chore: bump mooncake version to 0.3.13 (#36493)

docker/Dockerfile

@@ -15,7 +15,7 @@ ARG UBUNTU_MIRROR
 ARG GITHUB_ARTIFACTORY=github.com
 ARG INSTALL_FLASHINFER_JIT_CACHE=0
 ARG FLASHINFER_VERSION=0.6.17
-ARG MOONCAKE_VERSION=0.3.12.post1
+ARG MOONCAKE_VERSION=0.3.13
 ARG MSCCLPP_VERSION=sglang-v0.9.1
 
 ENV DEBIAN_FRONTEND=noninteractive \

scripts/ci/cuda/ci_install_dependency.sh

@@ -691,7 +691,7 @@ stabilize_flashinfer_jit_paths() {
 }
 
 install_extra_deps() {
-    MOONCAKE_VERSION="0.3.12.post1"
+    MOONCAKE_VERSION="0.3.13"
     NIXL_VERSION="1.3.0"
     # shellcheck source=scripts/ci/utils/sgl_eval_ref.sh
     source "${SCRIPT_DIR}/../utils/sgl_eval_ref.sh"

test/registered/disaggregation/test_kimi_linear_pd_dcp4.py

@@ -58,9 +58,22 @@ class TestKimiLinearPDDCP4(GSM8KMixin, PDDisaggregationServerBase):
     @classmethod
     def setUpClass(cls):
         super().setUpClass()
+        # Mooncake >=0.3.13 (kvcache-ai/Mooncake#2974) bounds TCP admission per
+        # peer and hard-fails the overflow instead of applying backpressure.
+        # This test's TP4/EP4 -> TP4/DCP4 fan-out blows past the 1024 + 1024
+        # defaults on hosts that fall back to TCP, killing every KV transfer.
+        # Set before launch_all() so the server subprocesses inherit it.
+        os.environ["MC_TCP_MAX_QUEUED_TRANSFERS_PER_PEER"] = "65535"
+        os.environ["MC_TCP_MAX_PENDING_ADMISSIONS_PER_PEER"] = "65535"
         cls._collect_monolithic_references()
         cls.launch_all()
 
+    @classmethod
+    def tearDownClass(cls):
+        os.environ.pop("MC_TCP_MAX_QUEUED_TRANSFERS_PER_PEER")
+        os.environ.pop("MC_TCP_MAX_PENDING_ADMISSIONS_PER_PEER")
+        super().tearDownClass()
+
     @classmethod
     def _monolithic_reference_args(cls):
         return [