[Fix] Account resident weight memory in KV sizing (#34053)

Signed-off-by: Schwinn Saereesitthipitak <schwinns@nvidia.com>

python/sglang/srt/model_executor/model_runner.py

@@ -844,6 +844,24 @@ class ModelRunner:
             max_rows = max(max_rows, max(capture_bs) * num_tokens_per_req)
         return max_rows
 
+    @property
+    def preloaded_weights_bytes(self) -> int:
+        value = self.loader.preloaded_weights_bytes
+        if isinstance(value, bool) or not isinstance(value, int) or value < 0:
+            raise ValueError(
+                "ModelLoader.preloaded_weights_bytes must be a non-negative int, "
+                f"got {value!r}"
+            )
+        return value
+
+    def account_preloaded_weights(self, preloaded_weights_bytes: int) -> None:
+        # Dist-init sampled B after the daemon already held weights, so slack
+        # (B * (1 - mem_fraction_static)) is too small. Add those bytes back
+        # onto the existing MIN'd baseline. Skip when nothing was preloaded.
+        if preloaded_weights_bytes == 0:
+            return
+        self.pre_model_load_memory += preloaded_weights_bytes / (1 << 30)
+
     def alloc_memory_pool(self, memory_pool_config: Optional[MemoryPoolConfig] = None):
         """Allocate KV cache memory pools only (no backends or cuda graphs)."""
         if memory_pool_config is not None: