[CI] KimiLinear PD in nightlies (#50266)
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai> Co-authored-by: Claude <noreply@anthropic.com>
Nicolò Lucchesi committed
Aug 3, 2026 at 14:20 UTC
ad0bf3963eef4c3cf267f78199324b1b340b6a12
2 files changed
+24
-2
.buildkite/test_areas/disaggregated.yaml
+20
@@ -241,3 +241,23 @@ steps:
241
commands:
242
- bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
243
- bash v1/kv_connector/nixl_integration/run_accuracy_test.sh
244
+
245
+- label: Kimi-Linear-48B-A3B Disaggregated DP EP
246
+ key: kimi-linear-disaggregated
247
+ timeout_in_minutes: 60
248
+ device: h200
249
+ optional: true
250
+ working_dir: "/vllm-workspace/tests"
251
+ num_devices: 4
252
+ env:
253
+ ENABLE_HMA_FLAG: "1"
254
+ DP_EP: "1"
255
+ GPU_MEMORY_UTILIZATION: "0.9"
256
+ VLLM_SSM_CONV_STATE_LAYOUT: "DS"
257
+ PREFILLER_TP_SIZE: "2"
258
+ DECODER_TP_SIZE: "2"
259
+ MODEL_NAMES: "moonshotai/Kimi-Linear-48B-A3B-Instruct"
260
+ VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--mamba-cache-mode align,--max-model-len 32768"
261
+ commands:
262
+ - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
263
+ - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh
tests/v1/kv_connector/nixl_integration/test_accuracy.py
+4
-2
@@ -59,7 +59,8 @@ def test_accuracy():
59
f"model={MODEL_NAME},"
60
f"base_url={BASE_URL}/chat/completions,"
61
f"num_concurrent={NUM_CONCURRENT},"
62
- "tokenizer_backend=huggingface"
62
+ "tokenizer_backend=huggingface,"
63
+ "trust_remote_code=True"
64
)
65
results = lm_eval.simple_evaluate(
66
model="local-chat-completions",
@@ -72,7 +73,8 @@ def test_accuracy():
73
model_args = (
74
f"model={MODEL_NAME},"
75
f"base_url={BASE_URL}/completions,"
75
- f"num_concurrent={NUM_CONCURRENT},tokenized_requests=False"
76
+ f"num_concurrent={NUM_CONCURRENT},tokenized_requests=False,"
77
+ "trust_remote_code=True"
78
)
79
results = lm_eval.simple_evaluate(
80
model="local-completions",