EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Fix] Add extra warmup run count for MC2 on specific SoC version (#4843)

Browse Source 
### What this PR does / why we need it?
We didn’t account for this earlier because we didn’t have A3 in CI, but
now that we do, this test case needs a few extra tweaks — please take a
look at `profile_run`.

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
This commit is contained in:
Yizhou
2025-12-09 21:37:38 +08:00
committed by GitHub
parent 4813cefc58
commit cd1c69ee0b
1 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/e2e/multicard/test_aclgraph_capture_replay.py
View File

@@ -25,6 +25,8 @@ import pytest
import torch import torch
from vllm.utils.network_utils import get_open_port from vllm.utils.network_utils import get_open_port
from vllm_ascend.utils import AscendDeviceType, get_ascend_device_type
MODELS = [ MODELS = [
"Qwen/Qwen3-0.6B", "Qwen/Qwen3-0.6B",
"vllm-ascend/DeepSeek-V2-Lite-W8A8", "vllm-ascend/DeepSeek-V2-Lite-W8A8",
@@ -212,6 +214,10 @@ def test_aclgraph_capture_replay_dp2(
# Part A: Warmup runs (Profile run + 2 runs per captured graph) # Part A: Warmup runs (Profile run + 2 runs per captured graph)
warmup_runs = 1 + (2 * max_batch_sizes) warmup_runs = 1 + (2 * max_batch_sizes)
soc_version = get_ascend_device_type()
if soc_version in {AscendDeviceType._910_93} and "DeepSeek" in model:
# An extra warmup run is needed for MC2 warmup here
warmup_runs += 1
# Part B: Alignment padding (Empty runs to hit the 32-step boundary) # Part B: Alignment padding (Empty runs to hit the 32-step boundary)
padding_runs = aligned_steps - total_steps padding_runs = aligned_steps - total_steps