[None][chore] BREAKING: Test KV cache manager v2 as the default

examples/disaggregated/slurm/service_discovery_example/launch.slurm

@@ -44,16 +44,14 @@ cat >${work_path}/ctx_config.yaml << EOL
 disable_overlap_scheduler: True
 internal_request_auth_key: ${internal_request_auth_key}
 cache_transceiver_config:
-  backend: UCX
-  transceiver_runtime: CPP
+  backend: NIXL
   max_tokens_in_buffer: 2048
 EOL
 
 cat >${work_path}/gen_config.yaml << EOL
 internal_request_auth_key: ${internal_request_auth_key}
 cache_transceiver_config:
-  backend: UCX
-  transceiver_runtime: CPP
+  backend: NIXL
   max_tokens_in_buffer: 2048
 EOL
 

examples/disaggregated/slurm/simple_example/ctx_extra-llm-api-config.yaml

@@ -2,6 +2,5 @@
 # not yet supported in disaggregated context server architectures.
 disable_overlap_scheduler: True
 cache_transceiver_config:
-  backend: UCX
-  transceiver_runtime: CPP
+  backend: NIXL
   max_tokens_in_buffer: 2048

examples/disaggregated/slurm/simple_example/gen_extra-llm-api-config.yaml

@@ -1,4 +1,3 @@
 cache_transceiver_config:
-  backend: UCX
-  transceiver_runtime: CPP
+  backend: NIXL
   max_tokens_in_buffer: 2048

examples/dwdp/reproduce.py

@@ -220,8 +220,7 @@ def build_worker_config(experiment: Dict[str, Any]) -> Dict[str, Any]:
             "backend": "CUTEDSL",
         },
         "cache_transceiver_config": {
-            "backend": "UCX",
-            "transceiver_runtime": "CPP",
+            "backend": "NIXL",
             "max_tokens_in_buffer": max_tokens_in_buffer,
         },
         "num_postprocess_workers": 4,
@@ -256,8 +255,7 @@ def build_worker_config(experiment: Dict[str, Any]) -> Dict[str, Any]:
             "free_gpu_memory_fraction": 0.3,
         },
         "cache_transceiver_config": {
-            "backend": "UCX",
-            "transceiver_runtime": "CPP",
+            "backend": "NIXL",
             "max_tokens_in_buffer": max_tokens_in_buffer,
         },
         "moe_config": {

examples/llm-api/quickstart_advanced.py

@@ -151,11 +151,11 @@ def add_llm_args(parser):
                         action='store_true')
     parser.add_argument(
         '--use_kv_cache_manager_v2',
-        default='auto',
+        default=True,
         type=_parse_kv_cache_manager_v2,
         metavar='{auto,true,false}',
         help=
-        'Whether to use KVCacheManagerV2 for KV cache management (PyTorch backend). Defaults to model-specific auto selection.',
+        'Whether to use KVCacheManagerV2 for KV cache management (PyTorch backend). Defaults to true; use auto for model-specific selection.',
     )
 
     # Runtime

examples/ray_orchestrator/disaggregated/disagg_serving_local.sh

@@ -8,7 +8,7 @@ ATTACH_MODE=false
 MODEL_DIR="TinyLlama/TinyLlama-1.1B-Chat-v1.0"
 TP_SIZE=1
 TRANSCEIVER_BACKEND="NIXL"
-TRANSCEIVER_RUNTIME="CPP"
+TRANSCEIVER_RUNTIME="PYTHON"
 USAGE="Usage: $0 [--executor ray|mpi] [--attach] [--model model_dir] [--tp_size N] [--transceiver_backend UCX|NIXL] [--transceiver_runtime CPP|PYTHON] [--help]"
 
 while [[ $# -gt 0 ]]; do
@@ -45,7 +45,7 @@ while [[ $# -gt 0 ]]; do
             echo "  --model model_dir    Model directory (default: TinyLlama/TinyLlama-1.1B-Chat-v1.0)"
             echo "  --tp_size N          Tensor parallel size (default: 1)"
             echo "  --transceiver_backend UCX|NIXL  Cache-transceiver backend (default: NIXL)"
-            echo "  --transceiver_runtime CPP|PYTHON  Cache transceiver runtime (default: CPP)"
+            echo "  --transceiver_runtime CPP|PYTHON  Cache transceiver runtime (default: PYTHON)"
             echo "  --help, -h           Show this help message"
             exit 0
             ;;

examples/wide_ep/slurm_scripts/kimi-k2-thinking.yaml

@@ -72,8 +72,8 @@ worker_config:
         num_slots: 416
         layer_updates_per_iter: 1
     cache_transceiver_config:
-      backend: UCX
-      transceiver_runtime: CPP
+      backend: NIXL
+      transceiver_runtime: PYTHON
       max_tokens_in_buffer: 8448
     stream_interval: 20
     num_postprocess_workers: 4
@@ -94,7 +94,7 @@ worker_config:
       free_gpu_memory_fraction: 0.75
       dtype: fp8
     cache_transceiver_config:
-      backend: UCX
-      transceiver_runtime: CPP
+      backend: NIXL
+      transceiver_runtime: PYTHON
       max_tokens_in_buffer: 8448
     trust_remote_code: true

tensorrt_llm/_torch/pyexecutor/_util.py

@@ -1399,6 +1399,9 @@ def try_prepare_estimation(self) -> bool:
                 self._skip_est = True
         model_config = self._model_engine.model.model_config
         if model_config.attn_backend == "VANILLA":
+            if (self._is_kv_cache_manager_v2
+                    and 'cp_type' not in self._mapping.cp_config):
+                self._skip_est = True
             estimating_kv_cache = False
             logger.info(
                 "KV cache size estimation is not supported for Vanilla attention backend, disable it."

tensorrt_llm/llmapi/llm_args.py

@@ -4305,7 +4305,7 @@ class KvCacheConfig(StrictBaseModel, PybindMirror):
         description="Configuration for reusable Mamba state snapshots.")
 
     use_kv_cache_manager_v2: bool | Literal["auto"] = Field(
-        default="auto",
+        default=True,
         status="prototype",
         description=
         "Whether to use the KV cache manager v2 (experimental). 'auto' uses "
@@ -4624,17 +4624,17 @@ class CacheTransceiverConfig(StrictBaseModel, PybindMirror):
             "The communication backend type to use for the cache transceiver.")
 
     transceiver_runtime: Optional[Literal["CPP", "PYTHON", "auto"]] = Field(
-        default="auto",
-        description=
-        "The runtime implementation. 'auto' (default) adopts the model's "
-        "preferred runtime when it declares one; otherwise it selects the "
-        "Python transceiver, falling back to the C++ transceiver only when "
-        "this config itself rules it out (non-NIXL backend or a null "
-        "kv_transfer_timeout_ms) — any other incompatibility fails at "
+        default="PYTHON",
+        description=
+        "The runtime implementation. 'PYTHON' (default) selects the Python "
+        "transceiver, while 'CPP' selects the C++ transceiver. 'auto' adopts "
+        "the model's preferred runtime when it declares one; otherwise it "
+        "selects the Python transceiver, falling back to the C++ transceiver "
+        "only when this config itself rules it out (non-NIXL backend or a "
+        "null kv_transfer_timeout_ms) — any other incompatibility fails at "
         "transceiver creation. The fallback is decided independently on "
         "each server and is only logged, not surfaced, so keep context and "
-        "generation server configurations consistent. 'CPP' selects the C++ "
-        "transceiver, 'PYTHON' the Python transceiver. None is equivalent "
+        "generation server configurations consistent. None is equivalent "
         "to 'CPP'. 'auto' is resolved on the PyTorch backend's standard "
         "model-loading path.")
 

tests/integration/defs/accuracy/test_disaggregated_serving.py

@@ -698,7 +698,7 @@ def run_parallel_test(model_name: str,
                       test_sets: List[LlmapiAccuracyTestHarness],
                       ctx_model: str = None,
                       gen_model: str = None,
-                      cache_transceiver_backend: str = "DEFAULT",
+                      cache_transceiver_backend: str = "NIXL",
                       trust_remote_code: bool = False,
                       quant_algo: str = None,
                       kv_cache_quant_algo: str = None,
@@ -896,6 +896,7 @@ def test_auto_dtype_with_helix(self, comms_medium, cuda_graph_config,
             "enable_block_reuse": False,
             "enable_partial_reuse": False,
             "tokens_per_block": 32,
+            "use_kv_cache_manager_v2": False,
         }
         ctx_server_config = {
             "pipeline_parallel_size": 1,
@@ -1390,7 +1391,7 @@ def test_auto_dtype(self, overlap_scheduler, enable_partial_reuse):
             "disable_overlap_scheduler": True,
             "cuda_graph_config": None,
             "cache_transceiver_config": {
-                "backend": "DEFAULT",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 4096
             },
             "kv_cache_config": kv_cache_config,
@@ -1399,7 +1400,7 @@ def test_auto_dtype(self, overlap_scheduler, enable_partial_reuse):
             "disable_overlap_scheduler": overlap_scheduler,
             "cuda_graph_config": None,
             "cache_transceiver_config": {
-                "backend": "DEFAULT",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 4096
             },
             "kv_cache_config": kv_cache_config,
@@ -1432,7 +1433,7 @@ def _test_chunked_prefill_helper(self, *, ctx_pp: int):
             "disable_overlap_scheduler": True,
             "cuda_graph_config": None,
             "cache_transceiver_config": {
-                "backend": "DEFAULT",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 4096
             },
             "enable_chunked_prefill": True,
@@ -1443,7 +1444,7 @@ def _test_chunked_prefill_helper(self, *, ctx_pp: int):
         gen_server_config = {
             "cuda_graph_config": None,
             "cache_transceiver_config": {
-                "backend": "DEFAULT",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 4096
             },
             "max_batch_size": max_batch_size,
@@ -1487,9 +1488,11 @@ def _run_helix_test(self, comms_medium, cuda_graph_config, gen_pp, gen_tp,
             "enable_block_reuse": False,
             "enable_partial_reuse": False,
             "tokens_per_block": 32,
+            "use_kv_cache_manager_v2": False,
         }
         cache_transceiver_config = {
             "backend": "DEFAULT",
+            "transceiver_runtime": "CPP",
             "max_tokens_in_buffer": 8192,
         }
         ctx_server_config = {

tests/integration/defs/accuracy/test_dwdp_disaggregated_serving.py

@@ -1,3 +1,6 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
 """DWDP disaggregated serving accuracy tests.
 
 Separated from test_disaggregated_serving.py to isolate MPI-dependent test
@@ -228,8 +231,7 @@ def test_dwdp_accuracy(self):
                 "tokens_per_block": 32,
             },
             "cache_transceiver_config": {
-                "backend": "UCX",
-                "transceiver_runtime": "CPP",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 8192,
             },
             "moe_config": {
@@ -261,8 +263,7 @@ def test_dwdp_accuracy(self):
                 "tokens_per_block": 32,
             },
             "cache_transceiver_config": {
-                "backend": "UCX",
-                "transceiver_runtime": "CPP",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 8192,
             },
             "moe_config": {
@@ -345,8 +346,7 @@ def test_dwdp_accuracy_contention_opt(self):
                 "tokens_per_block": 32,
             },
             "cache_transceiver_config": {
-                "backend": "UCX",
-                "transceiver_runtime": "CPP",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 8192,
             },
             "moe_config": {
@@ -379,8 +379,7 @@ def test_dwdp_accuracy_contention_opt(self):
                 "tokens_per_block": 32,
             },
             "cache_transceiver_config": {
-                "backend": "UCX",
-                "transceiver_runtime": "CPP",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 8192,
             },
             "moe_config": {
@@ -482,8 +481,7 @@ def test_dwdp_accuracy_mode_b_overlap(self):
                 "tokens_per_block": 32,
             },
             "cache_transceiver_config": {
-                "backend": "UCX",
-                "transceiver_runtime": "CPP",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 8192,
             },
             "moe_config": {
@@ -519,8 +517,7 @@ def test_dwdp_accuracy_mode_b_overlap(self):
                 "tokens_per_block": 32,
             },
             "cache_transceiver_config": {
-                "backend": "UCX",
-                "transceiver_runtime": "CPP",
+                "backend": "NIXL",
                 "max_tokens_in_buffer": 8192,
             },
             "moe_config": {

tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp1cp2_deepseek_v3_lite_bf16_tllm_gen.yaml

@@ -10,6 +10,7 @@ context_servers:
     enable_block_reuse: false
     enable_partial_reuse: false
     tokens_per_block: 32
+    use_kv_cache_manager_v2: false
   tensor_parallel_size: 2
   pipeline_parallel_size: 1
   cache_transceiver_config:
@@ -28,6 +29,7 @@ generation_servers:
     enable_block_reuse: false
     enable_partial_reuse: false
     tokens_per_block: 32
+    use_kv_cache_manager_v2: false
   cache_transceiver_config:
     backend: UCX
     transceiver_runtime: CPP

tests/integration/defs/disaggregated/test_configs/disagg_config_llama4_kv_cache_overflow.yaml

@@ -16,10 +16,12 @@ context_servers:
   kv_cache_config:
     enable_block_reuse: false
     free_gpu_memory_fraction: 0.3
+    use_kv_cache_manager_v2: false
   disable_overlap_scheduler: true
   cuda_graph_config: null
   cache_transceiver_config:
     backend: UCX
+    transceiver_runtime: CPP
     # Intentionally small to reproduce buffer overflow bug
     max_tokens_in_buffer: 2048
 generation_servers:
@@ -37,9 +39,11 @@ generation_servers:
   kv_cache_config:
     enable_block_reuse: false
     free_gpu_memory_fraction: 0.3
+    use_kv_cache_manager_v2: false
   disable_overlap_scheduler: true
   cuda_graph_config: null
   cache_transceiver_config:
     backend: UCX
+    transceiver_runtime: CPP
     # Intentionally small to reproduce buffer overflow bug
     max_tokens_in_buffer: 2048

tests/integration/defs/disaggregated/test_configs/disagg_config_metrics.yaml

@@ -1,3 +1,7 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Pin v1 KVCM and the C++ transceiver for v1 timing metrics and CSV output.
 hostname: localhost
 model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
 free_gpu_memory_fraction: 0.25
@@ -12,6 +16,8 @@ context_servers:
   pipeline_parallel_size: 1
   return_perf_metrics: True
   perf_metrics_max_requests: 1000
+  kv_cache_config:
+    use_kv_cache_manager_v2: false
   cache_transceiver_config:
     backend: DEFAULT
     transceiver_runtime: CPP
@@ -21,6 +27,8 @@ generation_servers:
   pipeline_parallel_size: 1
   return_perf_metrics: True
   perf_metrics_max_requests: 1000
+  kv_cache_config:
+    use_kv_cache_manager_v2: false
   cache_transceiver_config:
     backend: DEFAULT
     transceiver_runtime: CPP