biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI][PD] Add Hybrid SSM integration tests to CI (#37657)

Browse Source 
Signed-off-by: NickLucche <nlucches@redhat.com>
This commit is contained in:
Nicolò Lucchesi
2026-03-23 16:58:19 +01:00
committed by GitHub
parent aceadb5ee1
commit 1cbbcfe8a3
3 changed files with 14 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
.buildkite/test_areas/distributed.yaml
View File

@@ -257,6 +257,17 @@ steps:
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
- CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs)
timeout_in_minutes: 20
working_dir: "/vllm-workspace/tests"
num_devices: 4
source_file_dependencies:
- vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
- tests/v1/kv_connector/nixl_integration/
commands:
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
- HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
timeout_in_minutes: 30
device: a100

4
tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
View File

@@ -19,9 +19,9 @@ dp_ep_configs=(
"DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA+P-TP2, D-DPEP=2 (TP=1)
)
hybrid_ssm_configs=(
"ENABLE_HMA_FLAG=1 GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 VLLM_SERVE_EXTRA_ARGS=--max-model-len,8192,--trust-remote-code"
"ENABLE_HMA_FLAG=1 GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=ibm-granite/granite-4.0-h-tiny VLLM_SERVE_EXTRA_ARGS=--max-model-len,8192,--trust-remote-code"
# TODO: (NickLucche) Address async scheduling issue with TP>1 separately as this may impact other models.
"ENABLE_HMA_FLAG=1 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 VLLM_SERVE_EXTRA_ARGS=--max-model-len,8192,--trust-remote-code,--no-async-scheduling"
"ENABLE_HMA_FLAG=1 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=ibm-granite/granite-4.0-h-tiny VLLM_SERVE_EXTRA_ARGS=--max-model-len,8192,--trust-remote-code,--no-async-scheduling"
)
# Select config array based on DP_EP env var

1
tests/v1/kv_connector/nixl_integration/test_accuracy.py
View File

@@ -19,6 +19,7 @@ EXPECTED_VALUES = {
"deepseek-ai/DeepSeek-V2-Lite-Chat": 0.65,
"google/gemma-3-4b-it": 0.74,
"nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8": 0.84,
"ibm-granite/granite-4.0-h-tiny": 0.80,
}
SIMPLE_PROMPT = (