args().rl_on_policy_target is not None:
+        if get_server_args().rl_on_policy_target is not None:
             self._forward_method = self.forward_native
         elif _use_aiter and envs.SGLANG_OPT_USE_AITER_SILU_MUL.get():
             self._forward_method = self.forward_aiter

python/sglang/srt/layers/attention/dsa/dsa_indexer.py

@@ -112,7 +112,7 @@ from sglang.srt.model_executor.forward_context import (
     get_token_to_kv_pool,
 )
 from sglang.srt.model_executor.runner import get_is_capture_mode
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 _use_ag_after_qlora = envs.SGLANG_USE_AG_AFTER_QLORA.get()
 if TYPE_CHECKING:
@@ -134,7 +134,7 @@ def _is_in_piecewise_or_breakable_cuda_graph() -> bool:
 
 def _uses_dsa_attention_backend(forward_batch: ForwardBatch) -> bool:
     attn_backend = get_attn_backend()
-    server_args = get_global_server_args()
+    server_args = get_server_args()
     prefill_backend, decode_backend = server_args.get_attention_backends()
     prefill_backend = (
         getattr(attn_backend, "prefill_attention_backend_str", None) or prefill_backend
@@ -394,7 +394,7 @@ class Indexer(MultiPlatformOp):
         if _is_cuda:
             self.sm_count = deep_gemm.get_num_sms()
             self.half_device_sm_count = ceil_align(self.sm_count // 2, 8)
-            pp_size = get_global_server_args().pp_size
+            pp_size = get_server_args().pp_size
             self.logits_with_pp_recv = pp_size > 1 and not get_pp_group().is_last_rank
         else:
             self.logits_with_pp_recv = False
@@ -446,7 +446,7 @@ class Indexer(MultiPlatformOp):
             base=rope_theta,  # type: ignore
             rope_scaling=rope_scaling,
             is_neox_style=is_neox_style,
-            device=get_global_server_args().device,
+            device=get_server_args().device,
         )
         self.block_size = block_size
         self.scale_fmt = scale_fmt
@@ -1032,7 +1032,7 @@ class Indexer(MultiPlatformOp):
         total_mem = torch.cuda.get_device_properties(device_index).total_memory
 
         total_mem_budget = int(total_mem * self._MQA_LOGITS_TOTAL_MEM_FRACTION)
-        mem_fraction_static = get_global_server_args().mem_fraction_static
+        mem_fraction_static = get_server_args().mem_fraction_static
         if mem_fraction_static is None:
             static_budget = total_mem_budget
         else:

python/sglang/srt/layers/attention/dsa/utils.py

@@ -15,8 +15,7 @@ from sglang.srt.model_executor.runner_backend_utils.breakable_cuda_graph import
 from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph import (
     is_in_tc_piecewise_cuda_graph,
 )
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import get_bool_env_var, is_cuda, is_hip
 from sglang.srt.utils.common import ceil_align, ceil_div
 
@@ -69,20 +68,20 @@ def compute_dsa_seqlens(original_seq_lens, dsa_index_topk: int):
 
 
 def is_dsa_enable_prefill_cp():
-    return get_global_server_args().enable_dsa_prefill_context_parallel
+    return get_server_args().enable_dsa_prefill_context_parallel
 
 
 def is_dsa_prefill_cp_in_seq_split():
     return (
         is_dsa_enable_prefill_cp()
-        and get_global_server_args().dsa_prefill_cp_mode == "in-seq-split"
+        and get_server_args().dsa_prefill_cp_mode == "in-seq-split"
     )
 
 
 def is_dsa_prefill_cp_round_robin_split():
     return (
         is_dsa_enable_prefill_cp()
-        and get_global_server_args().dsa_prefill_cp_mode == "round-robin-split"
+        and get_server_args().dsa_prefill_cp_mode == "round-robin-split"
     )
 
 

python/sglang/srt/layers/attention/dsv4/indexer.py

@@ -861,9 +861,9 @@ class C4Indexer(nn.Module):
         self.rotary_emb = rotary_emb
         self.freqs_cis = freqs_cis
         self.weight_scale: float = self.softmax_scale * self.n_heads**-0.5
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
-        self.use_fp4_indexer = get_global_server_args().enable_deepseek_v4_fp4_indexer
+        self.use_fp4_indexer = get_server_args().enable_deepseek_v4_fp4_indexer
         self.alt_streams = alt_streams
 
     def compute_q(

python/sglang/srt/layers/attention/flashattention_backend.py

@@ -28,7 +28,7 @@ from sglang.srt.layers.utils.cp_utils import (
 from sglang.srt.mem_cache.memory_pool import KVWriteLoc
 from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.speculative.spec_info import SpecInput, SpeculativeAlgorithm
 from sglang.srt.utils import get_compiler_backend
 
@@ -1342,7 +1342,7 @@ class FlashAttentionBackend(AttentionBackend):
             ):
                 # Do multi-head attention with chunked prefix cache
                 if forward_batch.attn_attend_prefix_cache:
-                    assert not get_global_server_args().disable_chunked_prefix_cache
+                    assert not get_server_args().disable_chunked_prefix_cache
                     # MHA for chunked prefix kv cache when running model with MLA
                     assert forward_batch.prefix_chunk_idx is not None
                     assert forward_batch.prefix_chunk_cu_seq_lens is not None

python/sglang/srt/layers/attention/flashinfer_mla_backend.py

@@ -34,8 +34,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMo
 from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph import (
     is_in_tc_piecewise_cuda_graph,
 )
-from sglang.srt.runtime_context import get_buffer
+from sglang.srt.runtime_context import get_buffer, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.spec_info import SpecInput
 from sglang.srt.speculative.spec_utils import (
     draft_kv_indices_buffer_width,
@@ -226,9 +225,9 @@ class FlashInferMLAAttnBackend(AttentionBackend):
         self.token_to_kv_pool = model_runner.token_to_kv_pool
         self.enable_chunk_kv = (
             not skip_prefill
-            and get_global_server_args().disaggregation_mode != "decode"
+            and get_server_args().disaggregation_mode != "decode"
-            and not get_global_server_args().disable_chunked_prefix_cache
+            and not get_server_args().disable_chunked_prefix_cache
-            and not get_global_server_args().flashinfer_mla_disable_ragged
+            and not get_server_args().flashinfer_mla_disable_ragged
         )
         self.page_size = model_runner.page_size
 
@@ -404,7 +403,7 @@ class FlashInferMLAAttnBackend(AttentionBackend):
             prefix_lens = forward_batch.extend_prefix_lens
             extend_no_prefix = not any(forward_batch.extend_prefix_lens_cpu)
             use_ragged = (
-                not get_global_server_args().flashinfer_mla_disable_ragged
+                not get_server_args().flashinfer_mla_disable_ragged
                 and extend_no_prefix
                 # Piecewise cuda graph should use paged prefill to be compatible with prefix cache
                 and not is_in_tc_piecewise_cuda_graph()

python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py

@@ -19,7 +19,7 @@ from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.mem_cache.memory_pool import HybridReqToTokenPool
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.model_executor.model_runner import ModelRunner
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.speculative.eagle_info import EagleDraftInput, EagleVerifyInput
 from sglang.srt.speculative.spec_info import SpecInput
 
@@ -246,7 +246,7 @@ class MambaAttnBackendBase(AttentionBackend):
         lens_to_track = (
             forward_batch.mamba_track_seqlens - forward_batch.extend_prefix_lens
         )
-        mamba_cache_chunk_size = get_global_server_args().mamba_cache_chunk_size
+        mamba_cache_chunk_size = get_server_args().mamba_cache_chunk_size
         aligned_len = (lens_to_track // mamba_cache_chunk_size) * mamba_cache_chunk_size
         start_indices = query_start_loc[:-1] + aligned_len - conv_state_len
         start_indices = start_indices[forward_batch.mamba_track_mask]
@@ -265,7 +265,7 @@ class MambaAttnBackendBase(AttentionBackend):
         """src/dst indices to track SSM states for prefix caching: aligned seqs
         cache last_recurrent_state, unaligned cache intermediate `h` at the last
         chunk boundary."""
-        mamba_cache_chunk_size = get_global_server_args().mamba_cache_chunk_size
+        mamba_cache_chunk_size = get_server_args().mamba_cache_chunk_size
         # CPU to avoid kernel launches for the masking ops
         mamba_track_mask = forward_batch.mamba_track_mask.cpu()
         extend_seq_lens = forward_batch.extend_seq_lens.cpu()
@@ -336,7 +336,7 @@ class MambaAttnBackendBase(AttentionBackend):
         """Per-row (length bs) bool flush mask = the radix track's seq_lens_cpu %
         mamba_track_interval == 0, so force-flush and snapshot fire on the same
         steps (no off-by-one)."""
-        interval = get_global_server_args().mamba_track_interval
+        interval = get_server_args().mamba_track_interval
         if seq_lens_cpu is None:
             # Should not happen for the supported config; stay safe and never flush.
             return torch.zeros((bs,), dtype=torch.bool)
@@ -748,8 +748,7 @@ class Mamba2AttnBackend(MambaAttnBackendBase):
         # Page-major stores state strided; only the stride-aware Triton causal-conv
         # reads it (CUDA causal_conv1d garbles it). A model may also force Triton.
         use_triton_causal_conv = (
-            use_triton_causal_conv
+            use_triton_causal_conv or get_server_args().enable_page_major_kv_layout
-            or get_global_server_args().enable_page_major_kv_layout
         )
         layer_cache = self.req_to_token_pool.mamba2_layer_cache(layer_id)
         mixer_out, intermediate_states = mixer.forward(

python/sglang/srt/layers/attention/trtllm_mla_backend.py

@@ -38,8 +38,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMo
 from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph import (
     is_in_tc_piecewise_cuda_graph,
 )
-from sglang.srt.runtime_context import get_buffer, get_parallel
+from sglang.srt.runtime_context import get_buffer, get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import is_flashinfer_available, is_float4_e2m1fn_x2
 
 if is_flashinfer_available():
@@ -199,7 +198,7 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
         self.forward_decode_metadata: Union[TRTLLMMLADecodeMetadata, None] = None
 
         self.disable_chunked_prefix_cache = (
-            get_global_server_args().disable_chunked_prefix_cache
+            get_server_args().disable_chunked_prefix_cache
         )
 
         self.num_draft_tokens = model_runner.server_args.speculative_num_draft_tokens

python/sglang/srt/layers/attention/vision.py

@@ -77,7 +77,7 @@ from sglang.srt.layers.linear import (
 )
 from sglang.srt.layers.quantization import QuantizationConfig
 from sglang.srt.layers.rotary_embedding import apply_rotary_pos_emb
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import add_prefix, get_bool_env_var
 
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
@@ -964,10 +964,7 @@ class VisionAttention(nn.Module):
         # Select attention backend via a unified method
         _passed_backend = qkv_backend
         qkv_backend = self._determine_attention_backend(_passed_backend)
-        if (
+        if get_server_args().mm_attention_backend is None and _passed_backend is None:
-            get_global_server_args().mm_attention_backend is None
-            and _passed_backend is None
-        ):
             print_info_once(f"Multimodal attention backend not set. Use {qkv_backend}.")
         print_info_once(f"Using {qkv_backend} as multimodal attention backend.")
 
@@ -1047,7 +1044,7 @@ class VisionAttention(nn.Module):
                 weight_dtype=torch.float32,
                 cast_x_before_out_mul=True,
             )
-            if get_global_server_args().rl_on_policy_target is not None
+            if get_server_args().rl_on_policy_target is not None
             else {}
         )
         q_norm = RMSNorm(
@@ -1075,7 +1072,7 @@ class VisionAttention(nn.Module):
         - CUDA (other): "triton_attn"
         - Non-CUDA: "sdpa"
         """
-        override_backend = get_global_server_args().mm_attention_backend
+        override_backend = get_server_args().mm_attention_backend
         if override_backend is not None:
             backend = override_backend
         elif passed_backend is not None:
@@ -1179,7 +1176,7 @@ class VisionAttention(nn.Module):
             x = x.unsqueeze(0)
         assert x.dim() == 3, x.shape
         if (
-            get_global_server_args().rl_on_policy_target is not None
+            get_server_args().rl_on_policy_target is not None
             and position_embeddings is not None
         ):
             assert isinstance(position_embeddings, tuple), (

python/sglang/srt/layers/attention/xpu_backend.py

@@ -12,10 +12,10 @@ from sglang.srt.layers.attention.flashattention_backend import (
     merge_state_v2_wrapper,
     prepare_swa_spec_page_table_triton,
 )
-from sglang.srt.managers.schedule_batch import get_global_server_args
 from sglang.srt.mem_cache.memory_pool import KVWriteLoc
 from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
+from sglang.srt.runtime_context import get_server_args
 
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
@@ -638,7 +638,7 @@ class XPUAttentionBackend(AttentionBackend):
             ):
                 # Do multi-head attention with chunked prefix cache
                 if forward_batch.attn_attend_prefix_cache:
-                    assert not get_global_server_args().disable_chunked_prefix_cache
+                    assert not get_server_args().disable_chunked_prefix_cache
                     # MHA for chunked prefix kv cache when running model with MLA
                     assert forward_batch.prefix_chunk_idx is not None
                     assert forward_batch.prefix_chunk_cu_seq_lens is not None

python/sglang/srt/layers/communicator.py

@@ -72,8 +72,7 @@ from sglang.srt.model_executor.cuda_graph_config import (
     check_cuda_graph_backend,
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.runtime_context import get_forward, get_parallel
+from sglang.srt.runtime_context import get_forward, get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.utils import (
     get_bool_env_var,
@@ -171,7 +170,7 @@ def apply_flashinfer_allreduce_fusion(batch_size: int):
         and batch_size > 0
         and batch_size <= FUSE_ALLREDUCE_MAX_BATCH_SIZE
         and not is_dp_attention_enabled()
-        and get_global_server_args().flashinfer_allreduce_fusion_backend is not None
+        and get_server_args().flashinfer_allreduce_fusion_backend is not None
         and not is_flashinfer_allreduce_unavailable()
     )
 
@@ -187,7 +186,7 @@ def apply_aiter_all_reduce_fusion(input_tensor: torch.Tensor):
         and total_bytes <= 8 * 1024 * 8192
         and get_parallel().tp_size != 6
         and not is_dp_attention_enabled()
-        and get_global_server_args().enable_aiter_allreduce_fusion
+        and get_server_args().enable_aiter_allreduce_fusion
     )
 
 
@@ -266,7 +265,7 @@ class AttnTpContext:
     def init_context(self, q_lora_rank, is_dsa):
         self.is_dsa = is_dsa
         self.allow_input_scattered = (
-            get_global_server_args().enable_attn_tp_input_scattered
+            get_server_args().enable_attn_tp_input_scattered
             and (_is_cuda or _is_npu)
             and q_lora_rank is not None
             and not is_dsa
@@ -275,9 +274,9 @@ class AttnTpContext:
             and get_moe_a2a_backend().is_none()
             and not enable_moe_dense_fully_dp()
             and not check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE)
-            and get_global_server_args().speculative_algorithm != "EAGLE3"
+            and get_server_args().speculative_algorithm != "EAGLE3"
         )
-        if get_global_server_args().enable_attn_tp_input_scattered:
+        if get_server_args().enable_attn_tp_input_scattered:
             if not self.allow_input_scattered:
                 logging.info(
                     "attn_tp_input_scattered is not enabled while other conditions are not met"
@@ -407,7 +406,7 @@ class LayerScatterModes:
             not context.is_layer_sparse
             and context.is_next_layer_sparse
             and enable_moe_dense_fully_dp()
-            and get_global_server_args().enable_two_batch_overlap
+            and get_server_args().enable_two_batch_overlap
         )
 
     @classmethod
@@ -434,7 +433,7 @@ class LayerScatterModes:
 
 
 def enable_moe_dense_fully_dp():
-    return get_global_server_args().moe_dense_tp_size == 1
+    return get_server_args().moe_dense_tp_size == 1
 
 
 class LayerCommunicator:
@@ -463,7 +462,7 @@ class LayerCommunicator:
         )
         self._post_init_communicate()
         self._speculative_algo = SpeculativeAlgorithm.from_string(
-            get_global_server_args().speculative_algorithm
+            get_server_args().speculative_algorithm
         )
 
     def _post_init_communicate(self):
@@ -811,7 +810,7 @@ class LayerCommunicator:
                     and get_parallel().tp_size != 6
                     and not is_dp_attention_enabled()
                     and get_moe_a2a_backend().is_none()
-                    and get_global_server_args().enable_aiter_allreduce_fusion
+                    and get_server_args().enable_aiter_allreduce_fusion
                 )
             )
             and (not self.is_last_layer)
@@ -1116,7 +1115,7 @@ class CommunicateWithAllReduceAndLayerNormFn:
             if not handled:
                 quantize_communications = (
                     not forward_batch.forward_mode.is_decode_or_idle()
-                    and get_global_server_args().enable_quant_communications
+                    and get_server_args().enable_quant_communications
                 )
                 if quantize_communications:
                     hidden_states = attention_tensor_model_parallel_quant_all_reduce(

python/sglang/srt/layers/cp/base.py

@@ -195,9 +195,9 @@ class ContextParallelStrategy(ABC):
 
 
 def _is_dsa_active() -> bool:
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
-    sa = get_global_server_args()
+    sa = get_server_args()
     return bool(
         getattr(sa, "enable_prefill_cp", False)
         and getattr(sa, "_is_dsa_model_arch", False)
@@ -247,10 +247,10 @@ def get_cp_strategy() -> Optional[ContextParallelStrategy]:
     global _STRATEGY
 
     if _STRATEGY is None:
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
         try:
-            server_args = get_global_server_args()
+            server_args = get_server_args()
         except ValueError:
             return None
         if server_args is not None and getattr(server_args, "enable_prefill_cp", False):

python/sglang/srt/layers/cp/zigzag.py

@@ -204,10 +204,10 @@ class ZigzagCPStrategy(ContextParallelStrategy):
             actual_seq_q_prev_list.append(block_sizes[cp_rank])
             actual_seq_q_next_list.append(block_sizes[cp_segment_num - cp_rank - 1])
 
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
         try:
-            device = torch.device(get_global_server_args().device)
+            device = torch.device(get_server_args().device)
         except Exception:
             device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         cu_prev = [0] + list(accumulate(actual_seq_q_prev_list))

python/sglang/srt/layers/dcp/__init__.py

@@ -46,7 +46,7 @@ from sglang.srt.layers.dcp.layout import (
 from sglang.srt.layers.dcp.metadata import DecodeContextParallelMetadata
 
 # NOTE: planner.py is intentionally NOT imported here. It depends on server_args
-# (get_global_server_args), whereas this package-init executes at module-load time
+# (get_server_args), whereas this package-init executes at module-load time
 # for every eager importer of the DCP primitives — triton_backend,
 # mem_cache.memory_pool, mem_cache.triton_ops.mla_buffer, mem_cache.kv_cache_builder,
 # the FlashInfer-MLA / FlashMLA backends, and the deepseek forward methods. Keeping

python/sglang/srt/layers/dcp/planner.py

@@ -27,8 +27,7 @@ from sglang.srt.layers.dcp.kernels import (
 )
 from sglang.srt.layers.dcp.layout import update_local_kv_lens_for_dcp
 from sglang.srt.layers.dcp.metadata import DecodeContextParallelMetadata
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 
 
 def prepare_decode_context_parallel_metadata(
@@ -54,12 +53,12 @@ def prepare_decode_context_parallel_metadata(
     extend_prefix_starts = torch.zeros(
         len(seq_lens),
         dtype=torch.int32,
-        device=get_global_server_args().device,
+        device=get_server_args().device,
     )
     extend_cu_prefix_lens = torch.zeros(
         len(seq_lens) + 1,
         dtype=torch.int32,
-        device=get_global_server_args().device,
+        device=get_server_args().device,
     )
     extend_cu_prefix_lens[1:] = torch.cumsum(extend_prefix_lens, dim=0)
     extend_cu_prefix_lens = extend_cu_prefix_lens[:-1]
@@ -68,7 +67,7 @@ def prepare_decode_context_parallel_metadata(
     dcp_prefix_kv_indices = torch.empty(
         sum(extend_prefix_lens_cpu),
         dtype=torch.int32,
-        device=get_global_server_args().device,
+        device=get_server_args().device,
     )
     create_chunked_prefix_cache_kv_indices_fn[(len(seq_lens),)](
         req_to_token,
@@ -82,20 +81,20 @@ def prepare_decode_context_parallel_metadata(
     dcp_kv_indptr = torch.zeros(
         len(seq_lens) + 1,
         dtype=torch.int32,
-        device=get_global_server_args().device,
+        device=get_server_args().device,
     )
     dcp_kv_indptr[1:] = seq_lens.cumsum(dim=0)
     dcp_kv_indptr = dcp_kv_indptr[: (len(seq_lens) + 1)]
     dcp_kv_indices = torch.zeros(
         seq_lens_sum,
         dtype=torch.int32,
-        device=get_global_server_args().device,
+        device=get_server_args().device,
     )
 
     extend_cu_lens = torch.zeros(
         len(seq_lens) + 1,
         dtype=torch.int32,
-        device=get_global_server_args().device,
+        device=get_server_args().device,
     )
     extend_cu_lens[1:] = torch.cumsum(extend_seq_lens, dim=0)
     extend_cu_lens = extend_cu_lens[:-1]

python/sglang/srt/layers/dp_attention.py

@@ -778,9 +778,9 @@ def get_moe_cp_size() -> int:
 
 def is_enable_moe_cp_allgather() -> bool:
     """True when moe_dp_size < attn_cp_size, requiring allgather across CP ranks before MoE."""
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
-    sa = get_global_server_args()
+    sa = get_server_args()
     return sa.attn_cp_size > sa.moe_dp_size
 
 

python/sglang/srt/layers/flashinfer_comm_fusion.py

@@ -13,8 +13,7 @@ from sglang.srt.distributed import (
     get_tp_group,
 )
 from sglang.srt.distributed.parallel_state import in_the_same_node_as
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     ceil_align,
     get_cuda_driver_bindings,
@@ -673,7 +672,7 @@ def ensure_workspace_initialized(
     token_num = token_num or max_token_num
     group_key = (device_group, cpu_group)
     effective_dtype = dtype or torch.bfloat16
-    server_args = get_global_server_args()
+    server_args = get_server_args()
     backend = resolve_flashinfer_allreduce_fusion_backend(server_args)
     if backend is None:
         return False

python/sglang/srt/layers/layernorm.py

@@ -31,8 +31,7 @@ from sglang.srt.model_executor.cuda_graph_config import (
     Phase,
     check_cuda_graph_backend,
 )
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
@@ -193,7 +192,7 @@ def _forward_with_allreduce_fusion(
                     return fused_result
 
             # For AITER route, preserve correctness when fused path is unavailable.
-            if _use_aiter and get_global_server_args().enable_aiter_allreduce_fusion:
+            if _use_aiter and get_server_args().enable_aiter_allreduce_fusion:
                 x = tensor_model_parallel_all_reduce(x)
                 return norm_module.forward(x, residual, None)
 
@@ -271,7 +270,7 @@ class RMSNorm(MultiPlatformOp):
             if (
                 residual is not None
                 or self.cast_x_before_out_mul
-                or get_global_server_args().rl_on_policy_target == "fsdp"
+                or get_server_args().rl_on_policy_target == "fsdp"
             ):
                 return self.forward_native(x, residual, post_residual_addition)
             return rms_norm_batch_invariant(
@@ -371,7 +370,7 @@ class RMSNorm(MultiPlatformOp):
             if (
                 residual is not None
                 or self.cast_x_before_out_mul
-                or get_global_server_args().rl_on_policy_target == "fsdp"
+                or get_server_args().rl_on_policy_target == "fsdp"
                 or (self._fused_pad_kernel is not None and self.x_pad_to_multiple > 0)
             ):
                 return self.forward_native(x, residual, post_residual_addition)
@@ -432,7 +431,7 @@ class RMSNorm(MultiPlatformOp):
             if (
                 residual is not None
                 or self.cast_x_before_out_mul
-                or get_global_server_args().rl_on_policy_target == "fsdp"
+                or get_server_args().rl_on_policy_target == "fsdp"
             ):
                 return self.forward_native(x, residual, post_residual_addition)
             return rms_norm_batch_invariant(
@@ -559,10 +558,7 @@ class RMSNorm(MultiPlatformOp):
         if self.variance_size_override is not None:
             return self.forward_native(x, residual, post_residual_addition)
         if is_batch_invariant_mode_enabled():
-            if (
+            if residual is not None or get_server_args().rl_on_policy_target == "fsdp":
-                residual is not None
-                or get_global_server_args().rl_on_policy_target == "fsdp"
-            ):
                 return self.forward_native(x, residual, post_residual_addition)
             return rms_norm_batch_invariant(
                 x,

python/sglang/srt/layers/linear.py

@@ -37,8 +37,7 @@ from sglang.srt.layers.parameter import (
     _ColumnvLLMParameter,
 )
 from sglang.srt.layers.utils import pad_or_narrow_weight
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import get_bool_env_var, is_cpu, is_hip, is_npu, set_weight_attrs
 
 if TYPE_CHECKING:
@@ -1545,7 +1544,7 @@ class RowParallelLinear(LinearBase):
                 quantize_communications = (
                     (
                         not forward_batch.forward_mode.is_decode_or_idle()
-                        and get_global_server_args().enable_quant_communications
+                        and get_server_args().enable_quant_communications
                     )
                     if forward_batch is not None
                     else False

python/sglang/srt/layers/logits_processor.py

@@ -48,7 +48,6 @@ from sglang.srt.model_executor.forward_batch_info import (
     ForwardMode,
 )
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils.common import (
     is_cpu,
     is_npu,
@@ -336,7 +335,7 @@ class LogitsProcessor(nn.Module):
         self.vocab_size = config.vocab_size
         self.logit_scale = logit_scale
         self.use_attn_tp_group = get_server_args().enable_dp_lm_head
-        self.use_fp32_lm_head = get_global_server_args().enable_fp32_lm_head
+        self.use_fp32_lm_head = get_server_args().enable_fp32_lm_head
         if self.use_attn_tp_group:
             self.attn_tp_size = get_parallel().attn_tp_size
             self.do_tensor_parallel_all_gather = (
@@ -360,7 +359,7 @@ class LogitsProcessor(nn.Module):
             self.final_logit_softcapping = None
 
         self.return_full_logits = return_full_logits
-        self.enable_mis = get_global_server_args().enable_mis
+        self.enable_mis = get_server_args().enable_mis
 
         self._logits_gatherer = triton_symm_mem_ag.MultimemAllGatherer(
             max_tokens=triton_symm_mem_ag.recommended_max_tokens(
@@ -970,7 +969,7 @@ class LogitsProcessor(nn.Module):
                     None,  # bias
                     True,  # is_vnni
                 )
-            elif get_global_server_args().rl_on_policy_target is not None:
+            elif get_server_args().rl_on_policy_target is not None:
                 # Due to tie-weight, we may not be able to change lm_head's weight dtype
                 logits = torch.matmul(
                     hidden_states.bfloat16(), lm_head.weight.T.bfloat16()

python/sglang/srt/layers/mhc.py

@@ -521,10 +521,10 @@ def prewarm_mhc_pre(
     the TileLang/DeepGEMM on-disk JIT cache, so this cost is paid only on a cold
     cache; later server runs hit the cache. Driven once per process from load_weights.
     """
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
     hc_mult, hidden_size = residual.shape[-2], residual.shape[-1]
-    max_num_tokens = get_global_server_args().chunked_prefill_size
+    max_num_tokens = get_server_args().chunked_prefill_size
     buckets = get_mhc_pre_token_count_representatives(
         max_num_tokens, hc_mult * hidden_size
     )

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -66,8 +66,7 @@ from sglang.srt.model_executor.runner_backend_utils.tc_piecewise_cuda_graph impo
     is_in_tc_piecewise_cuda_graph,
 )
 from sglang.srt.model_loader.weight_utils import narrow_padded_param_and_loaded_weight
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
@@ -276,7 +275,7 @@ class FusedMoE(torch.nn.Module):
         )
 
         self.quant_method: Optional[FusedMoEMethodBase] = None
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         kt_config = create_kt_config_from_server_args(server_args, layer_id)
         if kt_config is not None:
             if quant_config is not None:

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

@@ -44,11 +44,10 @@ class HashTopK(nn.Module):
     ):
         super().__init__()
         self.layer_id = layer_id
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
         self.enable_deepep_waterfill = (
-            num_fused_shared_experts > 0
+            num_fused_shared_experts > 0 and get_server_args().enable_deepep_waterfill
-            and get_global_server_args().enable_deepep_waterfill
         )
         self.deepep_waterfill_balancer = None
 

python/sglang/srt/layers/moe/moe_runner/flashinfer_cutedsl.py

@@ -244,14 +244,14 @@ def ensure_cutedsl_wrapper(layer: torch.nn.Module) -> None:
             "Install with: pip install flashinfer"
         ) from e
 
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
     assert layer.intermediate_size_per_partition > 0, (
         f"CuteDSL MoE: intermediate_size_per_partition must be > 0, "
         f"got {layer.intermediate_size_per_partition}. Check EP/TP configuration."
     )
 
-    server_args = get_global_server_args()
+    server_args = get_server_args()
     # CuteDSL wrapper preallocates CG buffers used by any captured graph
     # that routes through this MoE — decode and prefill alike.
     use_cuda_graph = not cuda_graph_fully_disabled()

python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe.py

@@ -17,7 +17,7 @@ from sglang.srt.batch_invariant_ops import is_batch_invariant_mode_enabled
 from sglang.srt.environ import envs
 from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
 from sglang.srt.layers.moe.utils import get_moe_padding_size
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
@@ -488,7 +488,7 @@ def _fused_moe_kernel_sequence(
         out_hidden_states = torch.empty_like(hidden_states)
 
     use_fused_moe_sum_all_reduce = (
-        get_global_server_args().enable_fused_moe_sum_all_reduce
+        get_server_args().enable_fused_moe_sum_all_reduce
         and (not no_combine)
         and (topk > 2)
         and (not use_int8_w8a16)

python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe_triton_config.py

@@ -9,7 +9,7 @@ from typing import Any, Dict, List, Optional, Tuple
 import torch
 import triton
 
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import get_device_name, is_hip
 
 logger = logging.getLogger(__name__)
@@ -69,7 +69,7 @@ def get_moe_configs(
     kernel on a given batch size bs, the closest batch size in the grid should
     be picked and the associated configuration chosen to invoke the kernel.
     """
-    if get_global_server_args().enable_deterministic_inference:
+    if get_server_args().enable_deterministic_inference:
         logger.warning(
             "Deterministic inference is enabled, using default MoE kernel config."
         )
@@ -170,7 +170,7 @@ def get_default_config(
     is_marlin: bool,
     block_shape: Optional[List[int]] = None,
 ) -> Dict[str, int]:
-    if get_global_server_args().enable_deterministic_inference:
+    if get_server_args().enable_deterministic_inference:
         config = {
             "BLOCK_SIZE_M": 64,
             "BLOCK_SIZE_N": 64,

python/sglang/srt/layers/moe/token_dispatcher/flashinfer.py

@@ -23,7 +23,7 @@ from sglang.srt.layers.moe.token_dispatcher.flashinfer_utils import (
 )
 from sglang.srt.layers.moe.topk import StandardTopKOutput, TopKOutput
 from sglang.srt.layers.moe.utils import get_moe_runner_backend
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.utils import get_int_env_var
 
@@ -119,7 +119,7 @@ class FlashinferDispatcher(BaseDispatcher):
         # max_running_requests is not yet resolved at model-construction time,
         # so we use 4096 as a floor to cover decode batches and _dummy_run
         # (which warms up at batch_size = req_to_token_pool.size).
-        cps = get_global_server_args().chunked_prefill_size
+        cps = get_server_args().chunked_prefill_size
         default_max_tokens = max(cps if cps and cps > 0 else 4096, 4096)
         self.max_num_tokens = get_int_env_var(
             "SGLANG_FLASHINFER_NUM_MAX_DISPATCH_TOKENS_PER_RANK",
@@ -128,7 +128,7 @@ class FlashinferDispatcher(BaseDispatcher):
 
         # Calculate workspace size. For eagle mode, use the larger workspace size since nextn layer will be unquantized.
         speculative_algo = SpeculativeAlgorithm.from_string(
-            get_global_server_args().speculative_algorithm
+            get_server_args().speculative_algorithm
         )
         if MOE_NVFP4_DISPATCH and not speculative_algo.is_eagle():
             total_dispatch_payload_size_per_token = (

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

@@ -395,11 +395,10 @@ class TopK(MultiPlatformOp):
             assert num_expert_group is not None and topk_group is not None
 
         self.layer_id = layer_id
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
         self.enable_deepep_waterfill = (
-            num_fused_shared_experts > 0
+            num_fused_shared_experts > 0 and get_server_args().enable_deepep_waterfill
-            and get_global_server_args().enable_deepep_waterfill
         )
 
         self.deepep_waterfill_balancer = None
@@ -475,9 +474,9 @@ class TopK(MultiPlatformOp):
         # ===== TO BE REFACTORED ====
         elif get_moe_runner_backend().is_experimental_sgl_trtllm():
             try:
-                from sglang.srt.server_args import get_global_server_args
+                from sglang.srt.runtime_context import get_server_args
 
-                use_standard_for_lora = bool(get_global_server_args().enable_lora)
+                use_standard_for_lora = bool(get_server_args().enable_lora)
             except ValueError:
                 use_standard_for_lora = False
             output_format = (
@@ -1256,10 +1255,10 @@ def _eplb_remap_enabled() -> bool:
     # initial expert placement is non-trivial, or there are redundant physical
     # experts. Otherwise the map is identity and the remap must be skipped (it is
     # both unnecessary and not well-defined over the padded region of topk_ids).
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
     try:
-        server_args = get_global_server_args()
+        server_args = get_server_args()
     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/layers/moe/utils.py

@@ -20,7 +20,7 @@ _is_npu = is_npu()
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
 
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 logger = logging.getLogger(__name__)
 
@@ -206,7 +206,7 @@ def get_deepep_output_dtype(self) -> DeepEPOutputDtype:
     """
 
     # 0. Parse server argument.
-    server_args = get_global_server_args()
+    server_args = get_server_args()
     if server_args and server_args.deepep_dispatcher_output_dtype != "auto":
         return DeepEPOutputDtype(server_args.deepep_dispatcher_output_dtype)
 

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -33,7 +33,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     w8a8_block_fp8_matmul_deepgemm,
     w8a8_block_fp8_matmul_triton,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     ceil_align,
     ceil_div,
@@ -1691,8 +1691,7 @@ def apply_fp8_linear(
         if (
             input_scale is not None
             and input_scale.numel() == 1
-            and get_global_server_args().cuda_graph_config.prefill.tc_compiler
+            and get_server_args().cuda_graph_config.prefill.tc_compiler == "inductor"
-            == "inductor"
         ):
             qinput = (
                 (input_2d * input_scale.reciprocal())

python/sglang/srt/layers/quantization/mxfp4.py

@@ -48,7 +48,7 @@ from sglang.srt.layers.quantization.base_config import (
     QuantizeMethodBase,
 )
 from sglang.srt.layers.quantization.utils import is_layer_skipped
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     is_cpu,
@@ -334,7 +334,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
         self.use_flashinfer = get_moe_runner_backend().is_flashinfer_mxfp4()
         self.use_marlin = get_moe_runner_backend().is_marlin()
         self.flashinfer_mxfp4_moe_precision = (
-            get_global_server_args().flashinfer_mxfp4_moe_precision
+            get_server_args().flashinfer_mxfp4_moe_precision
         )
         # When `flashinfer_mxfp4` is enabled, dispatch to one of two FlashInfer
         # entry points depending on the GPU:

python/sglang/srt/layers/quantization/mxfp4_flashinfer_trtllm_moe.py

@@ -15,7 +15,7 @@ from sglang.srt.distributed.device_communicators.pynccl_allocator import (
 )
 from sglang.srt.layers.dp_attention import is_allocation_symmetric
 from sglang.srt.layers.moe.utils import RoutingMethodType
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     is_flashinfer_available,
     log_info_on_rank0,
@@ -128,7 +128,7 @@ class Mxfp4FlashinferTrtllmMoEMethod:
         self._fp8 = fp8_method
         self.prefix = prefix
         self.flashinfer_mxfp4_moe_precision = (
-            get_global_server_args().flashinfer_mxfp4_moe_precision
+            get_server_args().flashinfer_mxfp4_moe_precision
         )
 
     def create_moe_runner(self, layer, moe_runner_config):

python/sglang/srt/layers/rotary_embedding/base.py

@@ -11,7 +11,7 @@ from sglang.srt.environ import envs
 from sglang.srt.layers.rotary_embedding.utils import apply_rotary_emb
 from sglang.srt.layers.utils import MultiPlatformOp
 from sglang.srt.platforms import current_platform
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
@@ -127,7 +127,7 @@ class RotaryEmbedding(MultiPlatformOp):
         self._apply_rotary_emb_wrapped = apply_rotary_emb
 
         # XXX (MUSA): Implement sgl_kernel.rotary_embedding support for MUSA backend
-        if get_global_server_args().rl_on_policy_target is not None or _is_musa:
+        if get_server_args().rl_on_policy_target is not None or _is_musa:
             self._forward_method = self.forward_native
             self._apply_rotary_emb_wrapped = torch.compile(
                 dynamic=True,
@@ -151,7 +151,7 @@ class RotaryEmbedding(MultiPlatformOp):
         # create the cache on GPU for faster initialization. This may cause
         # a slight numerical difference between the HF implementation and ours.
         init_device = (
-            "cpu" if get_global_server_args().rl_on_policy_target is not None else None
+            "cpu" if get_server_args().rl_on_policy_target is not None else None
         )
         inv_freq = 1.0 / (
             base
@@ -162,7 +162,7 @@ class RotaryEmbedding(MultiPlatformOp):
                 / self.rotary_dim
             )
         )
-        if get_global_server_args().rl_on_policy_target is not None:
+        if get_server_args().rl_on_policy_target is not None:
             inv_freq = inv_freq.cuda()
         return inv_freq
 

python/sglang/srt/layers/rotary_embedding/mrope.py

@@ -18,7 +18,7 @@ from sglang.srt.layers.rotary_embedding.yarn import (
     yarn_get_mscale_simple,
     yarn_linear_ramp_mask,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     is_cuda,
@@ -216,7 +216,7 @@ class MRotaryEmbedding(RotaryEmbedding):
             self.register_buffer("axis_map", axis_map, persistent=False)
         else:
             self.axis_map = None
-        if get_global_server_args().rl_on_policy_target is not None:
+        if get_server_args().rl_on_policy_target is not None:
             self._forward_method = self.forward_native
 
     def get_cos_sin_with_position(self, positions):

python/sglang/srt/layers/sampler.py

@@ -15,7 +15,6 @@ from sglang.srt.layers.utils.logprob import get_token_ids_logprobs, get_top_logp
 from sglang.srt.runtime_context import get_parallel, get_server_args
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.sampling.sampling_params import TOP_K_ALL
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils.async_probe import sanitize_nan_logits
 from sglang.srt.utils.common import (
     get_bool_env_var,
@@ -72,11 +71,9 @@ class Sampler(nn.Module):
         if is_dp_attention_enabled():
             self.tp_sync_group = get_parallel().attn_tp_group.device_group
 
-        self.rl_on_policy_target = get_global_server_args().rl_on_policy_target
+        self.rl_on_policy_target = get_server_args().rl_on_policy_target
         # In RL on-policy mode, deterministic inference is automatically enabled.
-        self.enable_deterministic = (
+        self.enable_deterministic = get_server_args().enable_deterministic_inference
-            get_global_server_args().enable_deterministic_inference
-        )
         # In RL on-policy mode, we use log_softmax to compute logprobs to match the trainer.
         self.use_log_softmax_logprob = self.rl_on_policy_target is not None
         self.use_ascend_backend = get_server_args().sampling_backend == "ascend"
@@ -460,7 +457,7 @@ def register_sampler_backend(backend: str, factory: Callable[[], "Sampler"]) ->
 def create_sampler(backend: Optional[str] = None) -> "Sampler":
     """Create a sampler honoring custom backend registrations."""
 
-    server_args = get_global_server_args()
+    server_args = get_server_args()
     backend = backend or (server_args.sampling_backend if server_args else None)
 
     if backend in _CUSTOM_SAMPLER_FACTORIES:

python/sglang/srt/layers/utils/cp_utils.py

@@ -15,8 +15,7 @@ from sglang.srt.layers.dp_attention import (
 from sglang.srt.layers.moe import get_moe_a2a_backend
 from sglang.srt.mem_cache.memory_pool import KVWriteLoc
 from sglang.srt.model_executor.forward_context import get_token_to_kv_pool
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 
 
 @dataclass
@@ -59,13 +58,13 @@ class ContextParallelMetadata:
 
 
 def is_prefill_context_parallel_enabled():
-    return get_global_server_args().enable_prefill_context_parallel
+    return get_server_args().enable_prefill_context_parallel
 
 
 def is_prefill_cp_in_seq_split():
     return (
         is_prefill_context_parallel_enabled()
-        and get_global_server_args().prefill_cp_mode == "in-seq-split"
+        and get_server_args().prefill_cp_mode == "in-seq-split"
     )
 
 
@@ -85,7 +84,7 @@ def get_cp_padding_align_size() -> int:
 
 
 def is_mla_prefill_cp_enabled() -> bool:
-    sa = get_global_server_args()
+    sa = get_server_args()
     return sa.enable_prefill_context_parallel and sa.use_mla_backend()
 
 

python/sglang/srt/managers/mm_utils.py

@@ -32,7 +32,7 @@ from sglang.srt.managers.schedule_batch import (
 from sglang.srt.mem_cache.multimodal_cache import EmbeddingResult, MultiModalStaticCache
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.multimodal.evs import EVSEmbeddingResult
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import flatten_nested_list, is_npu, print_warning_once
 from sglang.srt.utils.stale_shm_cleanup import make_shm_name
 from sglang.utils import logger
@@ -714,7 +714,7 @@ def _adjust_embedding_length(
             f"tokens from multimodal embeddings."
         )
         if num_mm_tokens_in_input_ids < num_mm_tokens_in_embedding:
-            chunked_prefill_size = get_global_server_args().chunked_prefill_size
+            chunked_prefill_size = get_server_args().chunked_prefill_size
             if chunked_prefill_size != -1:
                 logger.warning(
                     "You may want to avoid this issue by raising `chunked_prefill_size`, or disabling chunked prefill"
@@ -1073,7 +1073,7 @@ def general_mm_embed_routine(
                 for i, seq_len in enumerate(forward_batch.extend_seq_lens_cpu)
                 if forward_batch.mm_inputs[i] is not None
             ]
-            server_args = get_global_server_args()
+            server_args = get_server_args()
             if server_args and server_args.enable_adaptive_dispatch_to_encoder:
                 # Split by precomputed vs non-precomputed so get_embedding_and_mask only sees uniform batches
                 input_embeds, other_info = _embed_mm_inputs_with_split(
@@ -1119,7 +1119,7 @@ def general_mm_embed_routine(
                             feature = getattr(mm_item, "feature", None)
                             if isinstance(feature, torch.Tensor) and feature.is_cuda:
                                 mm_item.feature = feature.to("cpu", non_blocking=True)
-                            if get_global_server_args().language_only:
+                            if get_server_args().language_only:
                                 precomputed_embeddings = getattr(
                                     mm_item, "precomputed_embeddings", None
                                 )
@@ -1756,7 +1756,7 @@ def _get_is_default_transport():
         )
 
         _is_default_tensor_transport = (
-            _determine_tensor_transport_mode(get_global_server_args()) == "default"
+            _determine_tensor_transport_mode(get_server_args()) == "default"
         )
     return _is_default_tensor_transport
 
@@ -1798,7 +1798,7 @@ def wrap_shm_features(obj):
     """
     Scan the object for multimodal tensors and wrap them in SHM pointers.
     """
-    if _get_is_default_transport() or get_global_server_args().skip_tokenizer_init:
+    if _get_is_default_transport() or get_server_args().skip_tokenizer_init:
         return obj
 
     if obj.mm_inputs:
@@ -1859,7 +1859,7 @@ def unwrap_shm_features(obj):
     Restore ShmPointerMMData wrappers back into standard torch.Tensors.
     Handles both single requests and batch requests.
     """
-    if _get_is_default_transport() or get_global_server_args().skip_tokenizer_init:
+    if _get_is_default_transport() or get_server_args().skip_tokenizer_init:
         return obj
     # Handle batch requests
     if isinstance(obj, BaseBatchReq):

python/sglang/srt/managers/schedule_batch.py

@@ -106,10 +106,10 @@ from sglang.srt.observability.req_time_stats import (
     DPControllerReqTimeStats,
     SchedulerReqTimeStats,
 )
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.sampling.sampling_params import SamplingParams
-from sglang.srt.server_args import ServerArgs, get_global_server_args
+from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import flatten_nested_list
 from sglang.srt.utils.cuda_ipc_transport_utils import CudaIpcTensorTransportProxy
 
@@ -1035,7 +1035,7 @@ class Req(ReqDllmMixin):
         """Check if this request is prefill-only (no token generation needed)."""
         # NOTE: when spec is enabled, prefill_only optimizations are disabled
 
-        spec_alg = get_global_server_args().speculative_algorithm
+        spec_alg = get_server_args().speculative_algorithm
         return self.sampling_params.max_new_tokens == 0 and spec_alg is None
 
     @property
@@ -1056,7 +1056,7 @@ class Req(ReqDllmMixin):
     def _cache_commit_len(self) -> int:
         # Report only the prompt prefix so thinking + answer fall into the
         # overallocated range and are reclaimed by release_kv_cache. #22373.
-        if get_global_server_args().strip_thinking_cache and self.reasoning_tokens > 0:
+        if get_server_args().strip_thinking_cache and self.reasoning_tokens > 0:
             return min(self.kv_committed_len, len(self.origin_input_ids))
         return self.kv_committed_len
 
@@ -2205,7 +2205,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 req.already_computed = seq_len
             req.is_retracted = False
 
-            if get_global_server_args().enable_mamba_extra_buffer():
+            if get_server_args().enable_mamba_extra_buffer():
                 track_entry = self._mamba_radix_cache_v2_req_prepare_for_extend(req)
                 mamba_track_mask_cpu.append(track_entry.track_mask)
                 mamba_track_indices_cpu.append(track_entry.track_index)
@@ -2310,7 +2310,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.extend_logprob_start_lens = extend_logprob_start_lens
         self.extend_input_logprob_token_ids = extend_input_logprob_token_ids
 
-        if get_global_server_args().enable_mamba_extra_buffer():
+        if get_server_args().enable_mamba_extra_buffer():
             self.mamba_track_indices = torch.tensor(
                 mamba_track_indices_cpu,
                 dtype=torch.int64,
@@ -2344,7 +2344,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self,
         req: Req,
     ) -> _MambaRadixCacheV2TrackEntry:
-        mamba_cache_chunk_size = get_global_server_args().mamba_cache_chunk_size
+        mamba_cache_chunk_size = get_server_args().mamba_cache_chunk_size
 
         def _force_track_h(i: int) -> int:
             assert i % mamba_cache_chunk_size == 0
@@ -2395,7 +2395,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             # In lazy mode, skip the swap — the second ping-pong slot is not
             # allocated yet; it will be allocated on demand at the track boundary
             # in mamba_lazy_prealloc_at_boundary during prepare_for_decode.
-            if not get_global_server_args().enable_mamba_extra_buffer_lazy():
+            if not get_server_args().enable_mamba_extra_buffer_lazy():
                 req.mamba_next_track_idx = (
                     self.req_to_token_pool.get_mamba_ping_pong_other_idx(
                         req.mamba_next_track_idx
@@ -2736,15 +2736,15 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 self.req_pool_indices_cpu,
             )
 
-        if get_global_server_args().enable_mamba_extra_buffer():
+        if get_server_args().enable_mamba_extra_buffer():
-            mamba_track_interval = get_global_server_args().mamba_track_interval
+            mamba_track_interval = get_server_args().mamba_track_interval
 
             if len(self.reqs) == 0:
                 self.mamba_track_indices = torch.empty(
                     (0,), dtype=torch.int64, device=self.device
                 )
             else:
-                if get_global_server_args().enable_mamba_extra_buffer_lazy():
+                if get_server_args().enable_mamba_extra_buffer_lazy():
                     self.mamba_lazy_prealloc_at_boundary(mamba_track_interval)
                 set_mamba_track_indices_from_reqs(self)
 
@@ -2932,7 +2932,7 @@ 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_global_server_args()
+            server_args = get_server_args()
 
             release_leaf_lock = (
                 envs.SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW.get()

python/sglang/srt/managers/schedule_policy.py

@@ -56,7 +56,8 @@ from sglang.srt.mem_cache.multi_ended_allocator import (
     UnifiedMambaTokenToKVPoolAllocator,
 )
 from sglang.srt.mem_cache.radix_cache import RadixCache, RadixKey, TreeNode
-from sglang.srt.server_args import ServerArgs, get_global_server_args
+from sglang.srt.runtime_context import get_server_args
+from sglang.srt.server_args import ServerArgs
 
 if TYPE_CHECKING:
     from sglang.srt.mem_cache.allocator import BaseTokenToKVPoolAllocator
@@ -185,7 +186,7 @@ class SchedulePolicy:
         if (
             not isinstance(policy, CacheAwarePolicy)
             and self.tree_cache.supports_fast_match_prefix()
-            and get_global_server_args().disaggregation_mode != "decode"
+            and get_server_args().disaggregation_mode != "decode"
         ):
             for r in waiting_queue:
                 match_prefix_for_req(self.tree_cache, r, include_req=True)

python/sglang/srt/managers/scheduler.py

@@ -237,9 +237,9 @@ from sglang.srt.observability.trace import process_tracing_init, trace_set_threa
 from sglang.srt.parser.reasoning_parser import ReasoningParser
 from sglang.srt.platforms import current_platform
 from sglang.srt.plugins import load_plugins
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
-from sglang.srt.server_args import PortArgs, ServerArgs, get_global_server_args
+from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.session.session_controller import SessionController
 from sglang.srt.speculative.dflash_utils import validate_dflash_request
 from sglang.srt.speculative.eagle_utils import get_draft_recurrent_hidden_state_spec
@@ -892,8 +892,8 @@ class Scheduler(
             self.min_free_slots_delayer = MinFreeSlotsDelayer(
                 min_free_slots=min_free_slots
             )
-        if not get_global_server_args().pp_max_micro_batch_size:
+        if not get_server_args().pp_max_micro_batch_size:
-            get_global_server_args().override(
+            get_server_args().override(
                 "scheduler.pp_max_micro_batch_size_default",
                 pp_max_micro_batch_size=max(
                     self.max_running_requests // self.ps.pp_size, 1
@@ -2730,7 +2730,7 @@ class Scheduler(
         return ret
 
     def get_num_allocatable_reqs(self, running_bs):
-        res = get_global_server_args().pp_max_micro_batch_size - running_bs
+        res = get_server_args().pp_max_micro_batch_size - running_bs
         res = min(res, self.req_to_token_pool.available_size())
         return res
 
@@ -3766,7 +3766,7 @@ class Scheduler(
         return success
 
     def get_internal_state(self, recv_req: GetInternalStateReq):
-        ret = dict(vars(get_global_server_args()))  # vars returns a ref to obj.__dict__
+        ret = dict(vars(get_server_args()))  # vars returns a ref to obj.__dict__
         ret["last_gen_throughput"] = self.metrics_reporter.last_gen_throughput
         ret["memory_usage"] = {
             "weight": round(self.tp_worker.model_runner.weight_load_mem_usage, 2),
@@ -3833,11 +3833,10 @@ class Scheduler(
             self.metrics_reporter.spec_total_num_accept_tokens = (
                 self.metrics_reporter.spec_total_num_forward_ct
             ) = 0
-            for k, v in server_args_dict.items():
+            get_server_args().override(source="update_server_args", **server_args_dict)
-                setattr(get_global_server_args(), k, v)
+            logger.info(f"Global server args updated! {get_server_args()=}")
-            logger.info(f"Global server args updated! {get_global_server_args()=}")
 
-        server_args = dict(vars(get_global_server_args()))
+        server_args = dict(vars(get_server_args()))
         # This field is not serializable.
         server_args.pop("model_config", None)
         return SetInternalStateReqOutput(

python/sglang/srt/managers/scheduler_components/batch_result_processor.py

@@ -25,7 +25,7 @@ from sglang.srt.mem_cache.common import (
     maybe_cache_unfinished_req,
     release_kv_cache,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.state_capturer.indexer_topk import get_global_indexer_capturer
 from sglang.srt.state_capturer.routed_experts import get_global_experts_capturer
 
@@ -848,7 +848,7 @@ class SchedulerBatchResultProcessor:
                     prepare_release(req)
                 is_insert = (
                     req.mamba_lazy_is_insert
-                    if get_global_server_args().enable_mamba_extra_buffer_lazy()
+                    if get_server_args().enable_mamba_extra_buffer_lazy()
                     else True
                 )
                 release_kv_cache(req, self.tree_cache, is_insert=is_insert)
@@ -883,7 +883,7 @@ class SchedulerBatchResultProcessor:
         if req.mamba_ping_pong_track_buffer is None:
             return
 
-        lazy = get_global_server_args().enable_mamba_extra_buffer_lazy()
+        lazy = get_server_args().enable_mamba_extra_buffer_lazy()
         at_boundary, track_seqlen = self._mamba_check_track_boundary(
             req, batch, result, i
         )
@@ -915,7 +915,7 @@ class SchedulerBatchResultProcessor:
         For spec decode, the boundary is detected by comparing the
         accepted seq_len range against interval boundaries.
         """
-        interval = get_global_server_args().mamba_track_interval
+        interval = get_server_args().mamba_track_interval
 
         if batch.spec_algorithm.is_none():
             if req.kv_committed_len % interval == 0:

python/sglang/srt/managers/scheduler_components/profiler_manager.py

@@ -18,7 +18,7 @@ import torch
 from sglang.srt.environ import envs
 from sglang.srt.managers.io_struct import ProfileReq, ProfileReqOutput, ProfileReqType
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import is_mps, is_npu
 from sglang.srt.utils.profile_merger import ProfileMerger
 from sglang.srt.utils.profile_utils import ProfileManager
@@ -245,7 +245,7 @@ class SchedulerProfilerManager:
             self.profile_in_progress = True
 
         if "CUDA_PROFILER" in activities:
-            if self.ps.gpu_id == get_global_server_args().base_gpu_id:
+            if self.ps.gpu_id == get_server_args().base_gpu_id:
                 torch.cuda.cudart().cudaProfilerStart()
             self.profile_in_progress = True
 
@@ -355,7 +355,7 @@ class SchedulerProfilerManager:
             torch.cuda.memory._record_memory_history(enabled=None)
 
         if "CUDA_PROFILER" in self.profiler_activities:
-            if self.ps.gpu_id == get_global_server_args().base_gpu_id:
+            if self.ps.gpu_id == get_server_args().base_gpu_id:
                 torch.cuda.cudart().cudaProfilerStop()
 
         merge_message = self._merge_profile_traces()

python/sglang/srt/mem_cache/base_prefix_cache.py

@@ -216,9 +216,9 @@ class BasePrefixCache(ABC, PrefixCacheTrait):
     )
 
     def init_metrics_collector(self):
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         labels = {"cache_type": self.__class__.__name__}
         if server_args.extra_metric_labels:
             labels.update(server_args.extra_metric_labels)

python/sglang/srt/mem_cache/common.py

@@ -26,7 +26,7 @@ from sglang.srt.mem_cache.triton_ops.common import (
     write_req_to_token_pool_triton,
 )
 from sglang.srt.runtime_context import get_server_args
-from sglang.srt.server_args import ServerArgs, get_global_server_args
+from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import is_cuda, is_hip, is_npu, support_triton
 from sglang.srt.utils.common import ceil_align, is_pin_memory_available
 
@@ -214,7 +214,7 @@ def get_last_loc_torch(
 
 def get_alloc_len_per_decode(server_args: Optional[ServerArgs] = None) -> int:
     if server_args is None:
-        server_args = get_global_server_args()
+        server_args = get_server_args()
 
     if server_args.speculative_algorithm is None:
         return 1
@@ -444,7 +444,7 @@ def _alloc_page_size(batch: ScheduleBatch) -> int:
     # DCP swaps in an allocator whose page_size is server_args.page_size *
     # dcp_size, so it can be > 1 even when tree_cache.page_size is 1; branch on
     # the real allocator's page_size there. Elsewhere the two are equal.
-    if (_is_hip or _is_cuda) and get_global_server_args().dcp_size > 1:
+    if (_is_hip or _is_cuda) and get_server_args().dcp_size > 1:
         return batch.tree_cache.token_to_kv_pool_allocator.page_size
     return batch.tree_cache.page_size
 
@@ -658,7 +658,7 @@ def release_kv_cache(req: Req, tree_cache: BasePrefixCache, is_insert: bool = Tr
 
     start_p, end_p = req.pop_overallocated_kv_cache()
 
-    global_server_args = get_global_server_args()
+    global_server_args = get_server_args()
     page_size = global_server_args.page_size
     spec_algo = global_server_args.speculative_algorithm
 

python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py

@@ -21,7 +21,7 @@ from sglang.srt.layers.attention.dsv4.index_buf_accessor import NopeFp8RopeBf16P
 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.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import ceil_div, is_hip
 
 logger = logging.getLogger(__name__)
@@ -276,7 +276,7 @@ class DeepSeekV4IndexerPool(KVCache):
             end_layer,
         )
         self.index_head_dim = index_head_dim
-        self.use_fp4_indexer = get_global_server_args().enable_deepseek_v4_fp4_indexer
+        self.use_fp4_indexer = get_server_args().enable_deepseek_v4_fp4_indexer
 
         self._create_buffer()
 
@@ -569,7 +569,7 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
             self.swa_kv_pool = None
             self.c4_kv_pool = None
             self.c128_kv_pool = None
-            server_args = get_global_server_args()
+            server_args = get_server_args()
             spec_extra = (
                 (server_args.speculative_num_draft_tokens - 1)
                 if server_args.speculative_algorithm is not None
@@ -651,7 +651,7 @@ 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_global_server_args()
+        server_args = get_server_args()
         is_speculative = server_args.speculative_algorithm is not None
         return get_compress_state_ring_size(compress_ratio, is_speculative)
 

python/sglang/srt/mem_cache/hi_mamba_radix_cache.py

@@ -1030,9 +1030,9 @@ class HiMambaRadixCache(MambaRadixCache):
             node_update = node_update.parent
 
         if len(value) > best_value_len:
-            from sglang.srt.server_args import get_global_server_args
+            from sglang.srt.runtime_context import get_server_args
 
-            mamba_cache_chunk_size = get_global_server_args().mamba_cache_chunk_size
+            mamba_cache_chunk_size = get_server_args().mamba_cache_chunk_size
             mamba_cache_chunk_aligned_seqlen = (
                 sum(len(v) for v in value) // mamba_cache_chunk_size
             ) * mamba_cache_chunk_size
@@ -1272,10 +1272,10 @@ class HiMambaRadixCache(MambaRadixCache):
             }
             if extra_metric_labels:
                 labels.update(extra_metric_labels)
-            from sglang.srt.server_args import get_global_server_args
+            from sglang.srt.runtime_context import get_server_args
 
             storage_cls = resolve_collector_class(
-                get_global_server_args(),
+                get_server_args(),
                 STAT_LOGGER_ROLE_STORAGE,
                 StorageMetricsCollector,
             )

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -338,10 +338,10 @@ class HiRadixCache(RadixCache):
                 labels.update(extra_metric_labels)
             existing_collector = getattr(self, "storage_metrics_collector", None)
             if existing_collector is None:
-                from sglang.srt.server_args import get_global_server_args
+                from sglang.srt.runtime_context import get_server_args
 
                 storage_cls = resolve_collector_class(
-                    get_global_server_args(),
+                    get_server_args(),
                     STAT_LOGGER_ROLE_STORAGE,
                     StorageMetricsCollector,
                 )

python/sglang/srt/mem_cache/mamba_checkpoint_pool.py

@@ -313,10 +313,10 @@ def maybe_init_int8_mamba_checkpoint_pool(
     allocating, so an oversized ``--int8-mamba-ckpt-size`` fails with an actionable
     message instead of a cryptic mid-allocation CUDA OOM.
     """
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
     try:
-        _sa = get_global_server_args()
+        _sa = get_server_args()
     except ValueError:
         # Some unit-test / mock runners construct HybridReqToTokenPool directly
         # without a global server-args context. The int8 checkpoint pool is opt-in

python/sglang/srt/mem_cache/mamba_radix_cache.py

@@ -50,7 +50,7 @@ from sglang.srt.mem_cache.multi_ended_allocator import (
 )
 from sglang.srt.mem_cache.radix_cache import RadixKey
 from sglang.srt.mem_cache.utils import split_node_hash_value
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import Req
@@ -435,7 +435,7 @@ class MambaRadixCache(KVCacheEventMixin, BasePrefixCache):
         )
         self.req_to_token_pool: HybridReqToTokenPool = params.req_to_token_pool
         self.token_to_kv_pool_allocator = params.token_to_kv_pool_allocator
-        self.mamba_cache_chunk_size = get_global_server_args().mamba_cache_chunk_size
+        self.mamba_cache_chunk_size = get_server_args().mamba_cache_chunk_size
 
         self.page_size = params.page_size
         self.disable = params.disable

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.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 try:
     from lmcache.integration.sglang.multi_process_adapter import LMCacheMPConnector
@@ -108,7 +108,7 @@ class LMCRadixCache(RadixCache):
     ):
         super().__init__(params)
 
-        cli_lmc_cfg = get_global_server_args().lmcache_config_file or ""
+        cli_lmc_cfg = get_server_args().lmcache_config_file or ""
 
         kvcache = self.token_to_kv_pool_allocator.get_kvcache()
         connector_kwargs = dict(
@@ -438,7 +438,7 @@ class LMCRadixCache(RadixCache):
                 self.lmcache_connector.end_session(req.rid)
             return
 
-        global_server_args = get_global_server_args()
+        global_server_args = get_server_args()
         topk = global_server_args.speculative_eagle_topk
         enable_kv_committed_len = topk is None or topk == 1
         if enable_kv_committed_len:

python/sglang/srt/mem_cache/unified_cache_components/mamba_component.py

@@ -26,7 +26,7 @@ from sglang.srt.mem_cache.unified_cache_components.tree_component import (
     TreeComponent,
     get_and_increase_time_counter,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import Req
@@ -84,7 +84,7 @@ class MambaComponent(TreeComponent):
         # states. We temporarily skip branching-state fill in that mode and can
         # add a HiCache-aware branching policy later.
         if self.cache.cache_controller is None and len(value_chunks) > best_value_len:
-            chunk_size = get_global_server_args().mamba_cache_chunk_size
+            chunk_size = get_server_args().mamba_cache_chunk_size
             aligned_seqlen = (
                 sum(len(v) for v in value_chunks) // chunk_size
             ) * chunk_size

python/sglang/srt/mem_cache/unified_radix_cache.py

@@ -2280,10 +2280,10 @@ class UnifiedRadixCache(KVCacheEventMixin, BasePrefixCache):
                 labels.update(extra_metric_labels)
             existing_collector = self.storage_metrics_collector
             if existing_collector is None:
-                from sglang.srt.server_args import get_global_server_args
+                from sglang.srt.runtime_context import get_server_args
 
                 storage_cls = resolve_collector_class(
-                    get_global_server_args(),
+                    get_server_args(),
                     STAT_LOGGER_ROLE_STORAGE,
                     StorageMetricsCollector,
                 )

python/sglang/srt/model_executor/cuda_graph_config.py

@@ -16,7 +16,7 @@ cuda_graph_config, and the --cuda-graph-config JSON CLI parser.
 
 Module-level imports are pure stdlib — no torch / sglang.srt deps — so
 ServerArgs can import everything here without pulling in backend
-classes. check_cuda_graph_backend lazy-imports get_global_server_args
+classes. check_cuda_graph_backend lazy-imports get_server_args
 inside the function body to preserve that invariant.
 """
 
@@ -164,10 +164,10 @@ def check_cuda_graph_backend(phase: str, backend: str) -> bool:
     """True if cuda_graph_config[phase].backend == backend on the
     global server args. Returns False if the global server args have not
     been initialized yet (e.g. unit tests, early startup)."""
-    from sglang.srt.server_args import get_global_server_args
+    from sglang.srt.runtime_context import get_server_args
 
     try:
-        server_args = get_global_server_args()
+        server_args = get_server_args()
     except ValueError:
         return False
     cfg = server_args.cuda_graph_config

python/sglang/srt/model_executor/forward_batch_info.py

@@ -49,8 +49,7 @@ from sglang.srt.model_executor.forward_batch_deepseek_mha_mixin import (
     ForwardBatchDeepSeekMHAMixin,
 )
 from sglang.srt.model_executor.triton_ops.position import compute_position_triton
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     is_cuda,
     is_hip,
@@ -910,7 +909,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
             # --enable-mis: every request must carry delimiter indices (the score
             # endpoint always produces MIS-structured requests; consumers index
             # without None-checking).
-            if get_global_server_args().enable_mis and any(
+            if get_server_args().enable_mis and any(
                 r.multi_item_delimiter_indices is not None for r in batch.reqs
             ):
                 assert all(
@@ -1073,7 +1072,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
                 # 3 * N
                 if (
                     mm_input is None
-                    or get_global_server_args().rl_on_policy_target is not None
+                    or get_server_args().rl_on_policy_target is not None
                 ):
                     mrope_positions_list[batch_idx] = torch.full(
                         (3, 1),
@@ -1092,7 +1091,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
                 )
                 if (
                     mm_input is None
-                    or get_global_server_args().rl_on_policy_target is not None
+                    or get_server_args().rl_on_policy_target is not None
                 ):
                     # text only
                     mrope_positions = torch.tensor(

python/sglang/srt/model_executor/model_runner.py

@@ -780,9 +780,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         torchao_applied = getattr(self.model, "torchao_applied", False)
         # In layered loading, torchao may have been applied
         if not torchao_applied:
-            apply_torchao_config_to_model(
+            apply_torchao_config_to_model(self.model, get_server_args().torchao_config)
-                self.model, get_global_server_args().torchao_config
-            )
 
         # Apply torch TP if the model supports it
         supports_torch_tp = getattr(self.model, "supports_torch_tp", False)
@@ -1007,7 +1005,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
 
         set_global_experts_capturer(
             RoutedExpertsCapturer.create(
-                enable=get_global_server_args().enable_return_routed_experts,
+                enable=get_server_args().enable_return_routed_experts,
                 model_config=self.model_config,
                 num_fused_shared_experts=num_fused_shared_experts,
                 num_tokens=self.max_total_num_tokens + self.page_size,
@@ -1017,7 +1015,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         )
 
     def init_indexer_capturer(self):
-        enable = get_global_server_args().enable_return_indexer_topk
+        enable = get_server_args().enable_return_indexer_topk
         # Producer wiring is CUDA-only (Indexer.forward_cuda + MLA skip_topk
         # path); other backends would create a capturer but never feed it.
         if enable and self.device != "cuda":
@@ -1709,8 +1707,8 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             else:
                 # Load the missing weights from disk
                 self.update_weights_from_disk(
-                    get_global_server_args().model_path,
+                    get_server_args().model_path,
-                    get_global_server_args().load_format,
+                    get_server_args().load_format,
                     weight_name_filter=weight_name_filter,
                 )
 

python/sglang/srt/model_loader/loader.py

@@ -44,7 +44,7 @@ from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
     get_remote_instance_transfer_engine_info_per_rank,
     register_memory_region,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import get_available_gpu_memory
 
 # Try to import accelerate (optional dependency)
@@ -481,7 +481,7 @@ class DefaultModelLoader(BaseModelLoader):
         else:
             hf_folder = model_name_or_path
 
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         if server_args and server_args.model_checksum is not None:
             from sglang.srt.utils.model_file_verifier import verify
 
@@ -567,7 +567,7 @@ class DefaultModelLoader(BaseModelLoader):
                 hf_weights_files,
             )
         elif use_safetensors:
-            server_args = get_global_server_args()
+            server_args = get_server_args()
             weight_loader_disable_mmap = server_args.weight_loader_disable_mmap
             weight_loader_prefetch = server_args.weight_loader_prefetch_checkpoints
             prefetch_num_threads = server_args.weight_loader_prefetch_num_threads
@@ -866,9 +866,9 @@ class LayeredModelLoader(DefaultModelLoader):
         device_config: DeviceConfig,
     ) -> nn.Module:
         from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
-        torchao_config = get_global_server_args().torchao_config
+        torchao_config = get_server_args().torchao_config
         target_device = torch.device(device_config.device)
         quant_config = _get_quantization_config(model_config, self.load_config)
 
@@ -3078,7 +3078,7 @@ class RunaiModelStreamerLoader(BaseModelLoader):
             )
         )
 
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         if server_args and server_args.model_checksum is not None:
             from sglang.srt.utils.model_file_verifier import verify
 

python/sglang/srt/models/bailing_moe.py

@@ -78,7 +78,6 @@ from sglang.srt.models.utils import (
     enable_fused_set_kv_buffer,
 )
 from sglang.srt.runtime_context import get_parallel, get_server_args, get_stream
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, is_cuda, is_non_idle_and_non_empty, make_layers
 
 LoraConfig = None
@@ -209,7 +208,7 @@ class BailingMoESparseMoeBlock(nn.Module):
             self.router_dtype = torch.bfloat16
 
         # TODO global_server_args.ep_num_redundant_experts is used for eplb, not supported now
-        assert get_global_server_args().ep_num_redundant_experts == 0
+        assert get_server_args().ep_num_redundant_experts == 0
         # check group topk
         self.num_expert_group = getattr(config, "n_group", 0)
         self.topk_group = getattr(config, "topk_group", 0)
@@ -224,7 +223,7 @@ class BailingMoESparseMoeBlock(nn.Module):
             self.use_grouped_topk = False
 
         self.num_experts = (
-            config.num_experts + get_global_server_args().ep_num_redundant_experts
+            config.num_experts + get_server_args().ep_num_redundant_experts
         )
 
         self.gate = BailingMoEGate(

python/sglang/srt/models/bailing_moe_linear.py

@@ -59,7 +59,6 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA, DeepseekV2MLP, _is_hip
 from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.runtime_context import get_parallel, get_server_args, get_stream
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     BumpAllocator,
     add_prefix,
@@ -534,7 +533,7 @@ class BailingMoELinearAttention(nn.Module):
             base=self.rope_theta,
             rope_scaling=config.rope_scaling,
             is_neox_style=True,
-            device=get_global_server_args().device,
+            device=get_server_args().device,
             dtype=torch.float32,
         )
 
@@ -695,7 +694,7 @@ class BailingMoEAttention(nn.Module):
             max_position=self.max_position_embeddings,
             base=self.rope_theta,
             rope_scaling=config.rope_scaling,
-            device=get_global_server_args().device,
+            device=get_server_args().device,
         )
         self.attn = RadixAttention(
             self.num_heads,

python/sglang/srt/models/bert.py

@@ -16,8 +16,7 @@ from sglang.srt.layers.radix_attention import AttentionType, RadixAttention
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix
 
 BertConfig = None
@@ -367,9 +366,7 @@ class BertModel(nn.Module):
             prefix=add_prefix("encoder", prefix),
         )
         pooling_type = (
-            PoolingType.CLS
+            PoolingType.CLS if get_server_args().is_embedding else PoolingType.LAST
-            if get_global_server_args().is_embedding
-            else PoolingType.LAST
         )
         self.pooler = (
             BertPooler(config)

python/sglang/srt/models/deepseek_common/attention_backend_handler.py

@@ -11,7 +11,7 @@ from sglang.srt.models.deepseek_common.attention_forward_methods.forward_methods
     AttnForwardMethod,
 )
 from sglang.srt.models.deepseek_common.utils import _is_hip
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import use_intel_amx_backend
 
 MHA_ONE_SHOT_SUPPORTED_BACKENDS = ["fa3", "flashinfer", "flashmla"]
@@ -114,7 +114,7 @@ def handle_attention_flashinfer(attn, forward_batch):
 
 def handle_attention_fa3(attn, forward_batch):
     # when deterministic inference is enabled, use MLA
-    if get_global_server_args().enable_deterministic_inference:
+    if get_server_args().enable_deterministic_inference:
         return _dispatch_mla_subtype(attn, forward_batch)
     else:
         return _handle_attention_backend(attn, forward_batch, "fa3")
@@ -183,7 +183,7 @@ def handle_attention_triton(attn, forward_batch):
         return AttnForwardMethod.MLA
 
     # when deterministic inference is enabled, use MLA
-    if get_global_server_args().enable_deterministic_inference:
+    if get_server_args().enable_deterministic_inference:
         return _dispatch_mla_subtype(attn, forward_batch)
 
     if (

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

@@ -31,7 +31,7 @@ from sglang.srt.models.deepseek_common.utils import (
     _use_aiter_bpreshuffle_gfx95,
     _use_aiter_gfx95,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import BumpAllocator, get_bool_env_var, next_power_of_2
 
 _use_fp8_prefill_attn = (
@@ -115,7 +115,7 @@ class DeepseekMHAForwardMixin:
 
     def init_mha_forward(self: DeepseekV2AttentionMLA):
         self.disable_chunked_prefix_cache = (
-            get_global_server_args().disable_chunked_prefix_cache
+            get_server_args().disable_chunked_prefix_cache
         )
 
         # TODO: Design a finer way to determine the threshold
@@ -279,8 +279,8 @@ class DeepseekMHAForwardMixin:
                 self.use_dsa
                 and self.kv_cache_dtype == "fp8_e4m3"
                 and (
-                    not get_global_server_args().dsa_decode_backend == "trtllm"
+                    not get_server_args().dsa_decode_backend == "trtllm"
-                    or not get_global_server_args().dsa_prefill_backend == "trtllm"
+                    or not get_server_args().dsa_prefill_backend == "trtllm"
                 )
             ):
                 # FP8 path: dequantize DSA-specific FP8 format to BF16

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

@@ -67,7 +67,7 @@ from sglang.srt.models.deepseek_common.utils import (
     _use_aiter_bpreshuffle_gfx95,
     _use_aiter_gfx95,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.state_capturer.indexer_topk import (
     maybe_capture_indexer_topk,
 )
@@ -178,7 +178,7 @@ def _should_defer_dsa_cp_kv_gather(
 class DeepseekMLAForwardMixin:
     def init_mla_forward(self: DeepseekV2AttentionMLA):
         self.flashinfer_mla_disable_ragged = (
-            get_global_server_args().flashinfer_mla_disable_ragged
+            get_server_args().flashinfer_mla_disable_ragged
         )
 
     def should_run_indexer(
@@ -1002,8 +1002,8 @@ class DeepseekMLAForwardMixin:
         """
         if self.current_attention_backend in ("dsa", "nsa"):
             return (
-                get_global_server_args().dsa_decode_backend == "trtllm"
+                get_server_args().dsa_decode_backend == "trtllm"
-                or get_global_server_args().dsa_prefill_backend == "trtllm"
+                or get_server_args().dsa_prefill_backend == "trtllm"
             ) and get_attn_backend().kv_cache_dtype == torch.float8_e4m3fn
 
         return (
@@ -1020,7 +1020,7 @@ class DeepseekMLAForwardMixin:
         """
         Check if we should skip rope and use fused rope+cache path for TileLang DSA on gfx95.
         """
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         return (
             _use_aiter_gfx95
             and self.current_attention_backend in ("dsa", "nsa")

python/sglang/srt/models/deepseek_nextn.py

@@ -58,7 +58,6 @@ from sglang.srt.models.deepseek_common.utils import enable_nextn_moe_bf16_cast_t
 from sglang.srt.models.deepseek_v2 import DeepseekV2DecoderLayer, DeepseekV3ForCausalLM
 from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import BumpAllocator, add_prefix, is_cuda, is_npu
 
 logger = logging.getLogger(__name__)
@@ -119,7 +118,7 @@ class DeepseekModelNextN(nn.Module):
 
         self.rot_weight = None
         if _is_npu:
-            rot_weight_path = get_global_server_args().model_path + "/rot.safetensors"
+            rot_weight_path = get_server_args().model_path + "/rot.safetensors"
             if os.path.isfile(rot_weight_path):
                 self.rot_weight = load_file(rot_weight_path)
                 self.rot_weight = self.rot_weight["rot.weight"].npu()
@@ -132,8 +131,8 @@ class DeepseekModelNextN(nn.Module):
 
         layer_name = "decoder"
         if _is_npu and (
-            get_global_server_args().speculative_draft_model_path
+            get_server_args().speculative_draft_model_path
-            == get_global_server_args().model_path
+            == get_server_args().model_path
         ):
             layer_name = "layers." + str(config.num_hidden_layers)
 

python/sglang/srt/models/deepseek_v2.py

@@ -180,7 +180,6 @@ from sglang.srt.models.deepseek_common.utils import (
     _use_aiter_gfx95,
 )
 from sglang.srt.runtime_context import get_flags, get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.utils import (
     BumpAllocator,
@@ -484,7 +483,7 @@ class MoEGate(nn.Module):
                 True,  # is_vnni
             )
 
-        if get_global_server_args().enable_deterministic_inference:
+        if get_server_args().enable_deterministic_inference:
             return F.linear(hidden_states, self.weight, None)
 
         if (
@@ -621,7 +620,7 @@ class DeepseekV2MoE(nn.Module):
 
         self.experts = get_moe_impl_class(quant_config)(
             num_experts=num_experts_for_moe
-            + get_global_server_args().ep_num_redundant_experts,
+            + get_server_args().ep_num_redundant_experts,
             num_fused_shared_experts=self.num_fused_shared_experts,
             top_k=top_k_for_moe,
             hidden_size=config.hidden_size,
@@ -794,8 +793,7 @@ class DeepseekV2MoE(nn.Module):
             # TODO: we will support tp < ep in the future
             self.ep_size = get_parallel().moe_ep_size
             self.num_experts = (
-                config.n_routed_experts
+                config.n_routed_experts + get_server_args().ep_num_redundant_experts
-                + get_global_server_args().ep_num_redundant_experts
             )
             self.renormalize = config.norm_topk_prob
             self.topk_group = config.topk_group
@@ -836,7 +834,7 @@ class DeepseekV2MoE(nn.Module):
         self, hidden_states: torch.Tensor, server_args=None
     ) -> bool:
         if server_args is None:
-            server_args = get_global_server_args()
+            server_args = get_server_args()
         return (
             _enable_pcg_dsv2_dual_stream
             and (is_in_tc_piecewise_cuda_graph() or is_in_breakable_cuda_graph())
@@ -874,7 +872,7 @@ class DeepseekV2MoE(nn.Module):
             )
 
         if not self._enable_a2a_moe:
-            server_args = get_global_server_args()
+            server_args = get_server_args()
             if self._can_dual_stream_graph(hidden_states, server_args):
                 return dsv2_flashinfer_moe_dual_stream_graph(
                     hidden_states,
@@ -935,7 +933,7 @@ class DeepseekV2MoE(nn.Module):
         # into the decode CUDA graph and replays from null.
         current_stream = torch.cuda.current_stream()
         self.alt_stream.wait_stream(current_stream)
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         dispatch_info = (
             ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id)
             if server_args.enable_eplb
@@ -1032,7 +1030,7 @@ class DeepseekV2MoE(nn.Module):
             self.shared_experts.gate_up_proj
         ):
             return self.forward_cpu(hidden_states, should_allreduce_fusion)
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         dispatch_info = (
             ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id)
             if server_args.enable_eplb
@@ -1599,7 +1597,7 @@ class DeepseekV2AttentionMLA(
         self.scaling = self.qk_head_dim**-0.5
         self.rope_theta = rope_theta
         self.max_position_embeddings = max_position_embeddings
-        self.kv_cache_dtype = get_global_server_args().kv_cache_dtype
+        self.kv_cache_dtype = get_server_args().kv_cache_dtype
 
         # NOTE modification to rope_scaling must be done early enough, b/c e.g. Indexer needs it
         if rope_scaling:
@@ -1711,7 +1709,7 @@ class DeepseekV2AttentionMLA(
                 base=rope_theta,
                 rope_scaling=rope_scaling,
                 is_neox_style=is_neox_style,
-                device=get_global_server_args().device,
+                device=get_server_args().device,
             )
 
             if rope_scaling and rope_scaling.get("apply_yarn_scaling", True):
@@ -1798,7 +1796,7 @@ class DeepseekV2AttentionMLA(
         # Determine attention backend name for current forward batch: prefer the
         # name stamped per-runner on the backend object, else resolve from server args.
         backend = get_attn_backend()
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         default_prefill_str, default_decode_str = server_args.get_attention_backends()
         prefill_backend_str = (
             backend.prefill_attention_backend_str or default_prefill_str
@@ -2071,7 +2069,7 @@ class DeepseekV2DecoderLayer(nn.Module):
             rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            get_global_server_args().speculative_algorithm
+            get_server_args().speculative_algorithm
         )
         self.dsa_enable_prefill_cp = dsa_enable_prefill_cp
         self.mla_enable_prefill_cp = mla_enable_prefill_cp
@@ -2749,7 +2747,7 @@ class DeepseekV2ForCausalLM(nn.Module, DeepseekV2WeightLoaderMixin):
         self, architecture: str = "DeepseekV3ForCausalLM"
     ):
         self.num_fused_shared_experts = 0
-        server_args = get_global_server_args()
+        server_args = get_server_args()
 
         if get_server_args().disable_shared_experts_fusion:
             return

python/sglang/srt/models/deepseek_v4.py

@@ -135,7 +135,7 @@ if _is_xpu:
 else:
     from sglang.srt.layers.mhc import hc_split_sinkhorn, mhc_fused_post_pre, npu_hc_pre
 
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
     LazyValue,
     add_prefix,
@@ -346,7 +346,7 @@ class MQALayer(nn.Module):
             base=rope_base,
             rope_scaling=rope_scaling,
             is_neox_style=False,
-            device=get_global_server_args().device,
+            device=get_server_args().device,
         )
 
         from sglang.srt.layers.deepseek_v4_rope import precompute_freqs_cis
@@ -2210,7 +2210,7 @@ class DeepseekV4ForCausalLM(nn.Module):
             return
 
         disable_reason = None
-        if get_global_server_args().enforce_shared_experts_fusion:
+        if get_server_args().enforce_shared_experts_fusion:
             if self.config.n_shared_experts != 1:
                 raise ValueError(
                     "DeepSeek V4 shared-experts fusion expects exactly one shared "

python/sglang/srt/models/exaone_moe.py

@@ -63,7 +63,6 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTe
 from sglang.srt.model_executor.runner import get_is_capture_mode
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.runtime_context import get_parallel, get_server_args, get_stream
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import LazyValue, add_prefix, is_cuda, make_layers
 
 logger = logging.getLogger(__name__)
@@ -171,8 +170,7 @@ class ExaoneMoESparseMoEBlock(nn.Module):
         )
 
         self.experts = get_moe_impl_class(quant_config)(
-            num_experts=config.num_experts
+            num_experts=config.num_experts + get_server_args().ep_num_redundant_experts,
-            + get_global_server_args().ep_num_redundant_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
@@ -213,7 +211,7 @@ class ExaoneMoESparseMoEBlock(nn.Module):
         if get_moe_a2a_backend().is_deepep():
             self.ep_size = get_parallel().moe_ep_size
             self.num_experts = (
-                config.num_experts + get_global_server_args().ep_num_redundant_experts
+                config.num_experts + get_server_args().ep_num_redundant_experts
             )
             self.top_k = config.num_experts_per_tok
 

python/sglang/srt/models/gemma4_causal.py

@@ -58,8 +58,7 @@ from sglang.srt.models.gemma3_causal import Gemma3MLP, Gemma3TextScaledWordEmbed
 from sglang.srt.models.utils import (
     create_fused_set_kv_buffer_arg,
 )
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, make_layers
 
 logger = logging.getLogger(__name__)
@@ -255,8 +254,7 @@ class Gemma4MoE(nn.Module):
         experts_type = get_moe_impl_class(quant_config)
 
         self.experts = experts_type(
-            num_experts=config.num_experts
+            num_experts=config.num_experts + get_server_args().ep_num_redundant_experts,
-            + get_global_server_args().ep_num_redundant_experts,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
             layer_id=layer_id,
@@ -789,7 +787,7 @@ class Gemma4TextModel(PreTrainedModel):
         # combination until the runner becomes schema-aware; users can run
         # PP + PLE eagerly with --disable-cuda-graph.
         if self.pp_group.world_size > 1 and self.hidden_size_per_layer_input > 0:
-            sa = get_global_server_args()
+            sa = get_server_args()
             if sa is not None and not sa.disable_cuda_graph:
                 raise ValueError(
                     "Pipeline parallelism is currently incompatible with "

python/sglang/srt/models/gemma4_vision.py

@@ -181,9 +181,9 @@ class Gemma4VisionAttention(nn.Module):
     @staticmethod
     def _select_backend() -> str:
         """Mirror VisionAttention._determine_attention_backend for consistency."""
-        from sglang.srt.server_args import get_global_server_args
+        from sglang.srt.runtime_context import get_server_args
 
-        override = get_global_server_args().mm_attention_backend
+        override = get_server_args().mm_attention_backend
         if override is not None:
             return override
         if is_cuda():

python/sglang/srt/models/glm4_moe.py

@@ -88,7 +88,6 @@ from sglang.srt.runtime_context import (
     get_server_args,
     get_stream,
 )
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     add_prefix,
     cpu_has_amx_support,
@@ -530,8 +529,7 @@ class Glm4MoeSparseMoeBlock(nn.Module):
             # TODO: we will support tp < ep in the future
             self.ep_size = get_parallel().moe_ep_size
             self.num_experts = (
-                config.n_routed_experts
+                config.n_routed_experts + get_server_args().ep_num_redundant_experts
-                + get_global_server_args().ep_num_redundant_experts
             )
             self.renormalize = config.norm_topk_prob
             self.topk_group = config.topk_group

python/sglang/srt/models/glm4_moe_lite.py

@@ -75,7 +75,6 @@ from sglang.srt.models.deepseek_common.deepseek_weight_loader import (
 from sglang.srt.models.deepseek_common.utils import _is_cuda, _use_aiter
 from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA
 from sglang.srt.runtime_context import get_parallel, get_server_args, get_stream
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     BumpAllocator,
     LazyValue,
@@ -216,7 +215,7 @@ class Glm4MoeLiteSparseMoeBlock(nn.Module):
         self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.n_routed_experts
             + self.num_fused_shared_experts
-            + get_global_server_args().ep_num_redundant_experts,
+            + get_server_args().ep_num_redundant_experts,
             num_fused_shared_experts=self.num_fused_shared_experts,
             top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
             hidden_size=config.hidden_size,
@@ -284,8 +283,7 @@ class Glm4MoeLiteSparseMoeBlock(nn.Module):
             # TODO: we will support tp < ep in the future
             self.ep_size = get_parallel().moe_ep_size
             self.num_experts = (
-                config.n_routed_experts
+                config.n_routed_experts + get_server_args().ep_num_redundant_experts
-                + get_global_server_args().ep_num_redundant_experts
             )
             self.renormalize = config.norm_topk_prob
             self.topk_group = config.topk_group

python/sglang/srt/models/glm4_moe_lite_nextn.py

@@ -36,7 +36,6 @@ from sglang.srt.models.glm4_moe_lite import (
     Glm4MoeLiteForCausalLM,
 )
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import BumpAllocator, add_prefix, is_npu
 
 logger = logging.getLogger(__name__)
@@ -140,10 +139,7 @@ class Glm4MoeLiteForCausalLMNextN(Glm4MoeLiteForCausalLM):
         nn.Module.__init__(self)
         self.config = config
         self.tp_size = get_parallel().tp_size
-        if (
+        if is_npu() and get_server_args().speculative_draft_model_quantization is None:
-            is_npu()
-            and get_global_server_args().speculative_draft_model_quantization is None
-        ):
             quant_config = None
         self.quant_config = quant_config
 

python/sglang/srt/models/glm4_moe_nextn.py

@@ -33,7 +33,6 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.glm4_moe import Glm4MoeDecoderLayer, Glm4MoeForCausalLM
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, is_npu
 
 logger = logging.getLogger(__name__)
@@ -126,10 +125,7 @@ class Glm4MoeForCausalLMNextN(Glm4MoeForCausalLM):
         nn.Module.__init__(self)
         self.config = config
         self.tp_size = get_parallel().tp_size
-        if (
+        if is_npu() and get_server_args().speculative_draft_model_quantization is None:
-            is_npu()
-            and get_global_server_args().speculative_draft_model_quantization is None
-        ):
             quant_config = None
         self.quant_config = quant_config
 

python/sglang/srt/models/glm4v.py

@@ -53,8 +53,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTe
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.glm4 import Glm4Model
 from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, is_npu
 from sglang.srt.utils.hf_transformers_utils import get_processor
 
@@ -548,7 +547,7 @@ class Glm4vForConditionalGeneration(nn.Module):
 
         self.pp_group = get_pp_group()
         self.config = config
-        self.use_data_parallel = get_global_server_args().mm_enable_dp_encoder
+        self.use_data_parallel = get_server_args().mm_enable_dp_encoder
         vision_utils.update_vit_attn_dummy_heads_config(self.config)
         self.visual = Glm4vVisionModel(
             config.vision_config,

python/sglang/srt/models/glm4v_moe.py

@@ -19,7 +19,6 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.glm4_moe import Glm4MoeModel
 from sglang.srt.models.glm4v import Glm4vForConditionalGeneration, Glm4vVisionModel
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 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_processor
 
@@ -42,7 +41,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
 
         self.pp_group = get_pp_group()
         self.config = config
-        self.use_data_parallel = get_global_server_args().mm_enable_dp_encoder
+        self.use_data_parallel = get_server_args().mm_enable_dp_encoder
         vision_utils.update_vit_attn_dummy_heads_config(self.config)
         self.tp_size = get_parallel().tp_size
         self.quant_config = quant_config

python/sglang/srt/models/glm_ocr.py

@@ -50,7 +50,7 @@ from sglang.srt.models.glm4v import (
     Glm4vVisionModel,
     Glm4vVisionPatchEmbed,
 )
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import add_prefix
 from sglang.srt.utils.hf_transformers_utils import get_processor
 
@@ -272,7 +272,7 @@ class GlmOcrForConditionalGeneration(Glm4vForConditionalGeneration):
 
         self.pp_group = get_pp_group()
         self.config = config
-        self.use_data_parallel = get_global_server_args().mm_enable_dp_encoder
+        self.use_data_parallel = get_server_args().mm_enable_dp_encoder
         self.visual = GlmOcrVisionModel(
             vision_config=config.vision_config,
             text_config=config.text_config,

python/sglang/srt/models/gpt_oss.py

@@ -69,7 +69,6 @@ from sglang.srt.models.utils import (
     enable_fused_set_kv_buffer,
 )
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     LazyValue,
     add_prefix,
@@ -228,7 +227,7 @@ class GptOssSparseMoeBlock(nn.Module):
 
         self.experts = experts_type(
             num_experts=config.num_local_experts
-            + get_global_server_args().ep_num_redundant_experts,
+            + get_server_args().ep_num_redundant_experts,
             top_k=config.num_experts_per_tok,
             layer_id=layer_id,
             hidden_size=config.hidden_size,

python/sglang/srt/models/internvl.py

@@ -40,8 +40,7 @@ from sglang.srt.multimodal.internvl_vit_cuda_graph_runner import (
     InternViTCudaGraphRunner,
 )
 from sglang.srt.multimodal.mm_utils import run_dp_sharded_vision_model
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import is_cuda
 from sglang.utils import logger
 
@@ -498,7 +497,7 @@ class InternVLChatModel(nn.Module):
     ) -> None:
         super().__init__()
         self.config = config
-        self.use_data_parallel = get_global_server_args().mm_enable_dp_encoder
+        self.use_data_parallel = get_server_args().mm_enable_dp_encoder
         self.quant_config = quant_config
         vision_utils.update_vit_attn_dummy_heads_config(self.config)
         image_size = config.force_image_size or config.vision_config.image_size

python/sglang/srt/models/kimi_k25.py

@@ -31,7 +31,7 @@ from sglang.srt.models.deepseek_v2 import DeepseekV3ForCausalLM
 from sglang.srt.models.kimi_vl_moonvit import MLP2
 from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import add_prefix, is_npu
 
 logger = logging.getLogger(__name__)
@@ -633,7 +633,7 @@ class KimiK25ForConditionalGeneration(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.use_data_parallel = get_global_server_args().mm_enable_dp_encoder
+        self.use_data_parallel = get_server_args().mm_enable_dp_encoder
         # Create vision tower
         self.vision_tower = MoonViT3dPretrainedModel(
             config.vision_config,

python/sglang/srt/models/laguna.py

@@ -54,7 +54,6 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTe
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.utils import apply_qk_norm
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import LazyValue, add_prefix, make_layers
 
 logger = logging.getLogger(__name__)
@@ -161,8 +160,7 @@ class LagunaMoE(nn.Module):
         self.gate = LagunaMoEGate(config, prefix=add_prefix("gate", prefix))
 
         self.experts = get_moe_impl_class(quant_config)(
-            num_experts=config.num_experts
+            num_experts=config.num_experts + get_server_args().ep_num_redundant_experts,
-            + get_global_server_args().ep_num_redundant_experts,
             top_k=config.num_experts_per_tok,
             layer_id=layer_id,
             hidden_size=config.hidden_size,

python/sglang/srt/models/llada2.py

@@ -77,7 +77,6 @@ from sglang.srt.models.utils import (
     enable_fused_set_kv_buffer,
 )
 from sglang.srt.runtime_context import get_parallel, get_server_args, get_stream
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     add_prefix,
     is_cuda,
@@ -220,7 +219,7 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
             self.router_dtype = torch.bfloat16
 
         # TODO global_server_args.ep_num_redundant_experts is used for eplb, not supported now
-        assert get_global_server_args().ep_num_redundant_experts == 0
+        assert get_server_args().ep_num_redundant_experts == 0
         # check group topk
         self.num_expert_group = getattr(config, "n_group", 0)
         self.topk_group = getattr(config, "topk_group", 0)
@@ -235,7 +234,7 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
             self.use_grouped_topk = False
 
         self.num_experts = (
-            config.num_experts + get_global_server_args().ep_num_redundant_experts
+            config.num_experts + get_server_args().ep_num_redundant_experts
         )
 
         self.gate = LLaDA2MoeGate(

python/sglang/srt/models/llama_eagle3.py

@@ -38,7 +38,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.llama import LlamaDecoderLayer, LlamaForCausalLM, LlamaMLP
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 
 
 class LlamaDecoderLayer(LlamaDecoderLayer):
@@ -258,7 +258,7 @@ class LlamaForCausalLMEagle3(LlamaForCausalLM):
         # Cache draft SWA size from server args once; consumed both by the post-init
         # attention patch below and by `get_attention_sliding_window_size` later.
         self._draft_window_size: Optional[int] = (
-            get_global_server_args().speculative_draft_window_size
+            get_server_args().speculative_draft_window_size
         )
 
         self.model = LlamaModel(

python/sglang/srt/models/mimo_audio.py

@@ -21,7 +21,7 @@ from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
 from transformers.models.qwen2.modeling_qwen2 import Qwen2Model
 
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import is_cuda
 
 if is_cuda():
@@ -1223,7 +1223,7 @@ class AudioEncoderMixin:
         else:
             raise ValueError(f"Invalid projection layers: {config.projection_layers}")
 
-        model_path = get_global_server_args().model_path
+        model_path = get_server_args().model_path
         if not os.path.isdir(model_path):
             from huggingface_hub import snapshot_download
 

python/sglang/srt/models/mimo_v2.py

@@ -77,7 +77,6 @@ from sglang.srt.model_loader.weight_utils import (
 from sglang.srt.models.mimo_audio import AudioEncoderMixin, MiMoAudioEncoderConfig
 from sglang.srt.models.mimo_vl import MiMoVisionTransformer, MiMoVLVisionConfig
 from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     LazyValue,
     add_prefix,
@@ -253,7 +252,7 @@ class MiMoV2MoE(nn.Module):
         experts_type = get_moe_impl_class(quant_config)
         self.experts = experts_type(
             num_experts=config.n_routed_experts
-            + get_global_server_args().ep_num_redundant_experts,
+            + get_server_args().ep_num_redundant_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
@@ -288,8 +287,7 @@ class MiMoV2MoE(nn.Module):
             # TODO: we will support tp < ep in the future
             self.ep_size = get_parallel().moe_ep_size
             self.num_experts = (
-                config.n_routed_experts
+                config.n_routed_experts + get_server_args().ep_num_redundant_experts
-                + get_global_server_args().ep_num_redundant_experts
             )
             self.renormalize = config.norm_topk_prob
             self.topk_group = config.topk_group

python/sglang/srt/models/mimo_vl.py

@@ -18,7 +18,7 @@ from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.quantization import QuantizationConfig
 from sglang.srt.models.qwen2_5_vl import Qwen2_5_VisionPatchMerger, Qwen2_5_VLMLP
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import add_prefix
 
 
@@ -232,7 +232,7 @@ class MiMoVisionTransformer(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
-        self.server_args = get_global_server_args()
+        self.server_args = get_server_args()
         self.vit_window_attn_types = vision_config.vit_window_attn_types
         patch_size: int = vision_config.patch_size
         temporal_patch_size: int = vision_config.temporal_patch_size

python/sglang/srt/models/minimax_m2.py

@@ -80,8 +80,7 @@ from sglang.srt.model_loader.weight_utils import (
     maybe_remap_kv_scale_name,
     narrow_padded_param_and_loaded_weight,
 )
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_server_args
-from sglang.srt.server_args import get_global_server_args
 
 # get_bool_env_var is defined in sglang.srt.utils.common, not sglang.srt.distributed.
 # Importing from the wrong module causes this file to fail import, which prevents the
@@ -426,7 +425,7 @@ class MiniMaxM2QKRMSNorm:
 
         props = torch.cuda.get_device_properties(device)
         # probe the maximum tokens for one prefill
-        server_args = get_global_server_args()
+        server_args = get_server_args()
         max_tokens = server_args.chunked_prefill_size
         if max_tokens is None:
             max_tokens = server_args.model_config.context_len
@@ -514,7 +513,7 @@ class MiniMaxM2MoE(nn.Module):
 
         self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.num_local_experts
-            + get_global_server_args().ep_num_redundant_experts,
+            + get_server_args().ep_num_redundant_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.intermediate_size,

python/sglang/srt/models/mllama4.py

@@ -33,7 +33,7 @@ from sglang.srt.managers.schedule_batch import (
     MultimodalInputs,
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import is_cpu
 
 _is_cpu = is_cpu()
@@ -477,7 +477,7 @@ class Llama4ForConditionalGeneration(nn.Module):
             )
 
         self.has_vision = (
-            self.has_vision_weights and get_global_server_args().enable_multimodal
+            self.has_vision_weights and get_server_args().enable_multimodal
         )
 
         if self.has_vision: