[V0 Deprecation] Remove multi-step scheduling (#22138)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
This commit is contained in:
Woosuk Kwon
2025-08-12 20:18:39 -07:00
committed by GitHub
parent e18859298d
commit 71683ca6f6
37 changed files with 57 additions and 3465 deletions
+3 -7
View File
@@ -644,11 +644,9 @@ def test_chunked_prefill_preempt():
assert out.num_batched_tokens == max_num_batched_tokens
@pytest.mark.parametrize("num_scheduler_steps", [1, 5])
def test_chunked_prefill_spec_prefill(num_scheduler_steps):
def test_chunked_prefill_spec_prefill():
"""Verify that the num_lookahead_slots is set appropriately for an all"""
"""prefill batch depending on whether multi-step scheduling is enabled"""
"""or not"""
"""prefill batch."""
block_size = 4
max_seqs = 30
max_model_len = 200
@@ -661,7 +659,6 @@ def test_chunked_prefill_spec_prefill(num_scheduler_steps):
max_model_len,
enable_chunked_prefill=True,
num_lookahead_slots=num_lookahead_slots,
num_scheduler_steps=num_scheduler_steps,
)
cache_config = CacheConfig(block_size, 1.0, 1, "auto")
cache_config.num_cpu_blocks = 16
@@ -679,8 +676,7 @@ def test_chunked_prefill_spec_prefill(num_scheduler_steps):
assert out.num_prefill_groups == 1
assert out.num_batched_tokens == max_num_batched_tokens
print(out.num_lookahead_slots)
assert out.num_lookahead_slots == (0 if (num_scheduler_steps == 1) else
num_lookahead_slots)
assert out.num_lookahead_slots == 0
def test_chunked_prefill_max_seqs():
+4 -20
View File
@@ -6,7 +6,6 @@ import pytest
from tests.conftest import VllmRunner
from tests.core.utils import create_dummy_prompt
from vllm.engine.llm_engine import LLMEngine
from vllm.platforms import current_platform
from vllm.sequence import SequenceGroup
MODEL = "JackFram/llama-160m"
@@ -17,32 +16,19 @@ def add_seq_group_to_engine(engine: LLMEngine, seq_group: SequenceGroup):
scheduler.add_seq_group(seq_group)
@pytest.mark.parametrize("num_scheduler_steps", [1, 8])
@pytest.mark.parametrize("enable_chunked_prefill", [False, True])
@pytest.mark.parametrize("enforce_eager", [False, True])
def test_num_computed_tokens_update(num_scheduler_steps: int,
enable_chunked_prefill: bool,
def test_num_computed_tokens_update(enable_chunked_prefill: bool,
enforce_eager: bool):
is_multi_step = num_scheduler_steps > 1
is_multi_step_chunked_prefill = is_multi_step and enable_chunked_prefill
if is_multi_step_chunked_prefill and current_platform.is_rocm():
pytest.skip("Multi-step with Chunked-Prefill does not support "
"rocm_flash_attn backend")
# Make a vllm engine
runner = VllmRunner(model_name=MODEL,
gpu_memory_utilization=0.7,
num_scheduler_steps=num_scheduler_steps,
enable_chunked_prefill=enable_chunked_prefill,
enforce_eager=enforce_eager)
engine: LLMEngine = runner.llm.llm_engine
# In multi-step + chunked-prefill there is no separate single prompt step.
# What is scheduled will run for num_scheduler_steps always.
num_prompt_steps = num_scheduler_steps \
if is_multi_step_chunked_prefill else 1
num_prompt_steps = 1
num_output_tokens_list = [4, 8, 12, 15, 16, 17]
@@ -73,10 +59,8 @@ def test_num_computed_tokens_update(num_scheduler_steps: int,
# Test correctness of num_computed_tokens after the decode steps
assert seq.data.get_num_computed_tokens(
) == prompt_num_computed_tokens + decode_step_counter
for _ in range(num_scheduler_steps):
# decode step
engine.step()
decode_step_counter += 1
engine.step()
decode_step_counter += 1
# Test correctness of num_computed_tokens after the sequence finish.
assert seq.data.get_num_computed_tokens(