fix(disagg): allow fake transfer with decode DCP

python/sglang/srt/arg_groups/pd_disaggregation_hook.py

@@ -35,10 +35,16 @@ def handle_pd_disaggregation(server_args: ServerArgs) -> None:
         )
 
     if server_args.disaggregation_mode == "decode" and server_args.dcp_size > 1:
-        if server_args.disaggregation_transfer_backend not in ("mooncake", "nixl"):
+        # Fake transfer moves no KV and is only used for synthetic decode
+        # benchmarks, so it does not need the DCP relayout from Mooncake/NIXL.
+        if server_args.disaggregation_transfer_backend not in (
+            "mooncake",
+            "nixl",
+            "fake",
+        ):
             raise ValueError(
                 "PD decode DCP requires --disaggregation-transfer-backend "
-                "mooncake or nixl, got "
+                "mooncake, nixl, or fake for synthetic benchmarking, got "
                 f"{server_args.disaggregation_transfer_backend!r}."
             )
         if server_args.disaggregation_decode_enable_radix_cache:

test/registered/unit/server_args/test_server_args.py

@@ -516,12 +516,22 @@ def test_pd_decode_dcp_rejects_unsupported_transfer_backend(self):
         server_args = ServerArgs(
             model_path="dummy",
             disaggregation_mode="decode",
-            disaggregation_transfer_backend="fake",
+            disaggregation_transfer_backend="mori",
             dcp_size=4,
         )
-        with self.assertRaisesRegex(ValueError, "mooncake or nixl"):
+        with self.assertRaisesRegex(
+            ValueError, "mooncake, nixl, or fake for synthetic benchmarking"
+        ):
             server_args._handle_pd_disaggregation()
 
+    def test_pd_decode_dcp_allows_fake_transfer_backend(self):
+        server_args = self._load_balance_args(
+            disaggregation_mode="decode",
+            disaggregation_transfer_backend="fake",
+            dcp_size=4,
+        )
+        self.assertTrue(server_args.disable_radix_cache)
+
     def test_pd_decode_dcp_rejects_radix_cache(self):
         server_args = ServerArgs(
             model_path="dummy",