config: the resolution callbacks into the record go to zero (#36972)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

python/sglang/srt/model_executor/model_runner_components/kv_pool_runtime.py

@@ -29,8 +29,10 @@ logger = logging.getLogger(__name__)
 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 (
-        server_args.post_capture_kv_sizing_planned()
+        post_capture_kv_sizing_planned(server_args)
         and current_platform.is_cuda()
         and not is_draft_worker
     )

python/sglang/srt/model_executor/pool_configurator.py

@@ -44,6 +44,7 @@ from sglang.srt.runtime_context import (
     get_parallel,
     get_schedule,
     get_spec,
+    max_speculative_num_draft_tokens,
 )
 from sglang.srt.utils.common import (
     ceil_align,
@@ -780,9 +781,7 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
         self.swa_page_size = cfg.window_size
         self.swa_ratio = get_schedule().swa_full_tokens_ratio
         self.is_speculative = get_spec().speculative_algorithm is not None
-        self.online_c128_mtp_max_draft_tokens = (
-            kvc.server_args.max_speculative_num_draft_tokens or 0
-        )
+        self.online_c128_mtp_max_draft_tokens = max_speculative_num_draft_tokens() or 0
         self.requested_max_running_requests_per_worker = (
             get_schedule().max_running_requests // kvc.ps.attn_dp_size
             if get_schedule().max_running_requests is not None
@@ -814,7 +813,7 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
         if self.is_speculative:
             # Ring is sized once here, so it must serve the largest adaptive tier.
             self._assert_ring_serves_draft_tokens(
-                kvc.server_args.max_speculative_num_draft_tokens or 0
+                max_speculative_num_draft_tokens() or 0
             )
 
         self.bytes_per_full_token = self._get_bytes_per_full_token()

python/sglang/srt/model_executor/runner/flashinfer_autotune.py

@@ -30,7 +30,9 @@ from sglang.srt.runtime_context import (
     get_disagg,
     get_exec,
     get_model,
+    get_schedule,
     get_spec,
+    max_prefill_buffer_tokens,
 )
 from sglang.srt.utils import empty_context, log_info_on_rank0
 
@@ -342,9 +344,7 @@ def maybe_flashinfer_autotune_extend(
     mr = runner.model_runner
     # Prefer the per-rank scheduler buffer while preserving the legacy ceiling
     # when chunked prefill is disabled.
-    num_tokens = (
-        mr.server_args.max_prefill_buffer_tokens() or mr.server_args.max_prefill_tokens
-    )
+    num_tokens = max_prefill_buffer_tokens() or get_schedule().max_prefill_tokens
     if num_tokens <= (decode_num_tokens or 0):
         return  # decode-shaped autotune already covered these buckets
     is_pd_prefill_target = (