[Fix] Keep the MiniCPM-SALA config reads visible to the resolution ratchets (#36178)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>

python/sglang/srt/arg_groups/overrides.py

@@ -1235,14 +1235,17 @@ def _minicpm_sala_overrides(server_args: Any, hf_config: Any) -> dict:
             "minicpm_flashattn": ("fa4" if is_blackwell_supported() else "fa3"),
             "minicpm_flashinfer": "flashinfer",
         }
-        for backend_field in (
-            "attention_backend",
-            "prefill_attention_backend",
-            "decode_attention_backend",
-        ):
-            dense_backend = dense_backends.get(getattr(server_args, backend_field))
-            if dense_backend is not None:
-                overrides[backend_field] = dense_backend
+        # Literal keys keep the written-field set statically derivable; a loop
+        # variable hides it from the census in test_chain_read_ratchet.py.
+        dense_attention = dense_backends.get(server_args.attention_backend)
+        if dense_attention is not None:
+            overrides["attention_backend"] = dense_attention
+        dense_prefill = dense_backends.get(server_args.prefill_attention_backend)
+        if dense_prefill is not None:
+            overrides["prefill_attention_backend"] = dense_prefill
+        dense_decode = dense_backends.get(server_args.decode_attention_backend)
+        if dense_decode is not None:
+            overrides["decode_attention_backend"] = dense_decode
     elif has_sparse_attention:
         uses_sparse_backend = server_args.is_attention_backend_not_set() or any(
             backend in ("minicpm_flashattn", "minicpm_flashinfer")

python/sglang/srt/layers/attention/minicpm/backend.py

@@ -17,7 +17,7 @@ from sglang.srt.layers.attention.minicpm.attention_adapter import (
 )
 from sglang.srt.layers.attention.minicpm.cache import attach_compressed_cache
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_parallel, get_schedule
 from sglang.srt.utils import is_blackwell_supported, next_power_of_2
 
 if TYPE_CHECKING:
@@ -273,7 +273,7 @@ class MiniCPMSparseBackend(AttentionBackend):
             "local_blocks": self.local_blocks,
             "dtype_str": dtype_str,
         }
-        chunked_prefill_size = model_runner.server_args.chunked_prefill_size
+        chunked_prefill_size = get_schedule().chunked_prefill_size
         if self.minicpm_fuse_topk and chunked_prefill_size <= 0:
             raise ValueError(
                 "MiniCPM fused top-k requires a positive --chunked-prefill-size."