A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 04:30 UTC
[CI][PD] Add Hybrid SSM integration tests to CI (#37657)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
Mar 23, 2026 at 16:58 UTC
1cbbcfe8a334bab004c43a60be201c8ab528e0d2
3 files changed
+14
-2
.buildkite/test_areas/distributed.yaml
+11
@@ -257,6 +257,17 @@ steps:
257
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
258
- CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
259
260
+- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs)
261
+ timeout_in_minutes: 20
262
+ working_dir: "/vllm-workspace/tests"
263
+ num_devices: 4
264
+ source_file_dependencies:
265
+ - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
266
+ - tests/v1/kv_connector/nixl_integration/
267
+ commands:
268
+ - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
269
+ - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
270
+
271
- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
272
timeout_in_minutes: 30
273
device: a100
tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+2
-2
@@ -19,9 +19,9 @@ dp_ep_configs=(
19
"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)
20
)
21
hybrid_ssm_configs=(
22
- "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"
22
+ "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"
23
# TODO: (NickLucche) Address async scheduling issue with TP>1 separately as this may impact other models.
24
- "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"
24
+ "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"
25
)
26
27
# Select config array based on DP_EP env var
tests/v1/kv_connector/nixl_integration/test_accuracy.py
+1
@@ -19,6 +19,7 @@ EXPECTED_VALUES = {
19
"deepseek-ai/DeepSeek-V2-Lite-Chat": 0.65,
20
"google/gemma-3-4b-it": 0.74,
21
"nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8": 0.84,
22
+ "ibm-granite/granite-4.0-h-tiny": 0.80,
23
}
24
25
SIMPLE_PROMPT = (