Skip to content

[Misc][v0.13.0]Removes unnecessary graph size re-initialization #6281

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wangxiyuan merged 2 commits into from
Jan 27, 2026
Merged

[Misc][v0.13.0]Removes unnecessary graph size re-initialization #6281

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9c11e94
bugfix_
Angazenn Jan 26, 2026
96615e7
fix
Angazenn Jan 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 1 addition & 4 deletions tests/ut/test_platform.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,15 +230,12 @@ def test_check_and_update_config_enforce_eager_mode(

@patch("vllm_ascend.utils.get_ascend_device_type",
return_value=AscendDeviceType.A3)
@patch("vllm_ascend.utils.update_default_aclgraph_sizes")
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch(
"vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config"
)
def test_check_and_update_config_unsupported_compilation_level(
self, mock_init_recompute, mock_init_ascend, mock_update_default,
mock_soc_version):
mock_update_default.return_value = MagicMock()
self, mock_init_recompute, mock_init_ascend, mock_soc_version):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config(
)
vllm_config = TestNPUPlatform.mock_vllm_config()
Expand Down
15 changes: 6 additions & 9 deletions vllm_ascend/platform.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,11 +30,12 @@
from vllm_ascend.utils import refresh_block_size

# isort: off
from vllm_ascend.utils import (
ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD,
COMPILATION_PASS_KEY, AscendDeviceType, enable_sp, get_ascend_device_type,
update_aclgraph_sizes, update_cudagraph_capture_sizes,
update_default_aclgraph_sizes, check_kv_extra_config)
from vllm_ascend.utils import (ASCEND_QUANTIZATION_METHOD,
COMPRESSED_TENSORS_METHOD, COMPILATION_PASS_KEY,
AscendDeviceType, enable_sp,
get_ascend_device_type, update_aclgraph_sizes,
update_cudagraph_capture_sizes,
check_kv_extra_config)

if TYPE_CHECKING:
from vllm.config import ModelConfig, VllmConfig
Expand Down Expand Up @@ -216,10 +217,6 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:

# set cudaprah sizes before extending `compilation_config.splitting_ops`
vllm_config._set_cudagraph_sizes()
# There are cases where default cudagraph_capture_sizes are not friendly
# to ascend ops && hardwares. We update these sizes here to improve
# default performance.
update_default_aclgraph_sizes(vllm_config)
# TODO delete graph size update here when compilation_config.pass_config.enable_sp
# is supported by vllm-ascend.
if vllm_config.parallel_config.tensor_parallel_size > 1 and not vllm_config.model_config.enforce_eager and \
Expand Down
55 changes: 0 additions & 55 deletions vllm_ascend/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -429,61 +429,6 @@ def update_cudagraph_capture_sizes(vllm_config: VllmConfig,
vllm_config.compilation_config.post_init_cudagraph_sizes()


def _is_default_capture_sizes(vllm_config: VllmConfig) -> bool:
"""
Check whether it is vLLM default capture sizes.
"""

max_cudagraph_capture_size = \
vllm_config.compilation_config.max_cudagraph_capture_size
cudagraph_capture_sizes = [
i for i in [1, 2, 4] if i <= max_cudagraph_capture_size
]
if max_cudagraph_capture_size >= 8:
# Step size 8 for small batch sizes, up to 256(not included)
cudagraph_capture_sizes += list(
range(8, min(max_cudagraph_capture_size + 1, 256), 8))
if max_cudagraph_capture_size >= 256:
# Step size 16 for larger batch sizes
cudagraph_capture_sizes += list(
range(256, max_cudagraph_capture_size + 1, 16))
# in newer version, vLLM use ascending order of cudagraph_capture_sizes.
target_cudagraph_capture_sizes = sorted(cudagraph_capture_sizes)
if target_cudagraph_capture_sizes == \
vllm_config.compilation_config.cudagraph_capture_sizes:
return True

return False


def update_default_aclgraph_sizes(vllm_config: VllmConfig) -> None:
"""
Update ACL graph default capture sizes, so that new sizes
are more friendly to ascend ops && hardware.
"""

if vllm_config.model_config is None or \
vllm_config.model_config.enforce_eager or \
not _is_default_capture_sizes(vllm_config):
return

# modify the default capture_sizes for Qwen3-MoE models on dp settings.
# this is mainly because performance of _npu_paged_attention might degrades
# on special shapes.
# TODO(Angazenn): we will remove this once _npu_paged_attention is fully
# replaced by npu_fused_infer_attention_score which does not contain such bugs.
if vllm_config.model_config and vllm_config.model_config.hf_text_config.model_type == "qwen3_moe" \
and vllm_config.parallel_config.tensor_parallel_size == 1 \
and vllm_config.parallel_config.data_parallel_size > 1 :

max_capture_size = vllm_config.compilation_config.max_cudagraph_capture_size
new_cudagraph_capture_sizes = [1, 2, 5, 10, 15, 20] + [
i for i in range(24, max_capture_size + 1, 8)
]
update_cudagraph_capture_sizes(vllm_config,
new_cudagraph_capture_sizes)


def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
"""Update ACL graph capture sizes based on hardware limitations"""
# NOTE: Currently, we can only capture 1800 graphs at most,
Expand Down