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, 13:15 UTC
[XPU] skip unsupported ut and update test_nixl_connector (#37179)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
liuzhenwei committed
Mar 18, 2026 at 13:32 UTC
86b7e3c95a91f4a97431f8fe4381a6ab524a0cc7
2 files changed
+10
-4
.buildkite/scripts/hardware_ci/run-xpu-test.sh
+3
-3
@@ -40,16 +40,16 @@ docker run \
40
python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
41
python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
42
python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
43
- python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager
43
+ python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192
44
python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2
45
python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
46
cd tests
47
pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
48
pytest -v -s v1/engine
49
pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
50
- pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
50
+ pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
51
pytest -v -s v1/structured_output
52
pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py
53
- pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py
53
+ pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py -k "not (test_register_kv_caches and FLASH_ATTN and True)"
54
pytest -v -s v1/test_serial_utils.py
55
'
tests/v1/kv_connector/unit/test_nixl_connector.py
+7
-1
@@ -1369,7 +1369,13 @@ def test_abort_timeout_on_prefiller(monkeypatch, distributed_executor_backend):
1369
"NIXL_TELEMETRY_ENABLE": "1",
1370
},
1371
}
1372
- ray.init(runtime_env=runtime_env)
1372
+ # On XPU/ROCm, vLLM expects Ray's device key to be "GPU".
1373
+ # Explicitly reserving GPU resources here prevents false negatives
1374
+ # when Ray cannot auto-detect accelerator resources in test envs.
1375
+ ray_init_kwargs: dict[str, Any] = {"runtime_env": runtime_env}
1376
+ if not current_platform.is_cuda():
1377
+ ray_init_kwargs["num_gpus"] = 1
1378
+ ray.init(**ray_init_kwargs)
1379
try:
1380
run_test_and_cleanup()
1381
finally: