[ROCm][CI] Enable kv_connector unit tests on ROCm (#45955)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Jun 21, 2026 at 21:08 UTC
f3df7a7231f352bb712624ace03f15b6a8058fdc
4 files changed
+16
.buildkite/scripts/install-kv-connectors.sh
+5
@@ -4,6 +4,11 @@
4
5
set -euo pipefail
6
7
+if python3 -c "import torch; raise SystemExit(0 if torch.version.hip is not None else 1)"; then
8
+ uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
9
+ exit 0
10
+fi
11
+
12
REQUIREMENTS_FILE="${KV_CONNECTORS_REQUIREMENTS:-/vllm-workspace/requirements/kv_connectors.txt}"
13
14
uv pip install --system -r "${REQUIREMENTS_FILE}"
.buildkite/test_areas/misc.yaml
+6
@@ -105,6 +105,12 @@ steps:
105
# Integration test for streaming correctness (requires special branch).
106
- pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
107
- pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
108
+ mirror:
109
+ amd:
110
+ device: mi325_1
111
+ timeout_in_minutes: 60
112
+ depends_on:
113
+ - image-build-amd
114
115
- label: V1 Others (CPU)
116
key: v1-others-cpu
tests/v1/kv_connector/unit/test_multi_connector.py
+1
@@ -219,6 +219,7 @@ def test_multi_example_connector_consistency():
219
enforce_eager=True,
220
gpu_memory_utilization=0.5,
221
kv_transfer_config=kv_transfer_config,
222
+ async_scheduling=False,
223
)
224
# Run generation - this should trigger saving KV cache
225
# Use a single prompt to avoid race conditions depending on the order of scheduling
tests/v1/kv_connector/unit/test_offloading_connector.py
+4
@@ -138,6 +138,10 @@ def _wait_for_prefix_cache_reset(llm: LLM) -> None:
138
139
140
def _latency_test(llm: LLM, subscriber: MockSubscriber | None):
141
+ # TODO: Reintroduce latency test on ROCm once MRV2 supports cross
142
+ # layer KV Cache. See https://github.com/vllm-project/vllm/pull/45947
143
+ if current_platform.is_rocm():
144
+ return
145
sampling_params = SamplingParams(max_tokens=1)
146
147
num_times_cpu_better_than_cold = 0