[refactor] Resolve config declarations onto server_args at the end of __post_init__ (#30297)

python/sglang/srt/arg_groups/hisparse_hook.py

@@ -43,21 +43,30 @@ def _hisparse_allowed_backends(kv_cache_dtype: str) -> set[str]:
 def validate_hisparse_dsa_backend(
     server_args: ServerArgs, attr: str, label: str
 ) -> None:
-    backend = getattr(server_args, attr)
+    from sglang.srt.arg_groups.overrides import resolved_view
-    allowed_backends = _hisparse_allowed_backends(server_args.kv_cache_dtype)
+
+    # Invoked after the DSA kv-cache-dtype / split-backend declarations:
+    # read the resolving state through the view.
+    view = resolved_view(server_args)
+    backend = getattr(view, attr)
+    kv_cache_dtype = view.kv_cache_dtype
+    allowed_backends = _hisparse_allowed_backends(kv_cache_dtype)
     if backend is not None and backend not in allowed_backends:
         raise ValueError(
             f"HiSparse supports DSA {label} backend(s) {sorted(allowed_backends)} "
-            f"on this platform with --kv-cache-dtype={server_args.kv_cache_dtype}, "
+            f"on this platform with --kv-cache-dtype={kv_cache_dtype}, "
             f"but got --dsa-{label}-backend={backend}. "
             f"Please use --dsa-{label}-backend="
-            f"{_hisparse_default_backend(server_args.kv_cache_dtype)} "
+            f"{_hisparse_default_backend(kv_cache_dtype)} "
             "or omit it."
         )
 
 
 def validate_hisparse_kv_cache_dtype(server_args: ServerArgs) -> None:
-    if server_args.kv_cache_dtype in HISPARSE_KV_CACHE_DTYPES:
+    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:
         return
 
     choices = " or ".join(
@@ -65,7 +74,7 @@ def validate_hisparse_kv_cache_dtype(server_args: ServerArgs) -> None:
     )
     raise ValueError(
         f"HiSparse requires one of {HISPARSE_KV_CACHE_DTYPES} KV cache dtypes, "
-        f"but got --kv-cache-dtype={server_args.kv_cache_dtype}. Please use {choices}."
+        f"but got --kv-cache-dtype={kv_cache_dtype}. Please use {choices}."
     )
 
 
@@ -112,7 +121,13 @@ def validate_hisparse(server_args: ServerArgs) -> None:
             )
         return
 
-    if server_args.kv_cache_dtype not in ("bfloat16", "auto", "fp8_e4m3"):
+    from sglang.srt.arg_groups.overrides import resolved_view
+
+    if resolved_view(server_args).kv_cache_dtype not in (
+        "bfloat16",
+        "auto",
+        "fp8_e4m3",
+    ):
         validate_hisparse_kv_cache_dtype(server_args)
 
     for attr, label in [

python/sglang/srt/arg_groups/overrides.py

@@ -117,14 +117,10 @@ def _invoke_provider(
 
 class ResolvedView:
     """Read-only view of the resolving configuration handed to post-process
-    passes.
+    passes: the accumulated declarations overlaid on the pristine
-
+    ``server_args`` (residual imperative writes of non-resolved fields show
-    During the dual-apply transition the view forwards every read to the live
+    through the fallthrough) — exactly the state the legacy handler at the
-    ``server_args`` — the pristine input plus the declarations replayed so far
+    same slot observed. Writes are rejected: passes return declarations.
-    plus any residual imperative writes — which is exactly the state the
-    legacy handler at the same slot observed. In the end state (dual-apply
-    retired) the same type overlays the accumulated declarations on the
-    pristine object. Writes are rejected: passes return declarations.
     """
 
     __slots__ = ("_server_args", "_overlay")
@@ -162,14 +158,27 @@ def register_post_process(fn: Callable[..., dict]) -> Callable[..., dict]:
     return fn
 
 
-def run_post_process_pass(server_args: Any, fn: Callable[..., dict]) -> None:
+def _declaration_overlay(server_args: Any) -> Dict[str, Any]:
-    """Transition-period invocation of one pass at its legacy handler slot.
+    """Accumulated declared values: declarations never mutate
+    ``server_args``, so mid-resolution readers overlay them from the
+    declaration stash (last writer wins, like the gate)."""
+    overlay: Dict[str, Any] = {}
+    for _source, declared in getattr(server_args, "_resolved_overrides", None) or ():
+        overlay.update(declared)
+    return overlay
 
-    Evaluates the pass on the live state (through a read-only view), appends
+
-    its declaration to the declaration stash, and dual-applies it in place —
+def run_post_process_pass(server_args: Any, fn: Callable[..., dict]) -> None:
-    byte-identical to the imperative handler write this replaces.
+    """Invoke one pass at its legacy handler slot.
+
+    Evaluates the pass on the resolving state (a read-only view with the
+    accumulated declarations overlaid from the stash) and appends its
+    declaration to the stash. During ``__post_init__`` the fields stay
+    untouched — ``materialize_declarations`` applies the whole stash once at
+    the end of resolution; a pass invoked after materialization (a post-init
+    slot) writes through immediately.
     """
-    declared = fn(ResolvedView(server_args))
+    declared = fn(ResolvedView(server_args, overlay=_declaration_overlay(server_args)))
     if not isinstance(declared, dict):
         raise TypeError(
             f"post-process pass {fn.__qualname__} must return a dict, "
@@ -186,19 +195,70 @@ def run_post_process_pass(server_args: Any, fn: Callable[..., dict]) -> None:
             # must sit at or after it in __post_init__ order.
             stash = server_args._resolved_overrides = []
         stash.append(entry)
-        apply_declarations_to_server_args(server_args, [entry])
+        validate_declarations(server_args, [entry])
+        if getattr(server_args, "_declarations_materialized", False):
+            for field, value in declared.items():
+                setattr(server_args, field, value)
+
+
+def materialize_declarations(server_args: Any) -> None:
+    """Apply the accumulated declarations onto ``server_args`` once, at the
+    end of ``__post_init__`` (gate order: last writer wins). After this the
+    fields carry the resolved configuration — every post-init reader, in any
+    process, reads them directly; ``resolved_view`` remains an internal
+    helper for mid-resolution code only."""
+    for _source, declared in getattr(server_args, "_resolved_overrides", None) or ():
+        for field, value in declared.items():
+            setattr(server_args, field, value)
+    server_args._declarations_materialized = True
+
+
+def resolved_view(server_args: Any) -> ResolvedView:
+    """Read-only view of the resolving configuration for mid-resolution code
+    that is not a pass (``__post_init__`` handlers and hooks). Internal to
+    the resolution pipeline: after ``materialize_declarations`` runs, the
+    fields themselves carry the resolved values — read them directly."""
+    return ResolvedView(server_args, overlay=_declaration_overlay(server_args))
+
+
+def attention_backends_of(cfg: Any) -> tuple:
+    """(prefill, decode) attention backends of a config-shaped object (a
+    ResolvedView mid-resolution, or pristine server_args at dispatch time):
+    split fields fall back to the base backend."""
+    prefill = (
+        cfg.prefill_attention_backend
+        if cfg.prefill_attention_backend
+        else cfg.attention_backend
+    )
+    decode = (
+        cfg.decode_attention_backend
+        if cfg.decode_attention_backend
+        else cfg.attention_backend
+    )
+    return prefill, decode
+
+
+def mamba_extra_buffer_of(cfg: Any) -> bool:
+    """Mid-resolution equivalent of runtime_context.mamba_extra_buffer_enabled:
+    reads the (possibly overlaid) strategy from a config-shaped object."""
+    return cfg.disable_radix_cache is False and cfg.mamba_radix_cache_strategy in (
+        "extra_buffer",
+        "extra_buffer_lazy",
+    )
 
 
 def declare_load_time_override(source: str, declared: Dict[str, Any]) -> None:
-    """Transition helper for load-time resolved fields (model-file config
+    """Declare a load-time resolved field (model-file config overrides,
-    overrides, weight-resolved dtypes): dual-apply the declaration onto the
+    weight-resolved dtypes): apply it onto the published ``server_args`` —
-    published ``server_args`` — byte-identical to the imperative write this
+    resolution has already materialized, so post-init declarations write
-    replaces — and record it into the flags tier through the runtime gate."""
+    through — and record it into the flags tier through the runtime gate."""
     from sglang.srt.runtime_context import get_context
 
     ctx = get_context()
     entry = (source, dict(declared))
-    apply_declarations_to_server_args(ctx.server_args, [entry])
+    validate_declarations(ctx.server_args, [entry])
+    for field, value in declared.items():
+        setattr(ctx.server_args, field, value)
     ctx.record_runtime_overrides([entry])
 
 
@@ -741,8 +801,11 @@ def _qwen3_5_hybrid_overrides(server_args: Any, hf_config: Any) -> dict:
         use_mla_backend=server_args.use_mla_backend(),
         model_config=server_args.get_model_config(),
     )
+    # The mamba radix-cache pass runs before this dispatch: read the
+    # declared strategy through the view (the legacy branch observed the
+    # already-written field here).
     if default_attn_backend == "trtllm_mha" and not (
-        not server_args.enable_mamba_extra_buffer()
+        not mamba_extra_buffer_of(resolved_view(server_args))
         and not server_args.disable_radix_cache
         and server_args.speculative_algorithm is None
     ):
@@ -1519,12 +1582,57 @@ def _mla_backend_page_constraints(view: Any) -> dict:
     return {}
 
 
+@register_post_process
+def _mla_kv_cache_dtype_checks(view: Any) -> dict:
+    """Read-only validation pass in the attention-backend compatibility
+    handler: the TRT-LLM and tokenspeed MLA backends constrain the resolved
+    kv-cache dtype (declarations never reach the field, so the checks read
+    the view)."""
+    if (
+        view.attention_backend == "trtllm_mla"
+        or view.decode_attention_backend == "trtllm_mla"
+    ):
+        if not is_blackwell_supported():
+            raise ValueError(
+                "TRTLLM MLA backend is only supported on Blackwell GPUs (SM100/SM12x). Please use a different backend."
+            )
+        if view.kv_cache_dtype not in ["fp8_e4m3", "fp4_e2m1", "bf16", "auto"]:
+            raise ValueError(
+                "TensorRT-LLM MLA backend only supports kv-cache-dtype of fp8_e4m3, fp4_e2m1, bf16, or auto."
+            )
+    if (
+        view.attention_backend == "tokenspeed_mla"
+        or view.decode_attention_backend == "tokenspeed_mla"
+    ):
+        if not is_blackwell_supported():
+            raise ValueError(
+                "tokenspeed_mla backend is only supported on Blackwell GPUs (SM100/SM12x)."
+            )
+        if view.kv_cache_dtype not in ["fp8_e4m3"]:
+            raise ValueError(
+                "tokenspeed_mla backend requires kv-cache-dtype=fp8_e4m3, "
+                f"got {view.kv_cache_dtype}."
+            )
+    return {}
+
+
+@register_post_process
+def _hisparse_validation(view: Any) -> dict:
+    """Read-only validation pass: --enable-hisparse constraints (model class,
+    radix cache, kv dtype, DSA backends) read the resolved values through the
+    view."""
+    from sglang.srt.arg_groups.hisparse_hook import validate_hisparse
+
+    validate_hisparse(view)
+    return {}
+
+
 @register_post_process
 def _cutedsl_prefill_backend_fill(view: Any) -> dict:
     """Slot pass in the attention-backend compatibility handler: CuteDSL MLA
     is decode-only, so validate the combination and default the prefill side
     to trtllm_mla. The trtllm_mha check that follows at the legacy slot reads
-    the dual-applied value."""
+    the resolved value through the view."""
     if not (
         view.attention_backend == "cutedsl_mla"
         or view.decode_attention_backend == "cutedsl_mla"
@@ -1610,7 +1718,7 @@ def _attention_backend_platform_fallbacks(view: Any) -> dict:
 
 @register_post_process
 def _intel_xpu_page_constraint(view: Any) -> dict:
-    _, decode_backend = view.get_attention_backends()
+    _, decode_backend = attention_backends_of(view)
     if decode_backend == "intel_xpu":
         if view.use_mla_backend():
             supported_page_sizes = [16, 32, 64, 128]
@@ -1674,6 +1782,17 @@ def _data_parallelism_defaults(view: Any) -> dict:
     return {}
 
 
+@register_post_process
+def _dp_lm_head_validation(view: Any) -> dict:
+    """Read-only validation pass: dp-attention is a prerequisite for the
+    dp LM head. Reads the mid-resolution values through the view."""
+    if view.enable_dp_lm_head:
+        assert (
+            view.enable_dp_attention
+        ), "Please enable dp attention when setting enable_dp_lm_head. "
+    return {}
+
+
 @register_post_process
 def _moe_runner_backend_quant_constraints(view: Any) -> dict:
     """The quantization-driven moe_runner_backend resolutions at the head of
@@ -1728,6 +1847,49 @@ def _moe_runner_backend_quant_constraints(view: Any) -> dict:
 
 
 @register_post_process
+def _moe_runner_fusion_disable(view: Any) -> dict:
+    """FlashInfer CuteDSL / TRT-LLM / TRT-LLM-routed MoE runners require the
+    shared-experts fusion disabled; declared at the legacy write slots in
+    _handle_moe_kernel_config (before the deprecated cutlass env override, so
+    the runner value observed is the pre-override one)."""
+    runner = view.moe_runner_backend
+    if runner == "flashinfer_cutedsl":
+        logger.warning(
+            "FlashInfer CuteDSL MoE is enabled. --disable-shared-experts-fusion is automatically set."
+        )
+        return {"disable_shared_experts_fusion": True}
+    if runner in ("flashinfer_trtllm", "experimental_sgl_trtllm"):
+        logger.warning(
+            "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."
+        )
+        return {"disable_shared_experts_fusion": True}
+    if runner == "flashinfer_trtllm_routed":
+        logger.warning(
+            "FlashInfer TRTLLM routed MoE is enabled. --disable-shared-experts-fusion is automatically set."
+        )
+        return {"disable_shared_experts_fusion": True}
+    return {}
+
+
+def _a2a_fusion_adjustments(view: Any) -> dict:
+    """A2A-backend-driven shared-experts fusion adjustments, declared at the
+    legacy write slots in _handle_a2a_moe: DeepEP Waterfill requires the
+    fusion enabled; FlashInfer A2A requires it disabled."""
+    if view.moe_a2a_backend == "deepep" and view.enable_deepep_waterfill:
+        if view.disable_shared_experts_fusion:
+            logger.warning(
+                "disable_shared_experts_fusion is overridden to False because DeepEP Waterfill requires shared expert fusion."
+            )
+            return {"disable_shared_experts_fusion": False}
+        return {}
+    if view.moe_a2a_backend == "flashinfer":
+        logger.warning(
+            "Flashinfer MoE A2A is enabled. --disable-shared-experts-fusion is automatically set."
+        )
+        return {"disable_shared_experts_fusion": True}
+    return {}
+
+
 def _cutlass_moe_env_override(view: Any) -> dict:
     from sglang.srt.environ import envs
 
@@ -1846,16 +2008,25 @@ def _dllm_overlap_disable(view: Any) -> dict:
 
 @register_post_process
 def _dllm_page_size(view: Any) -> dict:
-    if view.dllm_algorithm is None or view.disable_radix_cache:
+    if view.dllm_algorithm is None:
         return {}
     from sglang.srt.dllm.config import DllmConfig
 
     config = DllmConfig.from_server_args(view)
-    if view.page_size % config.block_size != 0:
+    if not view.disable_radix_cache and view.page_size % config.block_size != 0:
         logger.warning(
             f"Setting page size to {config.block_size} for diffusion LLM inference"
         )
         return {"page_size": config.block_size}
+    if view.page_size > config.block_size:
+        # Legacy scheduler-init fallback, folded into the pass: the page
+        # size must not exceed the dllm block size.
+        logger.warning(
+            "WARNING: "
+            f"The page size {view.page_size} should not be larger than dllm block size {config.block_size}."
+            f"Page size now falls back to {config.block_size}"
+        )
+        return {"page_size": config.block_size}
     return {}
 
 
@@ -1937,45 +2108,35 @@ def apply_model_overrides(
     return records
 
 
-def apply_declarations_to_server_args(
+def validate_declarations(
     server_args: Any,
     declarations: Sequence[Tuple[str, Dict[str, Any]]],
-    *,
-    terminal: Sequence[Tuple[str, Dict[str, Any]]] = (),
 ) -> None:
-    """Transition-period dual-apply: replay declarations onto ``server_args``
+    """Fail-fast whitelist check at declaration time: a registry typo or a
-    in gate order, byte-identical to the legacy imperative writes.
+    not-yet-resolvable field must be rejected at its slot, not only at
-
+    publish time. Declarations never mutate ``server_args``.
-    Retired per field once that field's readers have all flipped to the flags
-    tier (at which point the server_args field returns to pristine).
-
-    Validates against the same whitelist as the publish gate BEFORE any write:
-    a registry typo or a not-yet-resolvable field must fail fast here, not
-    mutate ``server_args`` and only be rejected at publish time.
     """
     # Non-dataclass fixtures carry no Arg metadata (mirrors the
     # resolvable_fields escape); only real ServerArgs is validated.
-    if dataclasses.is_dataclass(type(server_args)):
+    if not dataclasses.is_dataclass(type(server_args)):
+        return
     whitelist = resolvable_fields(type(server_args))
-        for source, decl in list(declarations) + list(terminal):
+    for source, decl in declarations:
         unknown = set(decl) - whitelist
         if unknown:
             raise ValueError(
-                    f"{source}: {sorted(unknown)} not model-overridable; the "
+                f"{source}: {sorted(unknown)} not model-overridable; "
-                    "transition dual-apply refuses fields the publish gate "
+                "declarations are limited to the fields the publish gate "
-                    "would reject."
+                "accepts."
             )
-    for _source, decl in list(declarations) + list(terminal):
-        for field, value in decl.items():
-            setattr(server_args, field, value)
 
 
 def refresh_declared_fields(server_args: Any, fields: Iterable[str]) -> None:
-    """Transition helper for legacy code that overwrites a resolved field
+    """Helper for legacy code that overwrites a resolved field AFTER
-    AFTER the override collection in ``__post_init__`` (e.g.
+    materialization (e.g. ``ModelRunner.model_specific_adjustment`` forcing
-    ``ModelRunner.model_specific_adjustment`` forcing ``attention_backend``
+    ``attention_backend`` for HRM-Text). Redeclares the live value so the
-    for HRM-Text). Redeclares the live value so publish parity holds and the
+    publish parity holds and the flags tier materializes the adjusted end
-    flags tier materializes the adjusted end state.
+    state.
     """
     _missing = object()
     declarations = server_args._resolved_overrides
@@ -1998,8 +2159,8 @@ def assert_flag_parity(
     *,
     leaf_map: Optional[Dict[str, str]] = None,
 ) -> None:
-    """Dual-apply drift guard: each migrated field's flag leaf must equal the
+    """Drift guard: each declared field's flag leaf must equal the
-    (dual-applied) ``server_args`` value."""
+    (materialized) ``server_args`` value."""
     mismatches = []
     for field in fields:
         owner, leaf = resolve_flag_leaf(flags, field, leaf_map=leaf_map)

python/sglang/srt/arg_groups/speculative_hook.py

@@ -135,7 +135,9 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None:
 
 
 def _handle_dflash(server_args: ServerArgs) -> None:
-    if server_args.enable_dp_attention:
+    from sglang.srt.arg_groups.overrides import resolved_view
+
+    if resolved_view(server_args).enable_dp_attention:
         raise ValueError(
             "Currently DFLASH speculative decoding does not support dp attention."
         )
@@ -264,7 +266,12 @@ def _resolve_dflash_draft_attention_backend(server_args: ServerArgs) -> None:
 
     draft_backend = server_args.speculative_draft_attention_backend
     if draft_backend is None:
-        draft_backend, _ = server_args.get_attention_backends()
+        from sglang.srt.arg_groups.overrides import (
+            attention_backends_of,
+            resolved_view,
+        )
+
+        draft_backend, _ = attention_backends_of(resolved_view(server_args))
     if draft_backend is None:
         draft_backend = fallback_backend
     elif draft_backend == "trtllm_mha":
@@ -305,9 +312,14 @@ def _handle_frozen_kv_mtp(server_args: ServerArgs) -> None:
 
 
 def _handle_eagle_family(server_args: ServerArgs) -> None:
+    from sglang.srt.arg_groups.overrides import (
+        attention_backends_of,
+        resolved_view,
+    )
+
     if (
         server_args.speculative_algorithm == "STANDALONE"
-        and server_args.enable_dp_attention
+        and resolved_view(server_args).enable_dp_attention
     ):
         # TODO: support dp attention for standalone speculative decoding
         raise ValueError(
@@ -320,7 +332,7 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
             "Max running requests is reset to 48 for speculative decoding. You can override this by explicitly setting --max-running-requests."
         )
 
-    if server_args.disable_overlap_schedule:
+    if resolved_view(server_args).disable_overlap_schedule:
         logger.warning(
             "Non-overlap (synchronous) spec v2 is used for eagle/eagle3/standalone "
             "speculative decoding."
@@ -375,11 +387,7 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
             server_args.speculative_num_draft_tokens,
         ) = _auto_choose_speculative_params(server_args, model_arch)
 
-    if (
+    if "trtllm_mha" in attention_backends_of(resolved_view(server_args)):
-        server_args.attention_backend == "trtllm_mha"
-        or server_args.decode_attention_backend == "trtllm_mha"
-        or server_args.prefill_attention_backend == "trtllm_mha"
-    ):
         if server_args.speculative_eagle_topk > 1:
             raise ValueError(
                 "trtllm_mha backend only supports topk = 1 for speculative decoding."
@@ -416,7 +424,9 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
                 "--enable-deterministic-inference; the sampling kernel draws "
                 "coins from the global RNG and is not batch-invariant."
             )
-        if server_args.enable_multi_layer_eagle:
+        from sglang.srt.arg_groups.overrides import resolved_view
+
+        if resolved_view(server_args).enable_multi_layer_eagle:
             raise NotImplementedError(
                 "--speculative-use-rejection-sampling is not supported with "
                 "multi-layer EAGLE (--enable-multi-layer-eagle)."
@@ -440,15 +450,16 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
     # pass + per-branch expand pass with prefix-tail dup). Only these backends implement
     # it; flashmla / trtllm_mla / cutlass_mla can't express the per-branch tree, so reject.
     _PAGE_TREE_SPEC_BACKENDS = ("flashinfer", "fa3", "triton")
+    view = resolved_view(server_args)
     if (
         server_args.speculative_eagle_topk > 1
-        and server_args.page_size > 1
+        and view.page_size > 1
-        and server_args.attention_backend not in _PAGE_TREE_SPEC_BACKENDS
+        and view.attention_backend not in _PAGE_TREE_SPEC_BACKENDS
     ):
         raise ValueError(
             f"speculative_eagle_topk > 1 with page_size > 1 is only supported on "
             f"{_PAGE_TREE_SPEC_BACKENDS}; got attention_backend="
-            f"{server_args.attention_backend!r}. Use page_size == 1 or one of those backends."
+            f"{view.attention_backend!r}. Use page_size == 1 or one of those backends."
         )
 
 
@@ -499,18 +510,21 @@ 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 (
         server_args.speculative_eagle_topk > 1
-        and server_args.page_size > 1
+        and view.page_size > 1
-        and server_args.attention_backend != "flashinfer"
+        and view.attention_backend != "flashinfer"
     ):
         raise ValueError(
             f"speculative_eagle_topk({server_args.speculative_eagle_topk}) > 1 "
-            f"with page_size({server_args.page_size}) > 1 is unstable "
+            f"with page_size({view.page_size}) > 1 is unstable "
             "and produces incorrect results for paged attention backends. "
             "This combination is only supported for the 'flashinfer' backend."
         )
-    if server_args.enable_dp_attention:
+    if view.enable_dp_attention:
         # TODO: support dp attention for ngram speculative decoding
         raise ValueError(
             "Currently ngram speculative decoding does not support dp attention."

python/sglang/srt/speculative/adaptive_spec_params.py

@@ -49,6 +49,8 @@ DEFAULT_ADAPTIVE_CONFIG: dict[str, dict] = {
 
 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 resolved_view
+
     if server_args.speculative_algorithm not in ("EAGLE", "EAGLE3"):
         return (
             f"speculative_algorithm={server_args.speculative_algorithm} "
@@ -62,12 +64,12 @@ def adaptive_unsupported_reason(server_args: ServerArgs) -> str | None:
             f"speculative_eagle_topk={server_args.speculative_eagle_topk} "
             "(only topk=1 is supported)"
         )
-    if server_args.enable_dp_attention:
+    if resolved_view(server_args).enable_dp_attention:
         return (
             "enable_dp_attention=True is not supported "
             "(adaptive tier decisions are not synchronized across DP ranks)"
         )
-    if server_args.enable_multi_layer_eagle:
+    if resolved_view(server_args).enable_multi_layer_eagle:
         return (
             "enable_multi_layer_eagle=True is not supported "
             "(MultiLayerEagleWorkerV2 does not implement adaptive)"

test/registered/mock_model/test_self_unit_install.py

@@ -17,6 +17,10 @@ register_amd_ci(est_time=60, suite="extra-a-test-1-gpu-small-amd")
 
 
 def _make_server_args(*, sampling_backend: str) -> SimpleNamespace:
+    # The install gate reads the resolved backend from the flags tier.
+    from sglang.srt.runtime_context import get_flags
+
+    get_flags().sampling_backend = sampling_backend
     return SimpleNamespace(sampling_backend=sampling_backend)
 
 

test/registered/unit/model_executor/test_pool_configurator.py

@@ -102,6 +102,10 @@ def _make_model_runner(
 
     sa = SimpleNamespace()
     sa.swa_full_tokens_ratio = swa_full_tokens_ratio
+    # The configurator reads the resolved ratio from the flags tier.
+    from sglang.srt.runtime_context import get_flags
+
+    get_flags().swa_full_tokens_ratio = swa_full_tokens_ratio
     sa.page_size = page_size
     sa.disable_radix_cache = disable_radix_cache
     sa.chunked_prefill_size = chunked_prefill_size

test/registered/unit/models/test_deepseek_v4_shared_expert_fusion.py

@@ -42,7 +42,7 @@ class TestDeepseekV4SharedExpertFusionPolicy(unittest.TestCase):
 
         self.assertEqual(model.num_fused_shared_experts, 0)
         self.assertTrue(get_flags().disable_shared_experts_fusion)
-        # dual-apply transition: the published config carries the value too
+        # post-init declaration writes through to the published config
         self.assertTrue(server_args.disable_shared_experts_fusion)
 
     def test_enables_shared_fusion_when_enforced(self):

test/registered/unit/server_args/test_server_args.py

@@ -354,7 +354,10 @@ class TestFa4PageSizeAutoForce(CustomTestCase):
 
         args._handle_attention_backend_compatibility()
 
-        self.assertEqual(args.page_size, 128)
+        from sglang.srt.arg_groups.overrides import resolved_view
+
+        self.assertEqual(args.page_size, 1)  # dual-apply retired: pristine
+        self.assertEqual(resolved_view(args).page_size, 128)
 
     @patch("sglang.srt.arg_groups.overrides.is_sm100_supported", return_value=True)
     @patch("sglang.srt.server_args.ServerArgs.use_mla_backend", return_value=False)
@@ -364,7 +367,10 @@ class TestFa4PageSizeAutoForce(CustomTestCase):
 
         args._handle_attention_backend_compatibility()
 
-        self.assertEqual(args.page_size, 128)
+        from sglang.srt.arg_groups.overrides import resolved_view
+
+        self.assertEqual(args.page_size, 1)  # dual-apply retired: pristine
+        self.assertEqual(resolved_view(args).page_size, 128)
 
 
 class TestContextParallelServerArgs(CustomTestCase):
@@ -1119,7 +1125,11 @@ class TestDeepEPWaterfillArgs(CustomTestCase):
         # dummy-model path short-circuits __post_init__; invoke the handler directly.
         server_args._handle_a2a_moe()
 
-        self.assertFalse(server_args.disable_shared_experts_fusion)
+        from sglang.srt.arg_groups.overrides import resolved_view
+
+        # dual-apply retired: the fields stay pristine, the declarations win
+        self.assertTrue(server_args.disable_shared_experts_fusion)
+        self.assertFalse(resolved_view(server_args).disable_shared_experts_fusion)
         self.assertTrue(server_args.enforce_shared_experts_fusion)
 
     def test_waterfill_overrides_moe_a2a_backend_to_deepep(self):
@@ -1131,7 +1141,10 @@ class TestDeepEPWaterfillArgs(CustomTestCase):
         # dummy-model path short-circuits __post_init__; invoke the handler directly.
         server_args._handle_a2a_moe()
 
-        self.assertEqual(server_args.moe_a2a_backend, "deepep")
+        from sglang.srt.arg_groups.overrides import resolved_view
+
+        self.assertEqual(server_args.moe_a2a_backend, "none")  # pristine
+        self.assertEqual(resolved_view(server_args).moe_a2a_backend, "deepep")
         self.assertTrue(server_args.enforce_shared_experts_fusion)
 
     def test_waterfill_supports_deepep_low_latency_mode(self):

test/registered/unit/test_model_overrides.py

@@ -19,11 +19,10 @@ from sglang.srt.arg_groups import overrides as overrides_module
 from sglang.srt.arg_groups.arg_utils import A, Arg, resolvable_fields
 from sglang.srt.arg_groups.overrides import (
     OverrideRecord,
-    apply_declarations_to_server_args,
     apply_model_overrides,
-    assert_flag_parity,
     collect_model_override_declarations,
     register_model_override,
+    validate_declarations,
 )
 from sglang.srt.runtime_context import (
     _StaticFlags,
@@ -207,7 +206,7 @@ class TestResolvedViewAndPasses(CustomTestCase):
         self.assertEqual(view.a, 10)
         self.assertEqual(view.b, 2)
 
-    def test_run_pass_appends_stash_and_dual_applies(self):
+    def test_run_pass_appends_stash_and_stays_pristine(self):
         from sglang.srt.arg_groups.overrides import run_post_process_pass
 
         live = SimpleNamespace(x=None, _resolved_overrides=[])
@@ -216,11 +215,12 @@ class TestResolvedViewAndPasses(CustomTestCase):
             return {"x": "filled"} if view.x is None else {}
 
         run_post_process_pass(live, _fill_x)
-        self.assertEqual(live.x, "filled")  # dual-applied in place
+        self.assertIsNone(live.x)  # never applied in place
         self.assertEqual(
             live._resolved_overrides, [(_fill_x.__qualname__, {"x": "filled"})]
         )
-        run_post_process_pass(live, _fill_x)  # now a no-op
+        # the next invocation sees the declared value through the overlay
+        run_post_process_pass(live, _fill_x)
         self.assertEqual(len(live._resolved_overrides), 1)
 
     def test_run_pass_rejects_non_dict(self):
@@ -415,7 +415,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_mistral_large3_forces_bfloat16(self):
         sa = self._construct("MistralLarge3ForCausalLM", "mistral")
-        self.assertEqual(sa.dtype, "bfloat16")  # dual-apply == legacy write
+        self.assertEqual(sa.dtype, "bfloat16")  # materialized at end of resolution
         self.assertIn(
             ("MODEL_OVERRIDES['MistralLarge3ForCausalLM']", {"dtype": "bfloat16"}),
             sa._resolved_overrides,
@@ -424,15 +424,15 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_pixtral_forces_bfloat16(self):
         sa = self._construct("PixtralForConditionalGeneration", "pixtral")
-        self.assertEqual(sa.dtype, "bfloat16")
+        self.assertEqual(sa.dtype, "bfloat16")  # materialized
         self.assertEqual(self._publish(sa).dtype, "bfloat16")
 
     def test_user_requested_dtype_is_still_overridden(self):
         # Legacy fidelity: the arch branch overwrote dtype unconditionally,
-        # so the declaration must too. The pristine request survives only on
+        # so the declaration must too. The pristine request survives on
-        # provenance (and, post-V3, as the un-overridden server_args field).
+        # provenance; the materialized field carries the override.
         sa = self._construct("MistralLarge3ForCausalLM", "mistral", dtype="float16")
-        self.assertEqual(sa.dtype, "bfloat16")
+        self.assertEqual(sa.dtype, "bfloat16")  # materialized
         self.assertEqual(self._publish(sa).dtype, "bfloat16")
 
     def test_control_arch_keeps_pristine_dtype(self):
@@ -446,7 +446,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_minimax_m2_enables_tf32_matmul(self):
         sa = self._construct("MiniMaxM2ForCausalLM", "llama")
-        self.assertTrue(sa.enable_tf32_matmul)  # dual-apply == legacy write
+        self.assertTrue(sa.enable_tf32_matmul)  # materialized
         self.assertIn(
             ("_minimax_m2_overrides", {"enable_tf32_matmul": True}),
             sa._resolved_overrides,
@@ -491,7 +491,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             config_extra=config_extra,
             enable_hierarchical_cache=True,
         )
-        # dual-apply == legacy writes
+        # materialized at the end of resolution
         self.assertEqual(sa.swa_full_tokens_ratio, 1.0)
         self.assertTrue(sa.disable_hybrid_swa_memory)
         flags = self._publish(sa)
@@ -500,7 +500,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_gemma2_disables_hybrid_swa_memory(self):
         sa = self._construct("Gemma2ForCausalLM", "llama")
-        self.assertTrue(sa.disable_hybrid_swa_memory)  # dual-apply == legacy
+        self.assertTrue(sa.disable_hybrid_swa_memory)  # materialized
         self.assertIn(
             ("_gemma2_gemma3_overrides", {"disable_hybrid_swa_memory": True}),
             sa._resolved_overrides,
@@ -509,7 +509,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_olmo2_disables_hybrid_swa_memory(self):
         sa = self._construct("Olmo2ForCausalLM", "llama")
-        self.assertTrue(sa.disable_hybrid_swa_memory)
+        self.assertTrue(sa.disable_hybrid_swa_memory)  # materialized
         self.assertTrue(self._publish(sa).disable_hybrid_swa_memory)
 
     def test_exaone_conditional_on_sliding_window_pattern(self):
@@ -520,7 +520,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             config_extra={"sliding_window_pattern": "LLLG"},
             attention_backend="fa3",
         )
-        self.assertTrue(sa.disable_hybrid_swa_memory)
+        self.assertTrue(sa.disable_hybrid_swa_memory)  # materialized
         self.assertTrue(self._publish(sa).disable_hybrid_swa_memory)
 
     def test_exaone_without_pattern_declares_nothing(self):
@@ -543,7 +543,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             "llama",
             config_extra={"quantization_config": {"quant_method": "mxfp4"}},
         )
-        self.assertEqual(sa.dtype, "bfloat16")  # dual-apply == legacy
+        self.assertEqual(sa.dtype, "bfloat16")  # materialized
         self.assertEqual(self._publish(sa).dtype, "bfloat16")
 
     def test_gpt_oss_without_mxfp4_keeps_pristine_dtype(self):
@@ -569,7 +569,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
         sa = self._construct("LlamaForCausalLM", "llama")
         expected = "flashinfer" if is_flashinfer_available() else "pytorch"
-        self.assertEqual(sa.sampling_backend, expected)
+        self.assertEqual(sa.sampling_backend, expected)  # materialized
         self.assertIn(
             ("_sampling_backend_default", {"sampling_backend": expected}),
             sa._resolved_overrides,
@@ -587,20 +587,20 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             "LlamaForCausalLM", "llama", enable_deterministic_inference=True
         )
         # two pass writers chain: default fill, then the deterministic force —
-        # last writer wins on the flags leaf and parity holds end-to-end.
+        # last writer wins; materialization lands the end state on the fields.
         self.assertEqual(sa.sampling_backend, "pytorch")
         flags = self._publish(sa)
         self.assertEqual(flags.sampling_backend, "pytorch")
         # the deterministic attention fill declared a compatible backend and
         # the compatibility default-fill then had nothing to do
-        self.assertIn(
+        deterministic_fills = [
-            (
+            decl["attention_backend"]
-                "_deterministic_attention_backend",
+            for source, decl in sa._resolved_overrides
-                {"attention_backend": sa.attention_backend},
+            if source == "_deterministic_attention_backend"
-            ),
+        ]
-            sa._resolved_overrides,
+        self.assertEqual(len(deterministic_fills), 1)
-        )
+        self.assertEqual(sa.attention_backend, deterministic_fills[0])
-        self.assertEqual(flags.attn.backend, sa.attention_backend)
+        self.assertEqual(flags.attn.backend, deterministic_fills[0])
 
     def test_deterministic_incompatible_backend_raises(self):
         from sglang.srt.arg_groups.overrides import (
@@ -631,13 +631,15 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_dllm_forces_flashinfer_with_cuda_graph(self):
         # CUDA path: cuda graph enabled by default -> dllm forces flashinfer.
+        # A real dllm arch: the page pass now runs regardless of the radix
+        # switch and builds DllmConfig for it.
         sa = self._construct(
-            "LlamaForCausalLM",
+            "SDARForCausalLM",
             "llama",
             dllm_algorithm="LowConfidence",
             disable_radix_cache=True,
         )
-        self.assertEqual(sa.attention_backend, "flashinfer")
+        self.assertEqual(sa.attention_backend, "flashinfer")  # materialized
         self.assertIn(
             ("_dllm_attention_backend", {"attention_backend": "flashinfer"}),
             sa._resolved_overrides,
@@ -647,24 +649,35 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_attention_backend_leaf_materializes_end_state(self):
         # The default-fill pass declares the platform-selected backend; the
-        # leaf must equal the final server_args value (publish parity).
+        # leaf must equal the last declared value while the server_args field
+        # stays pristine (dual-apply retired).
         sa = self._construct("LlamaForCausalLM", "llama")
-        declared = {f for _s, d in sa._resolved_overrides for f in d}
+        declared_values = [
-        self.assertIn("attention_backend", declared)  # default fill declared
+            d["attention_backend"]
-        self.assertEqual(self._publish(sa).attn.backend, sa.attention_backend)
+            for _s, d in sa._resolved_overrides
+            if "attention_backend" in d
+        ]
+        self.assertTrue(declared_values)  # default fill declared
+        self.assertEqual(sa.attention_backend, declared_values[-1])  # materialized
+        self.assertEqual(self._publish(sa).attn.backend, declared_values[-1])
 
-    def test_runner_side_adjustment_can_refresh_declaration(self):
+    def test_post_materialize_pass_writes_through(self):
-        from sglang.srt.arg_groups.overrides import refresh_declared_fields
+        from sglang.srt.arg_groups.overrides import run_post_process_pass
 
+        # A pass invoked after materialization (a post-init slot, like the
+        # legacy runner-side adjustments) declares AND writes through, so
+        # field readers and the publish see the same end state.
         sa = self._construct("LlamaForCausalLM", "llama")
-        declared = {f for _s, d in sa._resolved_overrides for f in d}
+        resolved_before = sa.attention_backend
-        self.assertIn("attention_backend", declared)
+
-        # Simulate a legacy runner-side overwrite between collection and publish
+        def _force_triton(view):
-        # (model_specific_adjustment forces attention_backend for HRM-Text).
+            if view.attention_backend != "triton":
-        sa.attention_backend = "fa3" if sa.attention_backend != "fa3" else "triton"
+                return {"attention_backend": "triton"}
-        with self.assertRaises(AssertionError):
+            return {}
-            self._publish(sa)  # stale declaration breaks parity
+
-        refresh_declared_fields(sa, ("attention_backend",))
+        run_post_process_pass(sa, _force_triton)
+        if resolved_before != "triton":
+            self.assertEqual(sa.attention_backend, "triton")
         self.assertEqual(self._publish(sa).attn.backend, sa.attention_backend)
 
     def test_attention_backend_user_choice_declares_nothing_extra(self):
@@ -798,9 +811,41 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             return_value=SimpleNamespace(block_size=32),
         ):
             self.assertEqual(_view() and _dllm_page_size(_view()), {"page_size": 32})
-            self.assertEqual(_dllm_page_size(_view(page_size=64)), {})  # aligned
+            # aligned but larger than the block: the scheduler-init fallback
-        self.assertEqual(_dllm_page_size(_view(dllm_algorithm=None)), {})
+            # (folded into this pass) still caps the page at the block size
+            self.assertEqual(_dllm_page_size(_view(page_size=64)), {"page_size": 32})
+            self.assertEqual(_dllm_page_size(_view(page_size=32)), {})  # equal
+            # radix disabled skips the alignment fill but keeps the cap
             self.assertEqual(_dllm_page_size(_view(disable_radix_cache=True)), {})
+            self.assertEqual(
+                _dllm_page_size(_view(disable_radix_cache=True, page_size=64)),
+                {"page_size": 32},
+            )
+        self.assertEqual(_dllm_page_size(_view(dllm_algorithm=None)), {})
+
+    def test_declaration_overlay_mechanics(self):
+        from sglang.srt.arg_groups.overrides import run_post_process_pass
+
+        live = SimpleNamespace(x="user", y=None, _resolved_overrides=[])
+
+        def _resolve_x(view):
+            return {"x": "resolved"} if view.x == "user" else {}
+
+        def _read_x(view):
+            return {"y": view.x}
+
+        run_post_process_pass(live, _resolve_x)
+        # declaration recorded, but server_args stays pristine
+        self.assertEqual(
+            live._resolved_overrides, [(_resolve_x.__qualname__, {"x": "resolved"})]
+        )
+        self.assertEqual(live.x, "user")
+        # a later pass sees the resolved value through the view overlay
+        run_post_process_pass(live, _read_x)
+        self.assertEqual(
+            live._resolved_overrides[-1], (_read_x.__qualname__, {"y": "resolved"})
+        )
+        self.assertIsNone(live.y)  # never applied in place
 
     def test_overlap_disable_passes(self):
         from sglang.srt.arg_groups.overrides import (
@@ -1531,9 +1576,12 @@ class TestGoldenModelOverrides(_IsolatedPublish):
 
     def test_page_size_leaf_materializes_end_state(self):
         sa = self._construct("LlamaForCausalLM", "llama")
-        declared = {f for _s, d in sa._resolved_overrides for f in d}
+        declared_values = [
-        self.assertIn("page_size", declared)  # default fill declared
+            d["page_size"] for _s, d in sa._resolved_overrides if "page_size" in d
-        self.assertEqual(self._publish(sa).page_size, sa.page_size)
+        ]
+        self.assertTrue(declared_values)  # default fill declared
+        self.assertEqual(sa.page_size, declared_values[-1])  # materialized
+        self.assertEqual(self._publish(sa).page_size, declared_values[-1])
 
     def test_qwen3_5_hybrid_coupled_declaration(self):
         from sglang.srt.arg_groups.overrides import _qwen3_5_hybrid_overrides
@@ -1544,7 +1592,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
                 _get_default_attn_backend=lambda **_: default_backend,
                 use_mla_backend=lambda: False,
                 get_model_config=lambda: None,
-                enable_mamba_extra_buffer=lambda: False,
+                mamba_radix_cache_strategy="auto",
                 disable_radix_cache=False,
                 speculative_algorithm=None,
             )
@@ -1571,6 +1619,24 @@ class TestGoldenModelOverrides(_IsolatedPublish):
                 ),
                 {},
             )
+            # the mamba pass ran before this dispatch and stashed the
+            # extra-buffer strategy: the callable must see it through the
+            # view (SM100 hybrid keeps trtllm_mha + page 64)
+            self.assertEqual(
+                _qwen3_5_hybrid_overrides(
+                    _args(
+                        "trtllm_mha",
+                        _resolved_overrides=[
+                            (
+                                "_mamba_radix_cache_declarations",
+                                {"mamba_radix_cache_strategy": "extra_buffer"},
+                            )
+                        ],
+                    ),
+                    None,
+                ),
+                {"attention_backend": "trtllm_mha", "page_size": 64},
+            )
         with patch.object(overrides_module, "is_sm100_supported", return_value=False):
             self.assertEqual(_qwen3_5_hybrid_overrides(_args("fa3"), None), {})
 
@@ -1738,7 +1804,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
         self.assertEqual(
             _intel_xpu_page_constraint(
                 _view(
-                    get_attention_backends=lambda: (None, "intel_xpu"),
+                    decode_attention_backend="intel_xpu",
                     use_mla_backend=lambda: False,
                 )
             ),
@@ -1747,7 +1813,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
         self.assertEqual(
             _intel_xpu_page_constraint(
                 _view(
-                    get_attention_backends=lambda: (None, "intel_xpu"),
+                    decode_attention_backend="intel_xpu",
                     use_mla_backend=lambda: True,
                     page_size=16,  # MLA decode accepts 16
                 )
@@ -2089,22 +2155,22 @@ class TestGoldenModelOverrides(_IsolatedPublish):
         self.assertEqual(_step3p_overrides(_args(), None), {})
 
 
-class TestDualApplyParity(CustomTestCase):
+class TestDeclarationValidation(CustomTestCase):
-    def test_dual_apply_replays_and_parity_holds(self):
+    def test_declarations_never_mutate_server_args(self):
         flags, args = _FakeFlags(), _FakeArgs()
         declarations = [("src", {"resolved_by_model": "dsv4", "also_resolved": 7})]
         apply_model_overrides(flags, args, declarations)
-        apply_declarations_to_server_args(args, declarations)
+        validate_declarations(args, declarations)
-        self.assertEqual(args.resolved_by_model, "dsv4")
+        # the leaves carry the declared values; the fields stay pristine
-        self.assertEqual(args.also_resolved, 7)
+        self.assertEqual(flags.resolved_by_model, "dsv4")
-        assert_flag_parity(flags, args, ["resolved_by_model", "also_resolved"])
+        self.assertEqual(flags.also_resolved, 7)
+        self.assertEqual(args.resolved_by_model, _FakeArgs.resolved_by_model)
+        self.assertEqual(args.also_resolved, _FakeArgs.also_resolved)
 
-    def test_parity_detects_drift(self):
+    def test_validation_rejects_unknown_fields(self):
-        flags, args = _FakeFlags(), _FakeArgs()
+        args = _FakeArgs()
-        apply_model_overrides(flags, args, [("src", {"resolved_by_model": "x"})])
+        with self.assertRaises(ValueError):
-        # dual-apply skipped -> server_args still pristine -> drift is caught
+            validate_declarations(args, [("src", {"nope": 1})])
-        with self.assertRaises(AssertionError):
-            assert_flag_parity(flags, args, ["resolved_by_model"])
 
 
 if __name__ == "__main__":

test/registered/unit/test_runtime_context.py

@@ -339,15 +339,6 @@ class TestRuntimeResolutionStages(_IsolatedServerArgs):
         self.assertEqual(get_flags().sampling_backend, "pytorch")
         self.assertEqual(get_flags().page_size, 64)  # earlier stage survives
 
-    def test_record_parity_failure_rolls_back(self):
-        self._publish(page_size=1)
-        flags_before = get_flags()
-        with self.assertRaises(AssertionError):
-            # declared value diverges from the live server_args (no dual-apply)
-            get_context().record_runtime_overrides([("bad", {"page_size": 64})])
-        self.assertIs(get_flags(), flags_before)  # previous flags intact
-        self.assertEqual(get_context()._runtime_overrides, [])  # rolled back
-
     def test_record_whitelist_violation_rolls_back(self):
         self._publish()
         with self.assertRaises(ValueError):
@@ -366,13 +357,15 @@ class TestRuntimeResolutionStages(_IsolatedServerArgs):
         finally:
             reset_context()
 
-    def test_declare_load_time_override_dual_applies_and_records(self):
+    def test_declare_load_time_override_applies_and_records(self):
         from sglang.srt.arg_groups.overrides import declare_load_time_override
 
         args = self._publish(page_size=1)
         declare_load_time_override("model.load_time", {"page_size": 64})
-        self.assertEqual(args.page_size, 64)  # dual-applied onto server_args
+        # post-init declaration: written through to the field and resolved
-        self.assertEqual(get_flags().page_size, 64)  # resolved into the leaf
+        # into the leaf
+        self.assertEqual(args.page_size, 64)
+        self.assertEqual(get_flags().page_size, 64)
         self.assertEqual(
             get_context()._runtime_overrides,
             [("model.load_time", {"page_size": 64})],
@@ -415,6 +408,21 @@ class TestRuntimeResolutionStages(_IsolatedServerArgs):
         finally:
             reset_context()
 
+    def test_declared_leaf_wins_over_stale_field(self):
+        # A stash entry always drives the leaf at publish, even if the field
+        # value diverged (e.g. a fixture that skipped materialization).
+        @dataclasses.dataclass
+        class _Args:
+            enable_dp_lm_head: A[bool, Arg(help="d", resolvable=True)] = True
+            _resolved_overrides: list = dataclasses.field(default_factory=list)
+
+        args = _Args()
+        args._resolved_overrides = [("dp", {"enable_dp_lm_head": False})]
+        args._declarations_materialized = True
+        args.enable_dp_lm_head = False
+        get_context().set_server_args(args)
+        self.assertFalse(get_flags().enable_dp_lm_head)
+
     def test_bare_dataclass_publish_skips_materialization(self):
         # object.__new__(ServerArgs) fixtures (no __init__, no field values)
         # must publish without touching the flags tier — dataclass defaults