config: the lazy imports that buy nothing become eager

python/sglang/srt/arg_groups/attention_hook.py

@@ -8,9 +8,25 @@
 from typing import Any
 
 from sglang.srt.arg_groups.overrides import (
+    _attention_backend_default,
+    _attention_backend_dual_chunk,
+    _attention_backend_fa3_fp8_fallback,
+    _attention_backend_platform_fallbacks,
+    _cutedsl_prefill_backend_fill,
+    _deterministic_allreduce_fusion_disable,
+    _deterministic_attention_backend,
+    _deterministic_sampling_backend,
+    _fa4_page_constraint,
+    _intel_xpu_page_constraint,
+    _mla_backend_page_constraints,
+    _mla_kv_cache_dtype_checks,
     declare_resolution,
+    mamba_extra_buffer_of,
+    model_config_of,
     resolved_view,
     resolving_view,
+    run_post_process_pass,
+    use_mla_backend,
 )
 from sglang.srt.connector import ConnectorType
 from sglang.srt.environ import envs
@@ -29,28 +45,14 @@
 
 
 def handle_attention_backend_compatibility(server_args: Any):
-    from sglang.srt.arg_groups.overrides import (
-        attention_backends_of,
-        model_config_of,
-        use_mla_backend,
-    )
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     cfg = resolving_view(server_args)
     model_config = model_config_of(server_args)
 
     # The attention_backend write clusters of this handler moved to the
     # resolution pipeline (arg_groups/overrides.py), each invoked below at
     # its legacy slot; the interleaved non-attention adjustments stay.
-    from sglang.srt.arg_groups.overrides import (
-        _attention_backend_default,
-        _attention_backend_dual_chunk,
-        _attention_backend_fa3_fp8_fallback,
-        _attention_backend_platform_fallbacks,
-        _fa4_page_constraint,
-        _intel_xpu_page_constraint,
-        _mla_backend_page_constraints,
-        run_post_process_pass,
-    )
 
     # Split-backend override + default fill.
     run_post_process_pass(server_args, _attention_backend_default)
@@ -122,14 +124,12 @@ def handle_attention_backend_compatibility(server_args: Any):
     # The TRT-LLM / tokenspeed MLA kv-dtype validations moved to the
     # resolution pipeline (arg_groups/overrides.py:
     # _mla_kv_cache_dtype_checks), invoked here at their legacy slot.
-    from sglang.srt.arg_groups.overrides import _mla_kv_cache_dtype_checks
 
     run_post_process_pass(server_args, _mla_kv_cache_dtype_checks)
 
     # The CuteDSL MLA validation + prefill fill moved to the resolution
     # pipeline (arg_groups/overrides.py: _cutedsl_prefill_backend_fill),
     # invoked here at its legacy slot.
-    from sglang.srt.arg_groups.overrides import _cutedsl_prefill_backend_fill
 
     run_post_process_pass(server_args, _cutedsl_prefill_backend_fill)
 
@@ -334,9 +334,6 @@ def handle_linear_attn_backend(server_args: Any):
                 "KDA, as the linear-attn decode backend; got "
                 f"--linear-attn-decode-backend={decode!r}."
             )
-        from sglang.srt.arg_groups.overrides import (
-            mamba_extra_buffer_of,
-        )
 
         if mamba_extra_buffer_of(resolved_view(server_args)):
             raise ValueError(
@@ -505,7 +502,6 @@ def handle_multi_item_scoring(server_args: Any):
 
 
 def handle_deterministic_inference(server_args: Any):
-    from sglang.srt.arg_groups.overrides import model_config_of
     from sglang.srt.server_args import (
         RADIX_SUPPORTED_DETERMINISTIC_ATTENTION_BACKEND,
     )
@@ -538,21 +534,12 @@ def handle_deterministic_inference(server_args: Any):
         # Moved to the resolution pipeline (arg_groups/overrides.py:
         # _deterministic_allreduce_fusion_disable), invoked here at its
         # legacy slot.
-        from sglang.srt.arg_groups.overrides import (
-            _deterministic_allreduce_fusion_disable,
-            run_post_process_pass,
-        )
 
         run_post_process_pass(server_args, _deterministic_allreduce_fusion_disable)
 
         # The forced-pytorch sampling write and the attention backend
         # fill/validation moved to the resolution pipeline
         # (arg_groups/overrides.py), invoked at their legacy slots.
-        from sglang.srt.arg_groups.overrides import (
-            _deterministic_attention_backend,
-            _deterministic_sampling_backend,
-            run_post_process_pass,
-        )
 
         run_post_process_pass(server_args, _deterministic_sampling_backend)
         is_deepseek_model = False

python/sglang/srt/arg_groups/cuda_graph_hook.py

@@ -8,6 +8,7 @@
 
 from sglang.srt.arg_groups.overrides import (
     declare_resolution,
+    model_config_of,
     resolved_view,
     resolving_view,
 )
@@ -110,7 +111,7 @@ def apply_cuda_graph_compatibility(server_args: Any):
     prefill backend (this folds in the old
     --enforce-piecewise-cuda-graph contract).
     """
-    from sglang.srt.arg_groups.overrides import attention_backends_of, model_config_of
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     cfg = resolving_view(server_args)
     if (Phase.PREFILL, "backend") in server_args._cuda_graph_config_locked:
@@ -153,7 +154,6 @@ def disable_tc_piecewise_cudagraph_if_incompatible(server_args: Any):
     """TcPiecewise (torch.compile + piecewise) is incompatible with
     these configurations. Most are torch.compile / dynamo limitations.
     """
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
 
@@ -244,7 +244,6 @@ def disable_breakable_cudagraph_if_incompatible(server_args: Any):
     memory-saver rejection in its own __init__; config-time rules can be
     added here as they're discovered.
     """
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     from sglang.srt.configs.model_config import is_deepseek_v4
@@ -331,7 +330,7 @@ def disable_prefill_cuda_graph_for_deepseek_trtllm_mla(server_args: Any):
     breakable) trtllm_mla falls back to FlashAttention for prefill and regresses
     performance, so disable whichever prefill graph backend is in effect.
     """
-    from sglang.srt.arg_groups.overrides import attention_backends_of, model_config_of
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     cfg = resolving_view(server_args)
 
@@ -403,7 +402,6 @@ def apply_inkling_prefill_cuda_graph_default(server_args: Any):
     auto-disabled for this multimodal arch, and declarative model overrides
     materialize too late to steer cuda-graph resolution. Honors an explicit
     --cuda-graph-backend-prefill / --disable-prefill-cuda-graph."""
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if (
@@ -425,7 +423,6 @@ def apply_inkling_prefill_cuda_graph_default(server_args: Any):
 
 
 def apply_muse_glimmer_prefill_cuda_graph_max_bs_default(server_args: Any):
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if (

python/sglang/srt/arg_groups/deepseek_v4_hook.py

@@ -4,8 +4,10 @@
 from typing import TYPE_CHECKING
 
 from sglang.srt.arg_groups.overrides import (
+    _deepseek_v4_kv_cache_dtype,
     declare_resolution,
     resolving_view,
+    run_post_process_pass,
 )
 from sglang.srt.environ import envs
 
@@ -129,10 +131,6 @@ def apply_deepseek_v4_defaults(server_args: ServerArgs, model_arch: str) -> None
     # The kv-cache dtype default moved to the resolution pipeline
     # (arg_groups/overrides.py: _deepseek_v4_kv_cache_dtype), invoked here at
     # its legacy slot.
-    from sglang.srt.arg_groups.overrides import (
-        _deepseek_v4_kv_cache_dtype,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _deepseek_v4_kv_cache_dtype)
 

python/sglang/srt/arg_groups/dllm_hook.py

@@ -7,8 +7,12 @@
 from typing import Any
 
 from sglang.srt.arg_groups.overrides import (
+    _dllm_attention_backend,
+    _dllm_overlap_disable,
+    _dllm_page_size,
     declare_resolution,
     resolving_view,
+    run_post_process_pass,
 )
 from sglang.srt.model_executor.cuda_graph_config import Backend, Phase, with_phase
 from sglang.srt.utils.common import is_hip
@@ -46,12 +50,6 @@ def handle_dllm_inference(server_args: Any):
                 ),
             )
 
-    from sglang.srt.arg_groups.overrides import (
-        _dllm_attention_backend,
-        _dllm_overlap_disable,
-        run_post_process_pass,
-    )
-
     run_post_process_pass(server_args, _dllm_attention_backend)
     run_post_process_pass(server_args, _dllm_overlap_disable)
 
@@ -60,7 +58,6 @@ def handle_dllm_inference(server_args: Any):
     # Invoked outside the radix gate: the alignment fill keeps its radix
     # gate inside the pass, the block-size cap applies regardless (it
     # replaces the unconditional scheduler-init fallback).
-    from sglang.srt.arg_groups.overrides import _dllm_page_size
 
     run_post_process_pass(server_args, _dllm_page_size)
 

python/sglang/srt/arg_groups/expert_pack_hook.py

@@ -9,7 +9,11 @@
 from pathlib import Path
 from typing import Any
 
-from sglang.srt.arg_groups.overrides import declare_resolution, resolving_view
+from sglang.srt.arg_groups.overrides import (
+    declare_resolution,
+    model_config_of,
+    resolving_view,
+)
 from sglang.srt.environ import envs
 from sglang.srt.model_executor.cuda_graph_config import (
     Backend,
@@ -27,7 +31,6 @@
 
 def handle_expert_pack(server_args: Any) -> None:
     """Normalize expert-pack settings and report all startup errors together."""
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if cfg.load_format != "expert_pack":

python/sglang/srt/arg_groups/hicache_hook.py

@@ -9,6 +9,7 @@
 from sglang.srt.arg_groups.overrides import (
     declare_resolution,
     resolving_view,
+    use_mla_backend,
 )
 
 logger = logging.getLogger(__name__)
@@ -68,7 +69,6 @@ def handle_hicache_ratio_default(server_args: Any):
 
 
 def resolve_hicache_dcp_compatibility(server_args: Any):
-    from sglang.srt.arg_groups.overrides import use_mla_backend
 
     cfg = resolving_view(server_args)
     if cfg.dcp_size <= 1 or not cfg.enable_hierarchical_cache:

python/sglang/srt/arg_groups/hisparse_hook.py

@@ -3,7 +3,11 @@
 import logging
 from typing import TYPE_CHECKING
 
-from sglang.srt.arg_groups.overrides import resolving_view
+from sglang.srt.arg_groups.overrides import (
+    model_config_of,
+    resolved_view,
+    resolving_view,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
@@ -46,7 +50,6 @@ def _hisparse_allowed_backends(kv_cache_dtype: str) -> set[str]:
 def validate_hisparse_dsa_backend(
     server_args: ServerArgs, attr: str, label: str
 ) -> None:
-    from sglang.srt.arg_groups.overrides import resolved_view
 
     # Invoked after the DSA kv-cache-dtype / split-backend declarations:
     # read the resolving state through the view.
@@ -65,7 +68,6 @@ def validate_hisparse_dsa_backend(
 
 
 def validate_hisparse_kv_cache_dtype(server_args: ServerArgs) -> None:
-    from sglang.srt.arg_groups.overrides import resolved_view
 
     kv_cache_dtype = resolved_view(server_args).kv_cache_dtype
     if kv_cache_dtype in HISPARSE_KV_CACHE_DTYPES:
@@ -82,7 +84,6 @@ def validate_hisparse_kv_cache_dtype(server_args: ServerArgs) -> None:
 
 def validate_hisparse(server_args: ServerArgs) -> None:
     """Validate --enable-hisparse constraints (model class, radix cache, DSA backend)."""
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if not cfg.enable_hisparse:
@@ -126,8 +127,6 @@ def validate_hisparse(server_args: ServerArgs) -> None:
             )
         return
 
-    from sglang.srt.arg_groups.overrides import resolved_view
-
     if resolved_view(server_args).kv_cache_dtype not in (
         "bfloat16",
         "auto",

python/sglang/srt/arg_groups/kv_cache_hook.py

@@ -10,6 +10,7 @@
     declare_resolution,
     resolved_view,
     resolving_view,
+    use_mla_backend,
 )
 from sglang.srt.environ import envs
 from sglang.srt.model_executor.cuda_graph_config import Backend
@@ -37,7 +38,7 @@ def handle_mxfp8_kv_cache_compatibility(server_args: Any) -> None:
 
 def handle_kv4_compatibility(server_args: Any) -> None:
     """Check FP4 KV cache compatibility with the attention backend"""
-    from sglang.srt.arg_groups.overrides import attention_backends_of, use_mla_backend
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     cfg = resolving_view(server_args)
 
@@ -279,7 +280,7 @@ def handle_page_major_kv_layout(server_args: Any):
     # The unified pool stores state in the page-major envelope-strided layout, so
     # enabling it implies --enable-page-major-kv-layout — routing it through the
     # single page-major path + stride-aware Triton asserts (set before the guard).
-    from sglang.srt.arg_groups.overrides import attention_backends_of, use_mla_backend
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     cfg = resolving_view(server_args)
     if cfg.enable_unified_memory:

python/sglang/srt/arg_groups/memory_hook.py

@@ -9,8 +9,11 @@
 
 from sglang.srt.arg_groups.overrides import (
     declare_resolution,
+    model_config_of,
+    post_capture_kv_sizing_planned,
     resolved_view,
     resolving_view,
+    use_mla_backend,
 )
 from sglang.srt.environ import envs
 from sglang.srt.model_executor.cuda_graph_config import Backend
@@ -47,11 +50,6 @@ def handle_gpu_memory_settings(server_args: Any, gpu_mem):
         generate_decode_cuda_graph_batch_sizes,
         generate_prefill_cuda_graph_batch_sizes,
     )
-    from sglang.srt.arg_groups.overrides import (
-        model_config_of,
-        post_capture_kv_sizing_planned,
-        use_mla_backend,
-    )
 
     cfg = resolving_view(server_args)
     # A copy, so an earlier declaration keeps the value it recorded.
@@ -278,7 +276,6 @@ def handle_gpu_memory_settings(server_args: Any, gpu_mem):
 
 
 def reserve_for_graph_mb(server_args: Any) -> float:
-    from sglang.srt.arg_groups.overrides import use_mla_backend
 
     cfg = resolving_view(server_args)
     decode_cuda_graph_config = cfg.cuda_graph_config.decode