From 270c5cb8cd01c4bb852e4f43a5e15bd10ddb39f6 Mon Sep 17 00:00:00 2001
From: LoganJane <42287016+LoganJane@users.noreply.github.com>
Date: Thu, 19 Mar 2026 11:02:29 +0800
Subject: [PATCH] [CI] Add nightly CI test cases for the Kimi-K2.5 (#7416)

### What this PR does / why we need it?
Add nightly CI test cases for the Kimi-K2.5.

- vLLM version: v0.17.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/4497431df654e46fb1fb5e64bf8611e762ae5d87

---------

Signed-off-by: LoganJane <loganJane73@hotmail.com>
Signed-off-by: LoganJane <42287016+LoganJane@users.noreply.github.com>
---
 .../workflows/schedule_nightly_test_a3.yaml   |  3 +
 .../single_node/models/configs/Kimi-K2.5.yaml | 62 +++++++++++++++++++
 2 files changed, 65 insertions(+)
 create mode 100644 tests/e2e/nightly/single_node/models/configs/Kimi-K2.5.yaml

diff --git a/.github/workflows/schedule_nightly_test_a3.yaml b/.github/workflows/schedule_nightly_test_a3.yaml
index e584b196..5d9c2b7c 100644
--- a/.github/workflows/schedule_nightly_test_a3.yaml
+++ b/.github/workflows/schedule_nightly_test_a3.yaml
@@ -261,6 +261,9 @@ jobs:
           - name: kimi-k2-thinking
             os: linux-aarch64-a3-16
             config_file_path: Kimi-K2-Thinking.yaml
+          - name: kimi-k2.5
+            os: linux-aarch64-a3-16
+            config_file_path: Kimi-K2.5.yaml
           - name: minimax-m2-5
             os: linux-aarch64-a3-16
             config_file_path: MiniMax-M2.5-A3.yaml
diff --git a/tests/e2e/nightly/single_node/models/configs/Kimi-K2.5.yaml b/tests/e2e/nightly/single_node/models/configs/Kimi-K2.5.yaml
new file mode 100644
index 00000000..b6ef8fbb
--- /dev/null
+++ b/tests/e2e/nightly/single_node/models/configs/Kimi-K2.5.yaml
@@ -0,0 +1,62 @@
+# ==========================================
+# Shared Configurations
+# ==========================================
+
+_envs: &envs
+  HCCL_BUFFSIZE: "512"
+  SERVER_PORT: "DEFAULT_PORT"
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  OMP_NUM_THREADS: "1"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_BALANCE_SCHEDULING: "1"
+
+_server_cmd: &server_cmd
+  - "--enable-expert-parallel"
+  - "--tensor-parallel-size"
+  - "8"
+  - "--data-parallel-size"
+  - "2"
+  - "--port"
+  - "$SERVER_PORT"
+  - "--max-model-len"
+  - "8192"
+  - "--max-num-batched-tokens"
+  - "8192"
+  - "--max-num-seqs"
+  - "32"
+  - "--async-scheduling"
+  - "--quantization"
+  - "ascend"
+  - "--trust-remote-code"
+  - "--gpu-memory-utilization"
+  - "0.9"
+  - "--additional-config"
+  - '{"multistream_overlap_shared_expert":true}'
+
+_benchmarks: &benchmarks
+  acc:
+    case_type: accuracy
+    dataset_path: vllm-ascend/gsm8k-lite
+    request_conf: vllm_api_general_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
+    max_out_len: 4096
+    batch_size: 8
+    baseline: 95
+    threshold: 5
+
+# ==========================================
+# ACTUAL TEST CASES
+# ==========================================
+
+test_cases:
+  - name: "Kimi-K2.5-W4A8-TP8-DP2-Case"
+    model: "Eco-Tech/Kimi-K2.5-W4A8"
+    envs:
+      <<: *envs
+    server_cmd: *server_cmd
+    server_cmd_extra:
+      - "--compilation-config"
+      - '{"cudagraph_capture_sizes": [1,2,4,8,16,32], "cudagraph_mode": "FULL_DECODE_ONLY"}'
+    benchmarks:
+      <<: *benchmarks
\ No newline at end of file