[Core] Ignore infeasible swap requests. (#4557)

tests/basic_correctness/test_preemption.py

@@ -7,6 +7,7 @@ pytest tests/basic_correctness/test_preemption.py`.
 """
 import pytest
 
+from vllm import SamplingParams
 from vllm.core.scheduler import (ARTIFICIAL_PREEMPTION_MAX_CNT,
                                  ENABLE_ARTIFICIAL_PREEMPT)
 
@@ -136,3 +137,87 @@ def test_swap(
             assert hf_output_ids[j] == vllm_output_ids[j], (
                 f"Test{i} output{j}:\nHF: {hf_output_ids}\n"
                 f"vLLM: {vllm_output_ids}")
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [96])
+@pytest.mark.parametrize("beam_width", [4])
+def test_swap_infeasible(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    beam_width: int,
+) -> None:
+    """Verify infeasible swap request will be ignored."""
+    BLOCK_SIZE = 16
+    prefill_blocks = 2
+    decode_blocks = max_tokens // BLOCK_SIZE
+    example_prompts = example_prompts[:1]
+
+    vllm_model = vllm_runner(
+        model,
+        dtype=dtype,
+        swap_space=10,
+        block_size=BLOCK_SIZE,
+        # Since beam search have more than 1 sequence, prefill + decode blocks
+        # are not enough to finish.
+        num_gpu_blocks_override=prefill_blocks + decode_blocks,
+        max_model_len=(prefill_blocks + decode_blocks) * BLOCK_SIZE,
+    )
+    sampling_params = SamplingParams(n=beam_width,
+                                     use_beam_search=True,
+                                     temperature=0.0,
+                                     max_tokens=max_tokens,
+                                     ignore_eos=True)
+    req_outputs = vllm_model.model.generate(
+        example_prompts,
+        sampling_params=sampling_params,
+    )
+    assert (vllm_model.model.llm_engine.scheduler.artificial_preempt_cnt <
+            ARTIFICIAL_PREEMPTION_MAX_CNT)
+    del vllm_model
+    # Verify the request is ignored and not hang.
+    assert req_outputs[0].outputs[0].finish_reason == "length"
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [96])
+def test_preemption_infeasible(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    """Verify infeasible preemption request will be ignored."""
+    BLOCK_SIZE = 16
+    prefill_blocks = 2
+    decode_blocks = max_tokens // BLOCK_SIZE
+    vllm_model = vllm_runner(
+        model,
+        dtype=dtype,
+        block_size=BLOCK_SIZE,
+        # Not enough gpu blocks to complete a single sequence.
+        # preemption should happen, and the sequence should be
+        # ignored instead of hanging forever.
+        num_gpu_blocks_override=prefill_blocks + decode_blocks // 2,
+        max_model_len=((prefill_blocks + decode_blocks // 2) * BLOCK_SIZE),
+    )
+    sampling_params = SamplingParams(max_tokens=max_tokens, ignore_eos=True)
+    req_outputs = vllm_model.model.generate(
+        example_prompts,
+        sampling_params=sampling_params,
+    )
+
+    assert (vllm_model.model.llm_engine.scheduler.artificial_preempt_cnt <
+            ARTIFICIAL_PREEMPTION_MAX_CNT)
+    del vllm_model
+    # Verify the request is ignored and not hang.
+    for req_output in req_outputs:
+        outputs = req_output.outputs
+        assert len(outputs) == 1
+        assert outputs[0].finish_reason == "length"