[Core] Ignore infeasible swap requests. (#4557)
This commit is contained in:
@@ -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"
|
||||
|
||||
@@ -224,7 +224,7 @@ def test_swap():
|
||||
|
||||
# Swap seq group from CPU -> GPU.
|
||||
cpu_blocks = block_manager.get_block_table(prompt)
|
||||
assert block_manager.can_swap_in(seq_group)
|
||||
assert block_manager.can_swap_in(seq_group) == AllocStatus.OK
|
||||
before_cpu_blocks = block_manager.get_num_free_cpu_blocks()
|
||||
before_gpu_blocks = block_manager.get_num_free_gpu_blocks()
|
||||
mapping = block_manager.swap_in(seq_group)
|
||||
|
||||
@@ -4,6 +4,7 @@ from unittest.mock import MagicMock
|
||||
import pytest # noqa
|
||||
|
||||
from vllm.config import CacheConfig, SchedulerConfig
|
||||
from vllm.core.interfaces import AllocStatus
|
||||
from vllm.core.scheduler import Scheduler
|
||||
from vllm.sequence import Logprob, SequenceGroup
|
||||
|
||||
@@ -410,7 +411,7 @@ def test_running_prefill_prioritized_over_swap():
|
||||
|
||||
# Add 1 more task. Swap is not possible, so prefill is running.
|
||||
scheduler.block_manager.can_swap_in = MagicMock()
|
||||
scheduler.block_manager.can_swap_in.return_value = False
|
||||
scheduler.block_manager.can_swap_in.return_value = AllocStatus.LATER
|
||||
|
||||
_, seq_group2 = create_dummy_prompt("2", prompt_length=60)
|
||||
scheduler.add_seq_group(seq_group2)
|
||||
@@ -423,7 +424,7 @@ def test_running_prefill_prioritized_over_swap():
|
||||
assert out.scheduled_seq_groups[0].seq_group == seq_group2
|
||||
|
||||
# Now although swap is possible, running prefill is prioritized.
|
||||
scheduler.block_manager.can_swap_in.return_value = True
|
||||
scheduler.block_manager.can_swap_in.return_value = AllocStatus.OK
|
||||
_, out = schedule_and_update_computed_tokens(scheduler)
|
||||
assert len(out.scheduled_seq_groups) == 1
|
||||
# 3 decodes. It is swapped in.
|
||||
|
||||
@@ -791,7 +791,7 @@ def test_schedule_swapped_cannot_swap_in():
|
||||
|
||||
# The last request should be swapped out.
|
||||
scheduler.block_manager.can_swap_in = MagicMock()
|
||||
scheduler.block_manager.can_swap_in.return_value = False
|
||||
scheduler.block_manager.can_swap_in.return_value = AllocStatus.LATER
|
||||
# Since we cannot swap in, none of the requests are swapped in.
|
||||
budget = create_token_budget()
|
||||
remaining_swapped, output = scheduler._schedule_swapped(
|
||||
@@ -803,6 +803,34 @@ def test_schedule_swapped_cannot_swap_in():
|
||||
assert len(output.prefill_seq_groups) == 0
|
||||
|
||||
|
||||
def test_infeasible_swap():
|
||||
scheduler = initialize_scheduler()
|
||||
swapped = deque()
|
||||
policy = PolicyFactory.get_policy(policy_name="fcfs")
|
||||
curr_loras = None
|
||||
blocks_to_swap_out = {}
|
||||
for _ in range(2):
|
||||
_, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
|
||||
scheduler._allocate_and_set_running(seq_group)
|
||||
append_new_token_seq_group(60, seq_group, 1)
|
||||
scheduler._swap_out(seq_group, blocks_to_swap_out)
|
||||
swapped.append(seq_group)
|
||||
|
||||
# The last request should be swapped out.
|
||||
scheduler.block_manager.can_swap_in = MagicMock()
|
||||
scheduler.block_manager.can_swap_in.return_value = AllocStatus.NEVER
|
||||
# Since we cannot swap in, none of the requests are swapped in.
|
||||
budget = create_token_budget()
|
||||
remaining_swapped, output = scheduler._schedule_swapped(
|
||||
swapped, budget, curr_loras, policy)
|
||||
assert len(remaining_swapped) == 0
|
||||
assert len(output.infeasible_seq_groups) == 2
|
||||
assert budget.num_batched_tokens == 0
|
||||
assert budget.num_curr_seqs == 0
|
||||
assert len(output.decode_seq_groups) == 0
|
||||
assert len(output.prefill_seq_groups) == 0
|
||||
|
||||
|
||||
def test_schedule_swapped_blocks_to_copy():
|
||||
scheduler = initialize_scheduler()
|
||||
swapped = deque()
|
||||
|
||||
Reference in New Issue
Block a user