Kernels for efficient KV cache IO (#7313)

python/sglang/srt/server_args.py

@@ -217,6 +217,7 @@ class ServerArgs:
     hicache_ratio: float = 2.0
     hicache_size: int = 0
     hicache_write_policy: str = "write_through_selective"
+    hicache_io_backend: str = ""
     flashinfer_mla_disable_ragged: bool = False
     disable_shared_experts_fusion: bool = False
     disable_chunked_prefix_cache: bool = False
@@ -1530,6 +1531,13 @@ class ServerArgs:
             default=ServerArgs.hicache_write_policy,
             help="The write policy of hierarchical cache.",
         )
+        parser.add_argument(
+            "--hicache-io-backend",
+            type=str,
+            choices=["direct", "kernel"],
+            default=ServerArgs.hicache_io_backend,
+            help="The IO backend for KV cache transfer between CPU and GPU",
+        )
         parser.add_argument(
             "--flashinfer-mla-disable-ragged",
             action="store_true",