[Bugfix][Spec Decode] Auto-enable async scheduling for draft models (#48341)

Signed-off-by: BWAAEEEK <jooho414@gmail.com>
This commit is contained in:
JooHo Lee
2026-08-06 13:51:40 +08:00
committed by GitHub
parent b976e980ef
commit c0202c5603
3 changed files with 25 additions and 39 deletions
+23
View File
@@ -407,6 +407,29 @@ def test_reconfigure_for_independent_dp_rank_on_multinode_dense_model():
assert parallel_config.world_size == 8
def test_draft_model_enables_async_scheduling_by_default():
parallel_config = ParallelConfig(distributed_executor_backend="uni")
model_config = ModelConfig("Qwen/Qwen3-0.6B", max_model_len=2048)
speculative_config = SpeculativeConfig(
method="draft_model",
model="Qwen/Qwen3-0.6B",
num_speculative_tokens=3,
target_model_config=model_config,
target_parallel_config=parallel_config,
)
cfg = VllmConfig(
model_config=model_config,
scheduler_config=SchedulerConfig(
max_model_len=2048,
is_encoder_decoder=False,
),
parallel_config=parallel_config,
speculative_config=speculative_config,
)
assert cfg.scheduler_config.async_scheduling is True
@dataclass
class _TestConfigFields:
a: int
@@ -25,10 +25,6 @@ from ..utils import (
)
class AsyncSchedulingNotEnabledError(AssertionError):
"""Raised when draft-model spec decode does not enable async scheduling."""
@dataclass
class ArgsTest:
target_model: str
@@ -93,22 +89,12 @@ cases = [
@pytest.mark.parametrize("args", cases)
@pytest.mark.parametrize("enforce_eager", [True, False])
@single_gpu_only
# TODO: Fix async_scheduling & engine initialization issues - see https://github.com/vllm-project/vllm/issues/38929
@pytest.mark.xfail(
raises=AsyncSchedulingNotEnabledError,
reason="draft_model does not yet enable async_scheduling: issue #38929",
)
def test_draft_model_correctness(args: ArgsTest, enforce_eager: bool):
args.enforce_eager = enforce_eager
assert_draft_model_correctness(args)
@single_gpu_only
# TODO: Fix async_scheduling and engine initialization issues - see https://github.com/vllm-project/vllm/issues/38929
@pytest.mark.xfail(
raises=AsyncSchedulingNotEnabledError,
reason="draft_model does not yet enable async_scheduling: issue #38929",
)
def test_draft_model_realistic_example():
args = ArgsTest(
target_model="Qwen/Qwen3-1.7B",
@@ -124,11 +110,6 @@ def test_draft_model_realistic_example():
@single_gpu_only
# TODO: Fix async_scheduling and engine initialization issues - see https://github.com/vllm-project/vllm/issues/38929
@pytest.mark.xfail(
raises=AsyncSchedulingNotEnabledError,
reason="draft_model does not yet enable async_scheduling: issue #38929",
)
def test_draft_model_parallel_drafting():
args = ArgsTest(
target_model="Qwen/Qwen3-1.7B",
@@ -155,11 +136,6 @@ def test_draft_model_parallel_drafting():
)
@pytest.mark.parametrize("enforce_eager", [True, False])
@single_gpu_only
# TODO: Fix async_scheduling and engine initialization issues - see https://github.com/vllm-project/vllm/issues/38929
@pytest.mark.xfail(
raises=AsyncSchedulingNotEnabledError,
reason="draft_model does not yet enable async_scheduling: issue #38929",
)
def test_draft_model_quantization(models: tuple[str, str], enforce_eager: bool):
tgt_model, draft_model = models
sd_case = ArgsTest(
@@ -172,11 +148,6 @@ def test_draft_model_quantization(models: tuple[str, str], enforce_eager: bool):
@multi_gpu_only(num_gpus=2)
# TODO: Fix async_scheduling and engine initialization issues - see https://github.com/vllm-project/vllm/issues/38929
@pytest.mark.xfail(
raises=AsyncSchedulingNotEnabledError,
reason="draft_model does not yet enable async_scheduling: issue #38929",
)
def test_draft_model_tensor_parallelism():
"""Ensure spec decode works when running with TP > 1."""
_skip_if_insufficient_gpus_for_tp(2)
@@ -372,17 +343,8 @@ def assert_draft_model_correctness(args: ArgsTest):
assert acceptance_rate >= args.expected_acceptance_rate
assert acceptance_len >= args.expected_acceptance_len
# draft_model supports async scheduling; assert it is active by default.
# Raise AsyncSchedulingNotEnabledError (a subclass of AssertionError) so that
# @pytest.mark.xfail(raises=AsyncSchedulingNotEnabledError) catches only this
# specific failure — leaving all other assertion failures (e.g. correctness or
# acceptance-rate checks above) visible as real test failures.
has_async = spec_llm.llm_engine.vllm_config.scheduler_config.async_scheduling
del spec_llm # CLEANUP
torch.accelerator.empty_cache()
cleanup_dist_env_and_memory()
if not has_async:
raise AsyncSchedulingNotEnabledError(
"Expected async_scheduling=True for draft_model spec decode, got False."
" See https://github.com/vllm-project/vllm/issues/38929"
)
assert has_async, "Expected async_scheduling=True for draft_model spec decode"
+1
View File
@@ -1126,6 +1126,7 @@ class VllmConfig:
self.speculative_config is not None
and self.speculative_config.method not in get_args(EagleModelTypes)
and self.speculative_config.method not in get_args(NgramGPUTypes)
and self.speculative_config.method != "draft_model"
and self.speculative_config.method != "dspark"
):
logger.warning_once(