Unit test for Hierarchical Caching (#4486)

python/sglang/srt/server_args.py

@@ -173,6 +173,7 @@ class ServerArgs:
     enable_custom_logit_processor: bool = False
     tool_call_parser: str = None
     enable_hierarchical_cache: bool = False
+    hicache_ratio: float = 2.0
     enable_flashinfer_mla: bool = False
     enable_flashmla: bool = False
     flashinfer_mla_disable_ragged: bool = False
@@ -1007,6 +1008,13 @@ class ServerArgs:
             action="store_true",
             help="Enable hierarchical cache",
         )
+        parser.add_argument(
+            "--hicache-ratio",
+            type=float,
+            required=False,
+            default=ServerArgs.hicache_ratio,
+            help="The ratio of the size of host KV cache memory pool to the size of device pool.",
+        )
 
         # Server warmups
         parser.add_argument(