From ad0bf3963eef4c3cf267f78199324b1b340b6a12 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nicol=C3=B2=20Lucchesi?= Date: Mon, 3 Aug 2026 14:20:41 +0200 Subject: [PATCH] [CI] KimiLinear PD in nightlies (#50266) Signed-off-by: NickLucche Co-authored-by: Claude --- .buildkite/test_areas/disaggregated.yaml | 20 +++++++++++++++++++ .../nixl_integration/test_accuracy.py | 6 ++++-- 2 files changed, 24 insertions(+), 2 deletions(-) diff --git a/.buildkite/test_areas/disaggregated.yaml b/.buildkite/test_areas/disaggregated.yaml index f1a89b3968..dd662168d2 100644 --- a/.buildkite/test_areas/disaggregated.yaml +++ b/.buildkite/test_areas/disaggregated.yaml @@ -241,3 +241,23 @@ steps: commands: - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh + +- label: Kimi-Linear-48B-A3B Disaggregated DP EP + key: kimi-linear-disaggregated + timeout_in_minutes: 60 + device: h200 + optional: true + working_dir: "/vllm-workspace/tests" + num_devices: 4 + env: + ENABLE_HMA_FLAG: "1" + DP_EP: "1" + GPU_MEMORY_UTILIZATION: "0.9" + VLLM_SSM_CONV_STATE_LAYOUT: "DS" + PREFILLER_TP_SIZE: "2" + DECODER_TP_SIZE: "2" + MODEL_NAMES: "moonshotai/Kimi-Linear-48B-A3B-Instruct" + VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--mamba-cache-mode align,--max-model-len 32768" + commands: + - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh + - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh diff --git a/tests/v1/kv_connector/nixl_integration/test_accuracy.py b/tests/v1/kv_connector/nixl_integration/test_accuracy.py index 711f27d209..cb7eef2014 100644 --- a/tests/v1/kv_connector/nixl_integration/test_accuracy.py +++ b/tests/v1/kv_connector/nixl_integration/test_accuracy.py @@ -59,7 +59,8 @@ def test_accuracy(): f"model={MODEL_NAME}," f"base_url={BASE_URL}/chat/completions," f"num_concurrent={NUM_CONCURRENT}," - "tokenizer_backend=huggingface" + "tokenizer_backend=huggingface," + "trust_remote_code=True" ) results = lm_eval.simple_evaluate( model="local-chat-completions", @@ -72,7 +73,8 @@ def test_accuracy(): model_args = ( f"model={MODEL_NAME}," f"base_url={BASE_URL}/completions," - f"num_concurrent={NUM_CONCURRENT},tokenized_requests=False" + f"num_concurrent={NUM_CONCURRENT},tokenized_requests=False," + "trust_remote_code=True" ) results = lm_eval.simple_evaluate( model="local-completions",