Fix --hicache-size allocating ~2x host memory on hybrid SWA (#32373)

Co-authored-by: cctry <cctry@fb.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>

test/registered/unit/mem_cache/test_hybrid_pool_assembler.py

@@ -0,0 +1,31 @@
+"""Unit test for hybrid HiCache fixed-size budget splitting."""
+
+import unittest
+
+from sglang.srt.mem_cache.hybrid_cache.hybrid_pool_assembler import _split_hicache_size
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=1, suite="base-a-test-cpu")
+
+
+class _Pool:
+    def __init__(self, kv_bytes):
+        self._kv_bytes = kv_bytes
+
+    def get_kv_size_bytes(self):
+        return self._kv_bytes
+
+
+class TestSplitHicacheSize(CustomTestCase):
+    def test_splits_total_budget_by_device_bytes(self):
+        # scalar and (k, v) tuple return shapes both supported
+        shares = _split_hicache_size(
+            100, (_Pool(75 * 10**9), _Pool((15 * 10**9, 10 * 10**9)))
+        )
+        self.assertEqual(shares, (75.0, 25.0))  # proportional to device KV bytes
+        self.assertEqual(sum(shares), 100)  # total budget preserved, not doubled
+
+
+if __name__ == "__main__":
+    unittest.main()