diff --git a/.buildkite/test_areas/misc.yaml b/.buildkite/test_areas/misc.yaml index a427ac5e77..4626f08ef6 100644 --- a/.buildkite/test_areas/misc.yaml +++ b/.buildkite/test_areas/misc.yaml @@ -90,6 +90,7 @@ steps: - tests/v1/kv_offload - tests/v1/simple_kv_offload - tests/v1/worker + - tests/v1/streaming_input - tests/v1/kv_connector/unit - tests/v1/ec_connector/unit - tests/v1/metrics @@ -103,6 +104,7 @@ steps: - pytest -v -s v1/kv_offload - pytest -v -s v1/simple_kv_offload - pytest -v -s v1/worker + - pytest -v -s v1/streaming_input - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit - pytest -v -s -m 'not cpu_test' v1/ec_connector/unit - pytest -v -s -m 'not cpu_test' v1/metrics diff --git a/tests/v1/streaming_input/test_gpu_model_runner_streaming.py b/tests/v1/streaming_input/test_gpu_model_runner_streaming.py index fd619610b7..a8f7221da8 100644 --- a/tests/v1/streaming_input/test_gpu_model_runner_streaming.py +++ b/tests/v1/streaming_input/test_gpu_model_runner_streaming.py @@ -16,7 +16,8 @@ from vllm.sampling_params import SamplingParams from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch from vllm.v1.worker.gpu_model_runner import GPUModelRunner -pytestmark = pytest.mark.cpu_test +# Not cpu_test: InputBatch allocates pinned (UVA) memory, which requires a +# CUDA device even though the batch tensors live on the CPU. @pytest.fixture @@ -28,6 +29,7 @@ def mock_model_runner_with_input_batch(): runner.requests = {} runner.max_num_reqs = 10 runner.max_model_len = 1024 + runner.late_interaction_runner = Mock() # Create a real InputBatch for e2e testing runner.input_batch = InputBatch( diff --git a/tests/v1/streaming_input/test_gpu_model_runner_v2_streaming.py b/tests/v1/streaming_input/test_gpu_model_runner_v2_streaming.py index 8fde0f117c..f598d0547f 100644 --- a/tests/v1/streaming_input/test_gpu_model_runner_v2_streaming.py +++ b/tests/v1/streaming_input/test_gpu_model_runner_v2_streaming.py @@ -16,7 +16,8 @@ from vllm.v1.core.sched.output import ( from vllm.v1.worker.gpu.model_runner import GPUModelRunner from vllm.v1.worker.gpu.states import RequestState -pytestmark = pytest.mark.cpu_test +# Not cpu_test: RequestState allocates pinned (UVA) memory, which requires a +# CUDA device even though the request state itself lives on the CPU. @pytest.fixture @@ -31,13 +32,12 @@ def mock_model_runner_with_req_states(): num_speculative_steps=0, vocab_size=32000, device=torch.device("cpu"), - model_dtype=torch.float32, - cache_draft_logits=False, ) runner.encoder_cache = None runner.model_state = Mock() runner.block_tables = Mock() runner.lora_state = Mock() + runner.pp_handler = None runner.sampler = None runner.prompt_logprobs_worker = None runner.is_last_pp_rank = False diff --git a/tests/v1/streaming_input/test_scheduler_streaming.py b/tests/v1/streaming_input/test_scheduler_streaming.py index 7d680895b8..822b4e49da 100644 --- a/tests/v1/streaming_input/test_scheduler_streaming.py +++ b/tests/v1/streaming_input/test_scheduler_streaming.py @@ -53,6 +53,8 @@ def create_scheduler() -> Scheduler: vllm_config.model_config = MagicMock() vllm_config.model_config.skip_tokenizer_init = True vllm_config.model_config.is_multimodal_model = False + vllm_config.model_config.is_encoder_decoder = False + vllm_config.model_config.is_diffusion = False vllm_config.model_config.max_model_len = 1024 vllm_config.model_config.enable_return_routed_experts = False vllm_config.cache_config = MagicMock() @@ -496,7 +498,9 @@ class TestStreamingScheduler(unittest.TestCase): eco_cycle2 = eco_dict_cycle2[session.client_index].outputs[0] assert eco_cycle2.finish_reason == FinishReason.STOP assert session.status == RequestStatus.WAITING_FOR_STREAMING_REQ - assert session in scheduler.waiting + # Sessions paused for streaming input are blocked-waiting, so they + # live in the skipped_waiting queue rather than the main waiting queue. + assert session in scheduler.skipped_waiting assert session._all_token_ids == [1, 2, 3, 10, STOP_TOKEN] # CRITICAL ASSERTION: Cached prompt_token_ids STILL must not have changed