From a03e45aadfbadceeb8c1568593afb6626a25bcab Mon Sep 17 00:00:00 2001
From: Cheng Wan <cheng.wan@radixark.ai>
Date: Sat, 29 Aug 2026 07:11:43 +0000
Subject: [PATCH] config: the lazy imports that buy nothing become eager

`arg_groups.overrides` is imported at module scope by thirty-five files
already, `server_args` among them, so importing it is not a cycle and not a
cost -- by the time any resolution step runs, the module is loaded. Ninety-three
`from ...overrides import` statements were nonetheless written inside the
functions that use them.

The ones that buy nothing become module-scope imports, including the twenty-three
that were **mixed**: a name a test patches, plus names riding along on the same
statement that nothing patches. Splitting those frees the riders -- the `_dsa_*`
pass callables are the shape of it, only ever handed to `run_post_process_pass`
as arguments, never looked up through the module.

Three groups were kept lazy for a test seam, which is the wrong way round, and
each is fixed at the test rather than in the code the runtime executes:

- The four declaration paths. `test_no_entry_changes_after_it_is_recorded`
  wrapped `declare_resolution`, `declare_late_resolution`,
  `declare_direct_writes` and `run_post_process_pass` by `setattr` on the
  module, purely to snapshot the stash after each append. The property it
  checks belongs to the stash, so the seam is now the stash: a list that
  snapshots on append, installed by a test-local `ServerArgs` subclass that
  survives the pipeline's own reset of it. The guard is strictly stronger --
  it records at the append rather than after the call -- and still fails on an
  injected in-place edit.
- `use_mla_backend`. Eight cases patched the answer; it is
  `model_config_of(server_args).attention_arch == AttentionArch.MLA`, so the
  fixtures now state the model instead. Five of the patches were redundant
  outright -- their stand-in already answered "not MLA" -- and the rest set
  `_model_config.attention_arch`, which is the input the value is derived from.
- `model_config_of` had no patch site at all and was lazy on a hunch.

What is left is load-bearing, for exactly two reasons and nothing else:
`attention_backends_of` (3 patch sites) and `supports_mamba_cache_extra_buffer`
(2) are still patched through the module, and `runtime_context` is imported by
`utils.common`, which `overrides` needs, so it sits *below* `overrides` and
cannot import it at module scope -- a cycle, not a preference.

A function-local `overrides` import that is neither of those: zero.

A dummy resolution is unchanged at ~9 ms and loads the same six `arg_groups`
modules; the 62-shape resolution probe is byte-identical to the base commit.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---
 .../sglang/srt/arg_groups/attention_hook.py   | 47 +++++-------
 .../sglang/srt/arg_groups/cuda_graph_hook.py  |  9 +--
 .../sglang/srt/arg_groups/deepseek_v4_hook.py |  6 +-
 python/sglang/srt/arg_groups/dllm_hook.py     | 11 ++-
 .../sglang/srt/arg_groups/expert_pack_hook.py |  7 +-
 python/sglang/srt/arg_groups/hicache_hook.py  |  2 +-
 python/sglang/srt/arg_groups/hisparse_hook.py | 11 ++-
 python/sglang/srt/arg_groups/kv_cache_hook.py |  5 +-
 python/sglang/srt/arg_groups/memory_hook.py   |  9 +--
 python/sglang/srt/arg_groups/model_hook.py    | 57 +++++----------
 .../sglang/srt/arg_groups/model_path_hook.py  |  6 +-
 python/sglang/srt/arg_groups/moe_hook.py      | 21 ++----
 python/sglang/srt/arg_groups/parallel_hook.py | 14 ++--
 .../srt/arg_groups/pd_disaggregation_hook.py  |  4 +-
 python/sglang/srt/arg_groups/serving_hook.py  |  2 +-
 .../sglang/srt/arg_groups/speculative_hook.py | 25 ++-----
 .../sglang/srt/arg_groups/validation_hook.py  |  6 +-
 python/sglang/srt/configs/model_config.py     |  2 +-
 python/sglang/srt/dllm/config.py              |  2 +-
 python/sglang/srt/entrypoints/grpc_server.py  |  3 +-
 .../sglang/srt/hardware_backend/npu/utils.py  |  7 +-
 .../layers/attention/attention_registry.py    |  6 +-
 python/sglang/srt/layers/cp/bcg.py            |  7 +-
 python/sglang/srt/layers/moe/kt_ep_wrapper.py |  3 +-
 .../srt/lora/marlin_lora_temp/policy.py       |  3 +-
 .../kv_pool_runtime.py                        |  2 +-
 .../srt/model_loader/expert_pack_runtime.py   | 11 ++-
 python/sglang/srt/runtime_context.py          |  4 +-
 python/sglang/srt/server_args.py              |  2 +-
 .../srt/speculative/adaptive_spec_params.py   |  6 +-
 .../dspark_components/dspark_config.py        |  2 +-
 python/sglang/srt/speculative/spec_info.py    |  2 +-
 .../sglang/srt/speculative/spec_registry.py   |  3 +-
 .../test_multimodal_piecewise_cuda_graph.py   | 11 ++-
 .../test_page_major_backend_allowlist.py      | 27 +++----
 .../test_resolution_declarations.py           | 71 +++++++++----------
 .../unit/server_args/test_server_args.py      | 29 ++------
 test/registered/unit/test_model_overrides.py  |  5 +-
 38 files changed, 176 insertions(+), 274 deletions(-)

diff --git a/python/sglang/srt/arg_groups/attention_hook.py b/python/sglang/srt/arg_groups/attention_hook.py
index df14894fac61..120724a14222 100644
--- a/python/sglang/srt/arg_groups/attention_hook.py
+++ b/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
@@ -29,11 +45,7 @@
 
 
 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)
@@ -41,16 +53,6 @@ def handle_attention_backend_compatibility(server_args: Any):
     # 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
diff --git a/python/sglang/srt/arg_groups/cuda_graph_hook.py b/python/sglang/srt/arg_groups/cuda_graph_hook.py
index 47ba48f76d72..9a958188d04f 100644
--- a/python/sglang/srt/arg_groups/cuda_graph_hook.py
+++ b/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 (
diff --git a/python/sglang/srt/arg_groups/deepseek_v4_hook.py b/python/sglang/srt/arg_groups/deepseek_v4_hook.py
index 8bf9800f8c00..ca96a1f0241e 100644
--- a/python/sglang/srt/arg_groups/deepseek_v4_hook.py
+++ b/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)
 
diff --git a/python/sglang/srt/arg_groups/dllm_hook.py b/python/sglang/srt/arg_groups/dllm_hook.py
index c590636026d0..bc3b631d4469 100644
--- a/python/sglang/srt/arg_groups/dllm_hook.py
+++ b/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)
 
diff --git a/python/sglang/srt/arg_groups/expert_pack_hook.py b/python/sglang/srt/arg_groups/expert_pack_hook.py
index 23a903f1fea9..ecb4aa5c90eb 100644
--- a/python/sglang/srt/arg_groups/expert_pack_hook.py
+++ b/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":
diff --git a/python/sglang/srt/arg_groups/hicache_hook.py b/python/sglang/srt/arg_groups/hicache_hook.py
index 28c1ddbb1b52..cd3c02df63b1 100644
--- a/python/sglang/srt/arg_groups/hicache_hook.py
+++ b/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:
diff --git a/python/sglang/srt/arg_groups/hisparse_hook.py b/python/sglang/srt/arg_groups/hisparse_hook.py
index e3754ae4b18e..e94a0510c16e 100644
--- a/python/sglang/srt/arg_groups/hisparse_hook.py
+++ b/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",
diff --git a/python/sglang/srt/arg_groups/kv_cache_hook.py b/python/sglang/srt/arg_groups/kv_cache_hook.py
index d8a65ad9df29..1a58fbe87004 100644
--- a/python/sglang/srt/arg_groups/kv_cache_hook.py
+++ b/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:
diff --git a/python/sglang/srt/arg_groups/memory_hook.py b/python/sglang/srt/arg_groups/memory_hook.py
index d060fca3f9c6..65ec050bfff8 100644
--- a/python/sglang/srt/arg_groups/memory_hook.py
+++ b/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
diff --git a/python/sglang/srt/arg_groups/model_hook.py b/python/sglang/srt/arg_groups/model_hook.py
index da135d50bd01..4db9df2528d4 100644
--- a/python/sglang/srt/arg_groups/model_hook.py
+++ b/python/sglang/srt/arg_groups/model_hook.py
@@ -7,10 +7,25 @@
 from typing import Any
 
 from sglang.srt.arg_groups.overrides import (
+    _deepseek_moe_quant_resolution,
+    _deepseek_spec_moe_resolution,
+    _dsa_kv_cache_dtype_default,
+    _dsa_split_backend_resolution,
+    _enforce_disable_allreduce_fusion,
+    _flashinfer_allreduce_fusion_auto_enable,
+    _hrm_text_attention_force,
+    _mamba_radix_cache_resolution,
+    _sparse_head_overlap_disable,
+    collect_model_override_declarations,
     declare_resolution,
     mamba_cache_chunk_size,
+    mamba_extra_buffer_of,
+    model_config_of,
     resolved_view,
     resolving_view,
+    run_post_process_pass,
+    use_mla_backend,
+    validate_declarations,
 )
 from sglang.srt.configs.embedding_model_spec import BCGPrefillPolicy
 from sglang.srt.configs.linear_attn_model_registry import get_linear_attn_spec_by_arch
@@ -35,11 +50,7 @@
 
 
 def handle_model_specific_adjustments(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)
     from sglang.srt.configs.model_config import (
@@ -107,10 +118,6 @@ def handle_model_specific_adjustments(server_args: Any):
     # server_args is never mutated — mid-resolution readers see the
     # declared values through resolved_view, runtime readers through the
     # flags tier.
-    from sglang.srt.arg_groups.overrides import (
-        collect_model_override_declarations,
-        validate_declarations,
-    )
 
     model_overrides = collect_model_override_declarations(
         model_arch, server_args, hf_config
@@ -210,11 +217,6 @@ def handle_model_specific_adjustments(server_args: Any):
                 import torch
 
                 major, _ = torch.cuda.get_device_capability()
-                from sglang.srt.arg_groups.overrides import (
-                    _dsa_kv_cache_dtype_default,
-                    _dsa_split_backend_resolution,
-                    run_post_process_pass,
-                )
 
                 run_post_process_pass(server_args, _dsa_kv_cache_dtype_default)
                 run_post_process_pass(server_args, _dsa_split_backend_resolution)
@@ -298,10 +300,6 @@ def handle_model_specific_adjustments(server_args: Any):
         # kv-cache-dtype default above must read the pristine
         # quantization). The HIP arm (fusion log + spec_moe writes, the
         # latter awaiting the speculative-hook migration) stays below.
-        from sglang.srt.arg_groups.overrides import (
-            _deepseek_moe_quant_resolution,
-            run_post_process_pass,
-        )
 
         run_post_process_pass(server_args, _deepseek_moe_quant_resolution)
         if is_hip():
@@ -324,9 +322,6 @@ def handle_model_specific_adjustments(server_args: Any):
             # resolution pipeline (arg_groups/overrides.py:
             # _deepseek_spec_moe_resolution), invoked here at its legacy
             # slot.
-            from sglang.srt.arg_groups.overrides import (
-                _deepseek_spec_moe_resolution,
-            )
 
             run_post_process_pass(server_args, _deepseek_spec_moe_resolution)
 
@@ -576,11 +571,6 @@ def handle_model_specific_adjustments(server_args: Any):
     # resolved before that tail write of disable_overlap_schedule.
     handle_mamba_radix_cache(server_args, model_arch)
 
-    from sglang.srt.arg_groups.overrides import (
-        _sparse_head_overlap_disable,
-        run_post_process_pass,
-    )
-
     run_post_process_pass(server_args, _sparse_head_overlap_disable)
 
     # The FlashInfer AllReduce Fusion auto-enable and the enforce-disable
@@ -588,10 +578,6 @@ def handle_model_specific_adjustments(server_args: Any):
     # _flashinfer_allreduce_fusion_auto_enable /
     # _enforce_disable_allreduce_fusion), invoked here at their legacy
     # slots.
-    from sglang.srt.arg_groups.overrides import (
-        _enforce_disable_allreduce_fusion,
-        _flashinfer_allreduce_fusion_auto_enable,
-    )
 
     run_post_process_pass(server_args, _flashinfer_allreduce_fusion_auto_enable)
     run_post_process_pass(server_args, _enforce_disable_allreduce_fusion)
@@ -604,15 +590,10 @@ def handle_model_capability_adjustments(server_args: Any):
     from sglang.srt.arg_groups.kv_cache_hook import (
         validate_prefill_only_disable_kv_cache_args,
     )
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if parse_connector_type(cfg.model_path) == ConnectorType.INSTANCE:
         return
-    from sglang.srt.arg_groups.overrides import (
-        _hrm_text_attention_force,
-        run_post_process_pass,
-    )
 
     model_config = model_config_of(server_args)
     hf_config = model_config.hf_config
@@ -833,11 +814,6 @@ def handle_mamba_radix_cache(server_args: Any, model_arch: str):
         validate_mamba_extra_buffer,
         validate_mamba_no_buffer,
     )
-    from sglang.srt.arg_groups.overrides import (
-        _mamba_radix_cache_resolution,
-        mamba_extra_buffer_of,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _mamba_radix_cache_resolution)
     view = resolved_view(server_args)
@@ -855,7 +831,6 @@ def handle_mamba_radix_cache(server_args: Any, model_arch: str):
 
 
 def handle_language_model_only(server_args: Any):
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if not cfg.language_model_only:
diff --git a/python/sglang/srt/arg_groups/model_path_hook.py b/python/sglang/srt/arg_groups/model_path_hook.py
index 083fe81e0bfb..0093fac31515 100644
--- a/python/sglang/srt/arg_groups/model_path_hook.py
+++ b/python/sglang/srt/arg_groups/model_path_hook.py
@@ -10,8 +10,10 @@
 from typing import Any, Optional
 
 from sglang.srt.arg_groups.overrides import (
+    _gguf_quantization,
     declare_resolution,
     resolving_view,
+    run_post_process_pass,
 )
 from sglang.srt.utils.common import is_remote_url
 from sglang.srt.utils.hf_transformers_utils import check_gguf_file
@@ -163,10 +165,6 @@ def handle_load_format(server_args: Any):
     # (arg_groups/overrides.py: _gguf_quantization); load_format itself is
     # genuine config (runtime user updates write it) and stays imperative.
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import (
-        _gguf_quantization,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _gguf_quantization)
     if (cfg.load_format == "auto" or cfg.load_format == "gguf") and check_gguf_file(
diff --git a/python/sglang/srt/arg_groups/moe_hook.py b/python/sglang/srt/arg_groups/moe_hook.py
index 792bee10ada8..15f48159cd60 100644
--- a/python/sglang/srt/arg_groups/moe_hook.py
+++ b/python/sglang/srt/arg_groups/moe_hook.py
@@ -8,12 +8,19 @@
 from typing import Any
 
 from sglang.srt.arg_groups.overrides import (
+    _a2a_backend_overrides,
+    _a2a_ep_size,
+    _a2a_fusion_adjustments,
+    _moe_runner_backend_quant_constraints,
+    _moe_runner_fusion_disable,
     cutedsl_moe_max_num_tokens,
     declare_resolution,
     max_prefill_buffer_tokens,
     max_speculative_num_draft_tokens,
+    model_config_of,
     resolved_view,
     resolving_view,
+    run_post_process_pass,
 )
 from sglang.srt.connector import ConnectorType
 from sglang.srt.environ import envs
@@ -27,14 +34,8 @@ def handle_moe_kernel_config(server_args: Any):
     # The quantization-driven runner resolutions moved to the pipeline
     # (arg_groups/overrides.py: _moe_runner_backend_quant_constraints);
     # the compatibility asserts and fusion writes stay below.
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import (
-        _moe_runner_backend_quant_constraints,
-        _moe_runner_fusion_disable,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _moe_runner_backend_quant_constraints)
 
@@ -121,15 +122,8 @@ def handle_a2a_moe(server_args: Any):
     # the resolution pipeline (arg_groups/overrides.py:
     # _a2a_backend_overrides / _a2a_ep_size); the per-backend logs,
     # asserts, fusion/deepep_mode/env/cuda-graph writes stay below.
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import (
-        _a2a_backend_overrides,
-        _a2a_ep_size,
-        _a2a_fusion_adjustments,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _a2a_backend_overrides)
     run_post_process_pass(server_args, _a2a_ep_size)
@@ -420,7 +414,6 @@ def validate_deepep_v2_dispatch_token_budget(server_args: Any) -> None:
 
 def validate_deepep_v2_model_architecture(server_args: Any) -> None:
     """Allow DeepEP v2 only where its model workflow is validated."""
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     if (
         parse_connector_type(resolved_view(server_args).model_path)
diff --git a/python/sglang/srt/arg_groups/parallel_hook.py b/python/sglang/srt/arg_groups/parallel_hook.py
index 1def83629be1..31b19ee99574 100644
--- a/python/sglang/srt/arg_groups/parallel_hook.py
+++ b/python/sglang/srt/arg_groups/parallel_hook.py
@@ -8,9 +8,14 @@
 from typing import Any
 
 from sglang.srt.arg_groups.overrides import (
+    _data_parallelism_defaults,
+    _dp_lm_head_validation,
+    _tp_lm_head_all_to_all_default,
     declare_resolution,
+    model_config_of,
     resolved_view,
     resolving_view,
+    run_post_process_pass,
     should_report_expert_balancedness,
 )
 from sglang.srt.connector import ConnectorType
@@ -22,7 +27,6 @@
 
 
 def handle_context_parallelism(server_args: Any):
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     if parse_connector_type(cfg.model_path) != ConnectorType.INSTANCE:
@@ -162,10 +166,6 @@ def handle_data_parallelism(server_args: Any):
     )
 
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import (
-        _data_parallelism_defaults,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _data_parallelism_defaults)
 
@@ -233,10 +233,6 @@ def handle_data_parallelism(server_args: Any):
 
     # Resolve the phase-aware TP LM-head default before validating the
     # resulting DP/TP LM-head configuration.
-    from sglang.srt.arg_groups.overrides import (
-        _dp_lm_head_validation,
-        _tp_lm_head_all_to_all_default,
-    )
 
     run_post_process_pass(server_args, _tp_lm_head_all_to_all_default)
     run_post_process_pass(server_args, _dp_lm_head_validation)
diff --git a/python/sglang/srt/arg_groups/pd_disaggregation_hook.py b/python/sglang/srt/arg_groups/pd_disaggregation_hook.py
index 456827dec28f..78bf2f755174 100644
--- a/python/sglang/srt/arg_groups/pd_disaggregation_hook.py
+++ b/python/sglang/srt/arg_groups/pd_disaggregation_hook.py
@@ -7,6 +7,8 @@
 
 from sglang.srt.arg_groups.overrides import (
     declare_resolution,
+    model_config_of,
+    resolved_view,
     resolving_view,
 )
 from sglang.srt.environ import envs
@@ -89,7 +91,6 @@ def handle_pd_disaggregation(server_args: ServerArgs) -> None:
                     "with speculative decoding "
                     f"(--speculative-algorithm {cfg.speculative_algorithm})"
                 )
-            from sglang.srt.arg_groups.overrides import resolved_view
 
             if resolved_view(server_args).enable_dp_attention:
                 logger.warning(
@@ -186,7 +187,6 @@ def _alias_bootstrap_port_to_api_port(server_args: ServerArgs) -> None:
 
 def handle_encoder_disaggregation(server_args: Any):
     from sglang.srt.arg_groups.model_hook import handle_language_model_only
-    from sglang.srt.arg_groups.overrides import model_config_of
     from sglang.srt.arg_groups.validation_hook import validate_ib_devices
     from sglang.srt.server_args import resolve_encoder_transfer_backend
 
diff --git a/python/sglang/srt/arg_groups/serving_hook.py b/python/sglang/srt/arg_groups/serving_hook.py
index 416a89285029..8219156cae0b 100644
--- a/python/sglang/srt/arg_groups/serving_hook.py
+++ b/python/sglang/srt/arg_groups/serving_hook.py
@@ -12,6 +12,7 @@
 
 from sglang.srt.arg_groups.overrides import (
     declare_resolution,
+    model_config_of,
     resolved_view,
     resolving_view,
 )
@@ -751,7 +752,6 @@ def handle_multimodal_feature_transport(server_args: Any):
     may still auto-select CUDA VMM. The legacy CUDA IPC flag and environment
     variable remain supported so existing deployments map to this policy.
     """
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
     requested_transport = cfg.mm_feature_transport
diff --git a/python/sglang/srt/arg_groups/speculative_hook.py b/python/sglang/srt/arg_groups/speculative_hook.py
index bd20e21f0d1a..8a2650b63119 100644
--- a/python/sglang/srt/arg_groups/speculative_hook.py
+++ b/python/sglang/srt/arg_groups/speculative_hook.py
@@ -6,9 +6,13 @@
 from typing import TYPE_CHECKING, Optional
 
 from sglang.srt.arg_groups.overrides import (
+    _speculative_moe_runner_default,
     declare_direct_writes,
     declare_resolution,
+    model_config_of,
+    resolved_view,
     resolving_view,
+    run_post_process_pass,
 )
 
 if TYPE_CHECKING:
@@ -86,10 +90,6 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None:
 
     # Moved to the resolution pipeline (arg_groups/overrides.py:
     # _speculative_moe_runner_default), invoked here at its legacy slot.
-    from sglang.srt.arg_groups.overrides import (
-        _speculative_moe_runner_default,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _speculative_moe_runner_default)
 
@@ -184,7 +184,6 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None:
 
 def _handle_dflash(server_args: ServerArgs) -> None:
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import resolved_view
 
     if not (cfg.device.startswith("cuda") or cfg.device == "npu"):
         raise ValueError(
@@ -337,7 +336,6 @@ def _handle_dflash(server_args: ServerArgs) -> None:
 
 
 def _target_checkpoint_bundles_dspark_draft(server_args: ServerArgs) -> bool:
-    from sglang.srt.arg_groups.overrides import model_config_of
     from sglang.srt.speculative.dspark_components.dspark_config import (
         checkpoint_bundles_dspark_draft,
     )
@@ -580,10 +578,7 @@ def _resolve_dflash_draft_attention_backend(server_args: ServerArgs) -> None:
 
     draft_backend = cfg.speculative_draft_attention_backend
     if draft_backend is None:
-        from sglang.srt.arg_groups.overrides import (
-            attention_backends_of,
-            resolved_view,
-        )
+        from sglang.srt.arg_groups.overrides import attention_backends_of
 
         draft_backend, _ = attention_backends_of(resolved_view(server_args))
     if draft_backend is None:
@@ -663,13 +658,9 @@ def _handle_frozen_kv_mtp(server_args: ServerArgs) -> None:
 
 
 def _handle_eagle_family(server_args: ServerArgs) -> None:
-    from sglang.srt.arg_groups.overrides import model_config_of
 
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import (
-        attention_backends_of,
-        resolved_view,
-    )
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     if (
         cfg.speculative_algorithm == "STANDALONE"
@@ -799,8 +790,6 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
                 "coins from the global RNG and is not batch-invariant."
             )
 
-        from sglang.srt.arg_groups.overrides import resolved_view
-
         if (
             resolved_view(server_args).enable_multi_layer_eagle
             and cfg.speculative_eagle_topk != 1
@@ -915,8 +904,6 @@ def _handle_ngram(server_args: ServerArgs) -> None:
         "using ngram speculative decoding."
     )
 
-    from sglang.srt.arg_groups.overrides import resolved_view
-
     view = resolved_view(server_args)
     if (
         cfg.speculative_eagle_topk > 1
diff --git a/python/sglang/srt/arg_groups/validation_hook.py b/python/sglang/srt/arg_groups/validation_hook.py
index c9fcb92427e4..6364f7ec9221 100644
--- a/python/sglang/srt/arg_groups/validation_hook.py
+++ b/python/sglang/srt/arg_groups/validation_hook.py
@@ -9,8 +9,10 @@
 from typing import Any, Dict, List, Optional
 
 from sglang.srt.arg_groups.overrides import (
+    _hisparse_validation,
     resolved_view,
     resolving_view,
+    run_post_process_pass,
 )
 from sglang.srt.distributed.device_communicators.mooncake_transfer_engine import (
     parse_ib_device_config,
@@ -154,10 +156,6 @@ def check_server_args(server_args: Any):
     # Check hisparse
     # Moved to the resolution pipeline (arg_groups/overrides.py:
     # _hisparse_validation), invoked here at its legacy slot.
-    from sglang.srt.arg_groups.overrides import (
-        _hisparse_validation,
-        run_post_process_pass,
-    )
 
     run_post_process_pass(server_args, _hisparse_validation)
 
diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py
index 952dda79b17f..32ea0fd3c96f 100644
--- a/python/sglang/srt/configs/model_config.py
+++ b/python/sglang/srt/configs/model_config.py
@@ -25,6 +25,7 @@
 import torch
 from transformers import PretrainedConfig
 
+from sglang.srt.arg_groups.overrides import resolving_view
 from sglang.srt.configs.embedding_model_spec import resolve_embedding_model_spec
 from sglang.srt.configs.linear_attn_model_registry import get_linear_attn_config
 from sglang.srt.environ import envs
@@ -604,7 +605,6 @@ def from_server_args(
         context_length: Optional[int] = None,
         **kwargs,
     ):
-        from sglang.srt.arg_groups.overrides import resolving_view
 
         cfg = resolving_view(server_args)
         quantization = (
diff --git a/python/sglang/srt/dllm/config.py b/python/sglang/srt/dllm/config.py
index a2206feb92a9..ac942c44b8c4 100644
--- a/python/sglang/srt/dllm/config.py
+++ b/python/sglang/srt/dllm/config.py
@@ -1,5 +1,6 @@
 from typing import Any
 
+from sglang.srt.arg_groups.overrides import resolving_view
 from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.server_args import ServerArgs
 
@@ -25,7 +26,6 @@ def __init__(
     def from_server_args(
         server_args: ServerArgs,
     ):
-        from sglang.srt.arg_groups.overrides import resolving_view
 
         cfg = resolving_view(server_args)
         if cfg.dllm_algorithm is None:
diff --git a/python/sglang/srt/entrypoints/grpc_server.py b/python/sglang/srt/entrypoints/grpc_server.py
index 0c2a8c559316..3c7fba8ea98e 100644
--- a/python/sglang/srt/entrypoints/grpc_server.py
+++ b/python/sglang/srt/entrypoints/grpc_server.py
@@ -17,6 +17,7 @@
 
 from aiohttp import web
 
+from sglang.srt.arg_groups.overrides import resolving_view
 from sglang.srt.managers.io_struct import ProfileReq, ProfileReqType
 from sglang.srt.utils.common import get_bool_env_var
 
@@ -165,8 +166,6 @@ async def serve_grpc(server_args, model_info=None):
             "version mismatch — see the chained exception above for details."
         ) from e
 
-    from sglang.srt.arg_groups.overrides import resolving_view
-
     # The integrated servicer builds an `Engine`, which validates and publishes
     # on its own. Validating here would run `check_server_args` twice, and the
     # LoRA normalization is not idempotent -- the second pass sees the `LoRARef`
diff --git a/python/sglang/srt/hardware_backend/npu/utils.py b/python/sglang/srt/hardware_backend/npu/utils.py
index df4fc2920f90..3fd179443620 100644
--- a/python/sglang/srt/hardware_backend/npu/utils.py
+++ b/python/sglang/srt/hardware_backend/npu/utils.py
@@ -6,7 +6,11 @@
 
 import torch
 
-from sglang.srt.arg_groups.overrides import declare_resolution
+from sglang.srt.arg_groups.overrides import (
+    declare_resolution,
+    resolving_view,
+    use_mla_backend,
+)
 from sglang.srt.environ import envs
 from sglang.srt.model_executor.cuda_graph_config import Phase, with_phase
 from sglang.srt.utils import get_npu_memory_capacity, is_npu
@@ -44,7 +48,6 @@ def set_default_server_args(args: "ServerArgs"):
     """
     Set default server arguments for NPU backend.
     """
-    from sglang.srt.arg_groups.overrides import resolving_view, use_mla_backend
 
     cfg = resolving_view(args)
 
diff --git a/python/sglang/srt/layers/attention/attention_registry.py b/python/sglang/srt/layers/attention/attention_registry.py
index 8f7b178c258b..a0615766dfed 100644
--- a/python/sglang/srt/layers/attention/attention_registry.py
+++ b/python/sglang/srt/layers/attention/attention_registry.py
@@ -2,6 +2,7 @@
 import warnings
 from typing import TYPE_CHECKING
 
+from sglang.srt.arg_groups.overrides import resolved_view
 from sglang.srt.configs.hybrid_arch import (
     hybrid_gdn_config,
     hybrid_lightning_config,
@@ -74,10 +75,7 @@ def create_trtllm_mla_backend(runner):
     if not runner.use_mla_backend:
         raise ValueError("trtllm_mla backend can only be used with MLA models.")
     if get_parallel().dcp_enabled and get_spec().speculative_algorithm is not None:
-        from sglang.srt.arg_groups.overrides import (
-            attention_backends_of,
-            resolved_view,
-        )
+        from sglang.srt.arg_groups.overrides import attention_backends_of
 
         _, decode_backend = attention_backends_of(resolved_view(runner.server_args))
         if decode_backend == "trtllm_mla":
diff --git a/python/sglang/srt/layers/cp/bcg.py b/python/sglang/srt/layers/cp/bcg.py
index df8f74f607d3..6c26099dd9b9 100644
--- a/python/sglang/srt/layers/cp/bcg.py
+++ b/python/sglang/srt/layers/cp/bcg.py
@@ -21,7 +21,10 @@
 
 import torch
 
-from sglang.srt.arg_groups.overrides import resolved_view
+from sglang.srt.arg_groups.overrides import (
+    resolved_view,
+    resolving_view,
+)
 from sglang.srt.layers.cp.base import get_cp_strategy
 from sglang.srt.layers.cp.padding import get_cp_padding_align_size
 from sglang.srt.layers.cp.utils import (
@@ -43,7 +46,7 @@
 
 def supports_prefill_cp_bcg(server_args: ServerArgs) -> bool:
     """Return whether the selected prefill-CP configuration supports BCG."""
-    from sglang.srt.arg_groups.overrides import attention_backends_of, resolving_view
+    from sglang.srt.arg_groups.overrides import attention_backends_of
 
     cfg = resolving_view(server_args)
     resolved = resolved_view(server_args)
diff --git a/python/sglang/srt/layers/moe/kt_ep_wrapper.py b/python/sglang/srt/layers/moe/kt_ep_wrapper.py
index 17c55559520a..8eb3d0fc89e3 100644
--- a/python/sglang/srt/layers/moe/kt_ep_wrapper.py
+++ b/python/sglang/srt/layers/moe/kt_ep_wrapper.py
@@ -12,6 +12,7 @@
 
 import torch
 
+from sglang.srt.arg_groups.overrides import model_config_of
 from sglang.srt.layers.quantization.base_config import FusedMoEMethodBase
 from sglang.srt.runtime_context import (
     get_exec,
@@ -77,8 +78,6 @@ def create_kt_config_from_server_args(
     if get_exec().moe.kt_weight_path is None:
         return None
 
-    from sglang.srt.arg_groups.overrides import model_config_of
-
     num_layers = getattr(
         model_config_of(server_args).hf_config, "num_hidden_layers", None
     )
diff --git a/python/sglang/srt/lora/marlin_lora_temp/policy.py b/python/sglang/srt/lora/marlin_lora_temp/policy.py
index c05838fa1365..eac576ee5cdc 100644
--- a/python/sglang/srt/lora/marlin_lora_temp/policy.py
+++ b/python/sglang/srt/lora/marlin_lora_temp/policy.py
@@ -9,12 +9,13 @@
 
 from typing import Any
 
+from sglang.srt.arg_groups.overrides import resolving_view
+
 
 def validate_experimental_sgl_marlin_server_args(
     server_args: Any, resolved_args: Any
 ) -> None:
     """Validate startup options before the experimental runner is constructed."""
-    from sglang.srt.arg_groups.overrides import resolving_view
 
     cfg = resolving_view(server_args)
 
diff --git a/python/sglang/srt/model_executor/model_runner_components/kv_pool_runtime.py b/python/sglang/srt/model_executor/model_runner_components/kv_pool_runtime.py
index 4b702e1518df..ac1c61418da0 100644
--- a/python/sglang/srt/model_executor/model_runner_components/kv_pool_runtime.py
+++ b/python/sglang/srt/model_executor/model_runner_components/kv_pool_runtime.py
@@ -6,6 +6,7 @@
 import msgspec
 import torch
 
+from sglang.srt.arg_groups.overrides import post_capture_kv_sizing_planned
 from sglang.srt.configs.hybrid_arch import mambaish_config
 from sglang.srt.distributed import get_world_group
 from sglang.srt.mem_cache.kv_cache_configurator import mm_runtime_reservation_gb
@@ -29,7 +30,6 @@
 def is_post_capture_kv_active(
     *, server_args: ServerArgs, is_draft_worker: bool
 ) -> bool:
-    from sglang.srt.arg_groups.overrides import post_capture_kv_sizing_planned
 
     return (
         post_capture_kv_sizing_planned(server_args)
diff --git a/python/sglang/srt/model_loader/expert_pack_runtime.py b/python/sglang/srt/model_loader/expert_pack_runtime.py
index 6e5e4c035ee3..5f5ea5b81047 100644
--- a/python/sglang/srt/model_loader/expert_pack_runtime.py
+++ b/python/sglang/srt/model_loader/expert_pack_runtime.py
@@ -15,7 +15,10 @@
 from pathlib import Path
 from typing import Any
 
-from sglang.srt.arg_groups.overrides import declare_resolution
+from sglang.srt.arg_groups.overrides import (
+    declare_resolution,
+    resolving_view,
+)
 
 METADATA_FORMAT_VERSION = 3
 GGUF_SHARD_SUFFIX_RE = re.compile(r"-\d{5}-of-\d{5}\.gguf$")
@@ -204,8 +207,6 @@ def prepare_raw_kimi_server_args(
 ) -> None:
     """Resolve a raw GGUF model path into the normal loader inputs."""
 
-    from sglang.srt.arg_groups.overrides import resolving_view
-
     cfg = resolving_view(server_args)
     model_path = Path(cfg.model_path).expanduser()
     if not model_path.is_file() or model_path.suffix.lower() != ".gguf":
@@ -516,8 +517,6 @@ def prepare_raw_deepseek_server_args(
 ) -> None:
     """Resolve a raw DeepSeek V4 GGUF into metadata and Expert Pack inputs."""
 
-    from sglang.srt.arg_groups.overrides import resolving_view
-
     cfg = resolving_view(server_args)
     source = Path(cfg.model_path).expanduser().resolve(strict=True)
     if not source.is_file():
@@ -564,8 +563,6 @@ def prepare_raw_expert_pack_server_args(
 ) -> None:
     """Dispatch a raw GGUF to the model-specific expert-pack preparation path."""
 
-    from sglang.srt.arg_groups.overrides import resolving_view
-
     cfg = resolving_view(server_args)
     source = Path(cfg.model_path).expanduser()
     if not source.is_file():
diff --git a/python/sglang/srt/runtime_context.py b/python/sglang/srt/runtime_context.py
index 02e331178b43..eaf7efab873b 100644
--- a/python/sglang/srt/runtime_context.py
+++ b/python/sglang/srt/runtime_context.py
@@ -864,12 +864,12 @@ def get_stream(self, name: str) -> Any:
         name (the keyed-lazy pattern of the persistent buffers). Creation is
         a driver call that must stay outside cuda-graph capture — call sites
         lease their stream at init/warmup time."""
+        from sglang.srt.arg_groups.overrides import resolution_result
+
         stream = self.resources.streams.get(name)
         if stream is None:
             import torch
 
-            from sglang.srt.arg_groups.overrides import resolution_result
-
             device = (
                 resolution_result(self._server_args, "device")
                 if self._server_args
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 009184acbf2d..7e1102e29ea3 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -56,6 +56,7 @@
     mamba_extra_buffer_lazy_of,
     mamba_extra_buffer_of,
     remote_instance_transfer_engine_of,
+    resolution_projection,
     resolving_view,
 )
 from sglang.srt.environ import envs
@@ -3697,7 +3698,6 @@ def resolved_dict(self) -> Dict[str, Any]:
         the way `asdict` expands them; the private resolution bookkeeping and the
         `model_config` memo are not fields and do not appear.
         """
-        from sglang.srt.arg_groups.overrides import resolution_projection
 
         return resolution_projection(self)
 
diff --git a/python/sglang/srt/speculative/adaptive_spec_params.py b/python/sglang/srt/speculative/adaptive_spec_params.py
index c3c252f82310..96997c1ffbf1 100644
--- a/python/sglang/srt/speculative/adaptive_spec_params.py
+++ b/python/sglang/srt/speculative/adaptive_spec_params.py
@@ -12,6 +12,10 @@
 from functools import cached_property
 from typing import TYPE_CHECKING
 
+from sglang.srt.arg_groups.overrides import (
+    resolved_view,
+    resolving_view,
+)
 from sglang.srt.utils import log_info_on_rank0
 
 if TYPE_CHECKING:
@@ -49,10 +53,8 @@
 
 def adaptive_unsupported_reason(server_args: ServerArgs) -> str | None:
     """Return why adaptive spec cannot run under the given server args, or None if supported."""
-    from sglang.srt.arg_groups.overrides import resolving_view
 
     cfg = resolving_view(server_args)
-    from sglang.srt.arg_groups.overrides import resolved_view
 
     if cfg.speculative_algorithm not in ("EAGLE", "EAGLE3"):
         return (
diff --git a/python/sglang/srt/speculative/dspark_components/dspark_config.py b/python/sglang/srt/speculative/dspark_components/dspark_config.py
index 88bc14405b3b..50909f9f7ae9 100644
--- a/python/sglang/srt/speculative/dspark_components/dspark_config.py
+++ b/python/sglang/srt/speculative/dspark_components/dspark_config.py
@@ -6,6 +6,7 @@
 
 import msgspec
 
+from sglang.srt.arg_groups.overrides import resolved_view
 from sglang.srt.runtime_context import (
     get_model,
     get_spec,
@@ -142,7 +143,6 @@ def read_draft_checkpoint_config(*, server_args: ServerArgs) -> DSparkDraftConfi
     silently drops the checkpoint's gamma and the cross-check with
     `--speculative-num-draft-tokens` along with it.
     """
-    from sglang.srt.arg_groups.overrides import resolved_view
     from sglang.srt.utils.hf_transformers_utils import get_config
 
     resolving = resolved_view(server_args)
diff --git a/python/sglang/srt/speculative/spec_info.py b/python/sglang/srt/speculative/spec_info.py
index 97ca92d31f0f..c2c494ed82fb 100644
--- a/python/sglang/srt/speculative/spec_info.py
+++ b/python/sglang/srt/speculative/spec_info.py
@@ -7,6 +7,7 @@
 
 import torch
 
+from sglang.srt.arg_groups.overrides import resolving_view
 from sglang.srt.runtime_context import get_spec as get_spec_config
 from sglang.srt.speculative.spec_registry import (
     CustomSpecAlgo,
@@ -254,7 +255,6 @@ def get_num_tokens_per_bs_for_target_verify(
     def create_worker(
         self, server_args: ServerArgs
     ) -> Optional[Union[Type[BaseSpecWorker], Type[TpModelWorker], Type[NGRAMWorker]]]:
-        from sglang.srt.arg_groups.overrides import resolving_view
 
         cfg = resolving_view(server_args)
         assert (
diff --git a/python/sglang/srt/speculative/spec_registry.py b/python/sglang/srt/speculative/spec_registry.py
index 933d4e05d7b0..20c0902bf88f 100644
--- a/python/sglang/srt/speculative/spec_registry.py
+++ b/python/sglang/srt/speculative/spec_registry.py
@@ -10,6 +10,8 @@
 
 import torch
 
+from sglang.srt.arg_groups.overrides import resolving_view
+
 if TYPE_CHECKING:
     from sglang.srt.managers.overlap_utils import FutureMap
     from sglang.srt.managers.schedule_batch import ScheduleBatch
@@ -108,7 +110,6 @@ def handle_server_args(self, server_args: ServerArgs) -> None:
         pass
 
     def create_worker(self, server_args: ServerArgs) -> Type:
-        from sglang.srt.arg_groups.overrides import resolving_view
 
         cfg = resolving_view(server_args)
         if not cfg.disable_overlap_schedule and not self.supports_overlap:
diff --git a/test/registered/unit/configs/test_multimodal_piecewise_cuda_graph.py b/test/registered/unit/configs/test_multimodal_piecewise_cuda_graph.py
index 98e9003b92f0..7f016d5fcf5d 100644
--- a/test/registered/unit/configs/test_multimodal_piecewise_cuda_graph.py
+++ b/test/registered/unit/configs/test_multimodal_piecewise_cuda_graph.py
@@ -9,6 +9,7 @@
 from sglang.srt.arg_groups.overrides import resolution_result
 from sglang.srt.configs.embedding_model_spec import resolve_embedding_model_spec
 from sglang.srt.configs.model_config import (
+    AttentionArch,
     is_multimodal_piecewise_cuda_graph_supported,
 )
 from sglang.srt.model_executor.cuda_graph_config import (
@@ -109,6 +110,7 @@ def test_trtllm_mla_stays_on_breakable(self):
             is_multimodal_piecewise_cuda_graph_supported=True,
             is_multimodal=False,
             is_multimodal_breakable_cuda_graph_supported=False,
+            attention_arch=AttentionArch.MLA,
             hf_config=SimpleNamespace(architectures=["DeepseekV2ForCausalLM"]),
         )
         args.cuda_graph_config = CudaGraphConfig(
@@ -116,12 +118,9 @@ def test_trtllm_mla_stays_on_breakable(self):
         )
         args._cuda_graph_config_locked = set()
 
-        with (
-            patch(
-                "sglang.srt.arg_groups.overrides.attention_backends_of",
-                return_value=("trtllm_mla", "trtllm_mla"),
-            ),
-            patch("sglang.srt.arg_groups.overrides.use_mla_backend", return_value=True),
+        with patch(
+            "sglang.srt.arg_groups.overrides.attention_backends_of",
+            return_value=("trtllm_mla", "trtllm_mla"),
         ):
             apply_cuda_graph_compatibility(args)
 
diff --git a/test/registered/unit/server_args/test_page_major_backend_allowlist.py b/test/registered/unit/server_args/test_page_major_backend_allowlist.py
index 3b185722010f..5cf9e3717cb5 100644
--- a/test/registered/unit/server_args/test_page_major_backend_allowlist.py
+++ b/test/registered/unit/server_args/test_page_major_backend_allowlist.py
@@ -30,9 +30,10 @@
 """
 
 import unittest
-from unittest import mock
+from types import SimpleNamespace
 
 from sglang.srt.arg_groups.kv_cache_hook import handle_page_major_kv_layout
+from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.server_args import ServerArgs
 from sglang.test.ci.ci_register import register_cpu_ci
 
@@ -67,18 +68,18 @@ def _accepts(
         "mamba_backend": "triton",
     }.items():
         object.__setattr__(sa, name, value)
-    # `use_mla_backend` asks the model configuration, which this stand-in has
-    # no room for; the case under test is what the handler does with the answer.
-    # The handler imports it inside the function, so the source module is
-    # where the patch has to go.
-    with mock.patch(
-        "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=use_mla
-    ):
-        try:
-            handle_page_major_kv_layout(sa)
-            return True
-        except AssertionError:
-            return False
+    object.__setattr__(
+        sa,
+        "_model_config",
+        SimpleNamespace(
+            attention_arch=AttentionArch.MLA if use_mla else AttentionArch.MHA
+        ),
+    )
+    try:
+        handle_page_major_kv_layout(sa)
+        return True
+    except AssertionError:
+        return False
 
 
 class TestPageMajorBackendAllowlist(unittest.TestCase):
diff --git a/test/registered/unit/server_args/test_resolution_declarations.py b/test/registered/unit/server_args/test_resolution_declarations.py
index 269300f2e45b..89ef72d69828 100644
--- a/test/registered/unit/server_args/test_resolution_declarations.py
+++ b/test/registered/unit/server_args/test_resolution_declarations.py
@@ -934,47 +934,42 @@ def restore():
         self.addCleanup(restore)
 
     def _resolve_recording_each_entry(self, **supplied):
-        """Resolve, deep-copying every stash entry the moment it is appended."""
-        from sglang.srt.arg_groups import overrides
+        """Resolve, deep-copying every stash entry the moment it is appended.
 
+        The property is about the stash, so the seam is the stash: a list that
+        snapshots on append. Every declaration path -- `declare_resolution`,
+        `declare_late_resolution`, `declare_direct_writes` and the passes --
+        reaches it through `.append`, whatever it was imported as.
+        """
         recorded = []
 
-        def watch(name):
-            original = getattr(overrides, name)
-
-            def wrapper(server_args, *args, **kwargs):
-                result = original(server_args, *args, **kwargs)
-                stash = getattr(server_args, "_resolved_overrides", None) or []
-                while len(recorded) < len(stash):
-                    index = len(recorded)
-                    recorded.append((index, copy.deepcopy(stash[index])))
-                return result
-
-            return original, wrapper
-
-        # Every path that appends to the stash.
-        patched = {}
-        for name in (
-            "declare_resolution",
-            "declare_late_resolution",
-            "declare_direct_writes",
-            "run_post_process_pass",
-        ):
-            original, wrapper = watch(name)
-            patched[name] = original
-            setattr(overrides, name, wrapper)
-        try:
-            path = tempfile.mkdtemp(prefix="declared_values_")
-            self.addCleanup(shutil.rmtree, path, ignore_errors=True)
-            with open(os.path.join(path, "config.json"), "w") as handle:
-                json.dump(_MINI_CONFIG, handle)
-            server_args = ServerArgs(
-                model_path=path, device="cuda", random_seed=42, **supplied
-            )
-            server_args.resolve_once()
-        finally:
-            for name, original in patched.items():
-                setattr(overrides, name, original)
+        class _SnapshotOnAppend(list):
+            def append(self, entry):
+                super().append(entry)
+                recorded.append((len(self) - 1, copy.deepcopy(entry)))
+
+        class _WatchedArgs(ServerArgs):
+            """Whatever list the pipeline installs, snapshot what lands in it.
+
+            The pipeline resets the stash at the start of a resolution, so the
+            seam has to survive that assignment rather than precede it.
+            """
+
+            def __setattr__(self, name, value):
+                if name == "_resolved_overrides" and not isinstance(
+                    value, _SnapshotOnAppend
+                ):
+                    value = _SnapshotOnAppend(value)
+                super().__setattr__(name, value)
+
+        path = tempfile.mkdtemp(prefix="declared_values_")
+        self.addCleanup(shutil.rmtree, path, ignore_errors=True)
+        with open(os.path.join(path, "config.json"), "w") as handle:
+            json.dump(_MINI_CONFIG, handle)
+        server_args = _WatchedArgs(
+            model_path=path, device="cuda", random_seed=42, **supplied
+        )
+        server_args.resolve_once()
         return server_args, recorded
 
     def test_no_entry_changes_after_it_is_recorded(self):
diff --git a/test/registered/unit/server_args/test_server_args.py b/test/registered/unit/server_args/test_server_args.py
index 2b0fed392ca7..dd60a4ff96ef 100644
--- a/test/registered/unit/server_args/test_server_args.py
+++ b/test/registered/unit/server_args/test_server_args.py
@@ -979,10 +979,7 @@ def _make_args(self, attention_backend, prefill=None, decode=None, page_size=1):
         return args
 
     @patch("sglang.srt.arg_groups.overrides.is_sm100_supported", return_value=True)
-    @patch("sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False)
-    def test_combined_attention_backend_fa4_forces_page_size_128(
-        self, _mock_mla, _mock_sm100
-    ):
+    def test_combined_attention_backend_fa4_forces_page_size_128(self, _mock_sm100):
         # `--attention-backend fa4` (combined): prefill/decode fields stay None.
         args = self._make_args(attention_backend="fa4")
 
@@ -994,8 +991,7 @@ def test_combined_attention_backend_fa4_forces_page_size_128(
         self.assertEqual(resolved_view(args).page_size, 128)
 
     @patch("sglang.srt.arg_groups.overrides.is_sm100_supported", return_value=True)
-    @patch("sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False)
-    def test_explicit_prefill_fa4_forces_page_size_128(self, _mock_mla, _mock_sm100):
+    def test_explicit_prefill_fa4_forces_page_size_128(self, _mock_sm100):
         # `--prefill-attention-backend fa4`: the previously-covered path.
         args = self._make_args(attention_backend=None, prefill="fa4", page_size=1)
 
@@ -1869,12 +1865,7 @@ def _handled_args(self, **overrides):
             is_multimodal=False,
             is_multimodal_piecewise_cuda_graph_supported=False,
         )
-        with (
-            patch("sglang.srt.utils.is_cuda", return_value=True),
-            patch(
-                "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False
-            ),
-        ):
+        with patch("sglang.srt.utils.is_cuda", return_value=True):
             handle_cuda_graph_config(args)
         return args
 
@@ -1944,12 +1935,7 @@ def _handled_args(self, **overrides):
             is_multimodal=False,
             is_multimodal_piecewise_cuda_graph_supported=False,
         )
-        with (
-            patch("sglang.srt.utils.is_cuda", return_value=True),
-            patch(
-                "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False
-            ),
-        ):
+        with patch("sglang.srt.utils.is_cuda", return_value=True):
             handle_cuda_graph_config(args)
         return args
 
@@ -2010,12 +1996,7 @@ def _handled_args(self, *, architectures, is_multimodal, allowlisted):
             is_multimodal_piecewise_cuda_graph_supported=False,
             is_multimodal_breakable_cuda_graph_supported=allowlisted,
         )
-        with (
-            patch("sglang.srt.utils.is_cuda", return_value=True),
-            patch(
-                "sglang.srt.arg_groups.overrides.use_mla_backend", return_value=False
-            ),
-        ):
+        with patch("sglang.srt.utils.is_cuda", return_value=True):
             handle_cuda_graph_config(args)
         return args
 
diff --git a/test/registered/unit/test_model_overrides.py b/test/registered/unit/test_model_overrides.py
index 73b9e4ce567b..16068f9919b0 100644
--- a/test/registered/unit/test_model_overrides.py
+++ b/test/registered/unit/test_model_overrides.py
@@ -2231,6 +2231,7 @@ def _args(default_backend, **kw):
             defaults.update(kw)
             args = SimpleNamespace(**defaults)
             args.default_backend_for_test = default_backend
+            args._model_config = SimpleNamespace(attention_arch=AttentionArch.MHA)
             return args
 
         with patch.object(
@@ -2239,10 +2240,6 @@ def _args(default_backend, **kw):
             overrides_module,
             "get_default_attn_backend",
             lambda server_args, **_: server_args.default_backend_for_test,
-        ), patch.object(
-            overrides_module, "use_mla_backend", return_value=False
-        ), patch.object(
-            overrides_module, "model_config_of", return_value=None
         ):
             # radix on + no extra buffer + no spec -> page_size=1 path
             self.assertEqual(