From f3df7a7231f352bb712624ace03f15b6a8058fdc Mon Sep 17 00:00:00 2001 From: Micah Williamson Date: Sun, 21 Jun 2026 21:08:44 -0500 Subject: [PATCH] [ROCm][CI] Enable kv_connector unit tests on ROCm (#45955) Signed-off-by: Micah Williamson --- .buildkite/scripts/install-kv-connectors.sh | 5 +++++ .buildkite/test_areas/misc.yaml | 6 ++++++ tests/v1/kv_connector/unit/test_multi_connector.py | 1 + tests/v1/kv_connector/unit/test_offloading_connector.py | 4 ++++ 4 files changed, 16 insertions(+) diff --git a/.buildkite/scripts/install-kv-connectors.sh b/.buildkite/scripts/install-kv-connectors.sh index 34c502e6b9..b1e024709e 100755 --- a/.buildkite/scripts/install-kv-connectors.sh +++ b/.buildkite/scripts/install-kv-connectors.sh @@ -4,6 +4,11 @@ set -euo pipefail +if python3 -c "import torch; raise SystemExit(0 if torch.version.hip is not None else 1)"; then + uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt + exit 0 +fi + REQUIREMENTS_FILE="${KV_CONNECTORS_REQUIREMENTS:-/vllm-workspace/requirements/kv_connectors.txt}" uv pip install --system -r "${REQUIREMENTS_FILE}" diff --git a/.buildkite/test_areas/misc.yaml b/.buildkite/test_areas/misc.yaml index 7db72be7b5..5c98006049 100644 --- a/.buildkite/test_areas/misc.yaml +++ b/.buildkite/test_areas/misc.yaml @@ -105,6 +105,12 @@ steps: # Integration test for streaming correctness (requires special branch). - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine + mirror: + amd: + device: mi325_1 + timeout_in_minutes: 60 + depends_on: + - image-build-amd - label: V1 Others (CPU) key: v1-others-cpu diff --git a/tests/v1/kv_connector/unit/test_multi_connector.py b/tests/v1/kv_connector/unit/test_multi_connector.py index 2d6fa834d2..292c3eea16 100644 --- a/tests/v1/kv_connector/unit/test_multi_connector.py +++ b/tests/v1/kv_connector/unit/test_multi_connector.py @@ -219,6 +219,7 @@ def test_multi_example_connector_consistency(): enforce_eager=True, gpu_memory_utilization=0.5, kv_transfer_config=kv_transfer_config, + async_scheduling=False, ) # Run generation - this should trigger saving KV cache # Use a single prompt to avoid race conditions depending on the order of scheduling diff --git a/tests/v1/kv_connector/unit/test_offloading_connector.py b/tests/v1/kv_connector/unit/test_offloading_connector.py index 7cf5272574..7d139f6ccd 100644 --- a/tests/v1/kv_connector/unit/test_offloading_connector.py +++ b/tests/v1/kv_connector/unit/test_offloading_connector.py @@ -138,6 +138,10 @@ def _wait_for_prefix_cache_reset(llm: LLM) -> None: def _latency_test(llm: LLM, subscriber: MockSubscriber | None): + # TODO: Reintroduce latency test on ROCm once MRV2 supports cross + # layer KV Cache. See https://github.com/vllm-project/vllm/pull/45947 + if current_platform.is_rocm(): + return sampling_params = SamplingParams(max_tokens=1) num_times_cpu_better_than_cold = 0