EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Test no vllm custom allreduce (#4210)

Browse Source
This commit is contained in:
Lianmin Zheng
2025-03-08 05:23:06 -08:00
committed by GitHub
parent 4a893d142d
commit 2cadd51d11
2 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.github/workflows/pr-test.yml vendored
View File

@@ -269,6 +269,8 @@ jobs:
cd test/srt
python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_moe_tp2_bs1
USE_VLLM_CUSTOM_ALLREDUCE=0 python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_moe_tp2_bs1
- name: Benchmark single latency + torch.compile (TP=2)
timeout-minutes: 10
run: |

6
test/srt/test_bench_one_batch.py
View File

@@ -11,7 +11,9 @@ from sglang.test.test_utils import (
class TestBenchOneBatch(unittest.TestCase):
def test_bs1(self):
output_throughput = run_bench_one_batch(DEFAULT_MODEL_NAME_FOR_TEST, [])
output_throughput = run_bench_one_batch(
DEFAULT_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
)
if is_in_ci():
write_github_step_summary(
@@ -22,7 +24,7 @@ class TestBenchOneBatch(unittest.TestCase):
def test_moe_tp2_bs1(self):
output_throughput = run_bench_one_batch(
DEFAULT_MOE_MODEL_NAME_FOR_TEST, ["--tp", "2"]
DEFAULT_MOE_MODEL_NAME_FOR_TEST, ["--tp", "2", "--cuda-graph-max-bs", "2"]
)
if is_in_ci():