[PD] Allow decode radix cache and HiCache L1/L2 with DCP (#40263)

Co-authored-by: Cursor <cursoragent@cursor.com>

python/sglang/srt/arg_groups/pd_disaggregation_hook.py

@@ -63,16 +63,6 @@ def handle_pd_disaggregation(server_args: ServerArgs) -> None:
                 "mooncake, nixl, or fake for synthetic benchmarking, got "
                 f"{cfg.disaggregation_transfer_backend!r}."
             )
-        if cfg.disaggregation_decode_enable_radix_cache:
-            raise ValueError(
-                "PD decode DCP currently requires chunk cache; "
-                "--disaggregation-decode-enable-radix-cache is not supported."
-            )
-        if cfg.enable_hierarchical_cache:
-            raise ValueError(
-                "PD decode DCP currently requires chunk cache; "
-                "--enable-hierarchical-cache is not supported."
-            )
 
     if cfg.disaggregation_mode == "decode":
         if cfg.disaggregation_decode_enable_radix_cache:
@@ -86,7 +76,7 @@ def handle_pd_disaggregation(server_args: ServerArgs) -> None:
                     "--disaggregation-decode-enable-radix-cache is incompatible "
                     "with --disaggregation-transfer-backend fake"
                 )
-            if cfg.speculative_algorithm is not None:
+            if cfg.speculative_algorithm not in (None, "DSPARK"):
                 raise ValueError(
                     "--disaggregation-decode-enable-radix-cache is incompatible "
                     "with speculative decoding "

python/sglang/srt/mem_cache/kv_cache_builder.py

@@ -257,8 +257,8 @@ def build_kv_cache(
 
     # Decode-side radix cache supports SWA only through the unified tree, whose
     # component pools preserve the full-attention prefix while transferring the
-    # SWA window fresh. The legacy SWA cache and hybrid SSM pools remain
+    # SWA window fresh. Hybrid SSM/KDA uses UnifiedRadixCache's Mamba
-    # incompatible with the prefix-match-and-lock allocation path.
+    # component (match + lock + CoW), the same path as colocated serving.
     if (
         get_disagg().disaggregation_decode_enable_radix_cache
         and get_disagg().disaggregation_mode == "decode"
@@ -287,11 +287,6 @@ def build_kv_cache(
                     "--disaggregation-decode-enable-radix-cache does not support "
                     "SWA-compress models (e.g. Gemma4 / MiMo-V2) yet."
                 )
-        if is_hybrid_ssm:
-            raise ValueError(
-                "--disaggregation-decode-enable-radix-cache is incompatible "
-                "with Mamba/SSM models"
-            )
 
     effective_chunked_prefill_size = get_schedule().chunked_prefill_size
     if model_config.is_multimodal and uses_transformers_backend:

test/registered/unit/server_args/test_server_args.py

@@ -1039,28 +1039,6 @@ class TestLoadBalanceMethod(unittest.TestCase):
         )
         self.assertTrue(resolution_result(server_args, "disable_radix_cache"))
 
-    def test_pd_decode_dcp_rejects_radix_cache(self):
-        server_args = ServerArgs(
-            model_path="dummy",
-            disaggregation_mode="decode",
-            disaggregation_transfer_backend="nixl",
-            disaggregation_decode_enable_radix_cache=True,
-            dcp_size=4,
-        )
-        with self.assertRaisesRegex(ValueError, "currently requires chunk cache"):
-            handle_pd_disaggregation(server_args)
-
-    def test_pd_decode_dcp_rejects_hierarchical_cache(self):
-        server_args = ServerArgs(
-            model_path="dummy",
-            disaggregation_mode="decode",
-            disaggregation_transfer_backend="nixl",
-            enable_hierarchical_cache=True,
-            dcp_size=4,
-        )
-        with self.assertRaisesRegex(ValueError, "--enable-hierarchical-cache"):
-            handle_pd_disaggregation(server_args)
-
     def test_pd_decode_radix_cache_rejects_hisparse(self):
         server_args = ServerArgs(
             model_path="dummy",