perf(qwen3-tts): cache decoder masks and compile pre-transformer (#5107)
Signed-off-by: Ziyang Zhang <hafuhafu@qq.com> Co-authored-by: Sy03 <1370724210@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
3 files changed
+249
-40
No files matched your search
@@ -12,15 +12,21 @@ import os
|
||||
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
|
||||
|
||||
import pytest
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
from tests.helpers.mark import hardware_test
|
||||
from tests.helpers.runtime import OmniServerParams
|
||||
from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
|
||||
from tests.helpers.stage_config import (
|
||||
get_deploy_config_path,
|
||||
get_deploy_config_stage,
|
||||
modify_stage_config,
|
||||
)
|
||||
|
||||
MODEL = "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice"
|
||||
|
||||
_DEFAULT_STAGE_CONFIG = get_deploy_config_path("qwen3_tts.yaml")
|
||||
_STAGE_CONFIG = modify_stage_config(
|
||||
get_deploy_config_path("qwen3_tts.yaml"),
|
||||
_DEFAULT_STAGE_CONFIG,
|
||||
updates={"stages": {0: {"default_sampling_params.max_tokens": 500}}},
|
||||
)
|
||||
|
||||
@@ -50,6 +56,39 @@ tts_server_params = [
|
||||
)
|
||||
]
|
||||
|
||||
default_tts_server_params = [
|
||||
pytest.param(
|
||||
OmniServerParams(
|
||||
model=MODEL,
|
||||
stage_config_path=_DEFAULT_STAGE_CONFIG,
|
||||
server_args=["--trust-remote-code"],
|
||||
),
|
||||
id="async_chunk",
|
||||
)
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.advanced_model
|
||||
@pytest.mark.tts
|
||||
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA Graph startup test requires CUDA")
|
||||
@hardware_test(res={"cuda": "L4"}, num_cards=1)
|
||||
@pytest.mark.parametrize("omni_server", default_tts_server_params, indirect=True)
|
||||
def test_default_cuda_graph_startup(omni_server) -> None:
|
||||
"""Verify both stages start with the shipped CUDA Graph configuration.
|
||||
|
||||
The fixture reaching this test is the smoke assertion: it waits for the
|
||||
server to become ready after both stages finish model initialization and
|
||||
CUDA Graph capture. The regression covered here exited during stage 1
|
||||
capture, before fixture setup could complete.
|
||||
"""
|
||||
for stage_id in (0, 1):
|
||||
stage = get_deploy_config_stage("qwen3_tts.yaml", stage_id=stage_id)
|
||||
assert stage.get("enforce_eager", False) is False
|
||||
|
||||
assert omni_server.proc is not None
|
||||
assert omni_server.proc.poll() is None
|
||||
|
||||
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.advanced_model
|
||||
|
||||
Reference in new issue
Block a user