[Nightly] Replace MiniMax-M2 with MiniMax-M2.5 (#20083)

Co-authored-by: Alison Shao <alisonshao@mac.lan>
This commit is contained in:
Alison Shao
2026-03-07 01:15:34 -08:00
committed by GitHub
parent 43d6a32045
commit 1aa6ab41de

View File

@@ -9,32 +9,30 @@ from sglang.test.test_utils import ModelLaunchSettings
# Runs on both H200 and B200 via nightly-8-gpu-common suite
register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
MINIMAX_M2_MODEL_PATH = "MiniMaxAI/MiniMax-M2"
MINIMAX_M25_MODEL_PATH = "MiniMaxAI/MiniMax-M2.5"
class TestMiniMaxM2(unittest.TestCase):
"""Unified test class for MiniMax-M2 performance and accuracy.
class TestMiniMaxM25(unittest.TestCase):
"""Unified test class for MiniMax-M2.5 performance and accuracy.
Single variant with TP=8 + EP=8 configuration.
MiniMax-M2 is a 230B MoE model with 10B active params.
Runs BOTH:
- Performance test (using NightlyBenchmarkRunner with extra_bench_args)
- Accuracy test (using run_eval with mgsm_en)
- Accuracy test (using run_eval with gsm8k)
"""
def test_minimax_m2(self):
"""Run performance and accuracy for MiniMax-M2."""
def test_minimax_m25(self):
"""Run performance and accuracy for MiniMax-M2.5."""
base_args = [
"--tp=8",
"--ep=8",
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--ep=8",
"--mem-fraction-static=0.85",
"--reasoning-parser=minimax-append-think",
]
variants = [
ModelLaunchSettings(
MINIMAX_M2_MODEL_PATH,
MINIMAX_M25_MODEL_PATH,
tp_size=8,
extra_args=base_args,
variant="TP8+EP8",
@@ -43,10 +41,10 @@ class TestMiniMaxM2(unittest.TestCase):
run_combined_tests(
models=variants,
test_name="MiniMax-M2",
test_name="MiniMax-M2.5",
accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.80),
performance_params=PerformanceTestParams(
profile_dir="performance_profiles_minimax_m2",
profile_dir="performance_profiles_minimax_m25",
),
)