From f757d84eae909bebd0a3fc63bae6c65ee6d69e53 Mon Sep 17 00:00:00 2001
From: Jin Li <59594262+liji-nv@users.noreply.github.com>
Date: Fri, 31 Jul 2026 03:01:44 -0700
Subject: [PATCH] [Kimi K2.5][feat] Default to KV cache manager V2

Use the C++ KV cache manager V2 as the model-specific default for Kimi K2.5. This keeps the default scoped to Kimi K2.5 without changing other Kimi implementations or the global LLM arguments.

The existing auto resolution remains responsible for transport compatibility: NIXL with the Python transceiver keeps ManagerV2, while non-NIXL or C++ transceiver routes fall back to the legacy manager unless the user explicitly requests V2.

Validation:

- pre-commit run --files tensorrt_llm/_torch/models/modeling_kimi_k25.py

- git diff --check

- B200 clean sqsh build (Slurm 1569850)

Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
---
 tensorrt_llm/_torch/models/modeling_kimi_k25.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/tensorrt_llm/_torch/models/modeling_kimi_k25.py b/tensorrt_llm/_torch/models/modeling_kimi_k25.py
index a216e916df0f..8eed050eafe2 100644
--- a/tensorrt_llm/_torch/models/modeling_kimi_k25.py
+++ b/tensorrt_llm/_torch/models/modeling_kimi_k25.py
@@ -1517,6 +1517,11 @@ class KimiK25ForConditionalGeneration(PreTrainedModel):
 
     _LANG_PREFIX = "language_model."
 
+    @classmethod
+    def get_model_defaults(cls, llm_args: Any) -> dict:
+        """Use the C++ KV cache manager V2 by default."""
+        return {"kv_cache_config": {"use_kv_cache_manager_v2": True}}
+
     @classmethod
     def get_preferred_transceiver_runtime(
         cls,