[AMD] Fix Janus-Pro crash and add Kimi-K2.5 nightly test (#18269)

This commit is contained in:
Michael
2026-02-10 22:33:13 -08:00
committed by GitHub
parent cd90346a2b
commit d84d2063d3
4 changed files with 250 additions and 10 deletions
@@ -0,0 +1,104 @@
"""AMD Kimi-K2.5 GSM8K Completion Evaluation Test (8-GPU)
Tests moonshotai/Kimi-K2.5 with GSM8K few-shot benchmark on MI325.
Registry: nightly-amd-accuracy-8-gpu-kimi-k25 suite
"""
import os
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
# Register for AMD CI - Kimi K2.5 accuracy test (~60 min)
register_amd_ci(
est_time=3600, suite="nightly-amd-accuracy-8-gpu-kimi-k25", nightly=True
)
KIMI_K25_MODEL_PATH = "moonshotai/Kimi-K2.5"
SERVER_LAUNCH_TIMEOUT = 3600
ACCURACY_THRESHOLD = 0.92
TP_SIZE = 8
class TestKimiK25EvalAMD(CustomTestCase):
"""Kimi-K2.5 GSM8K Completion Evaluation Test for AMD MI325."""
@classmethod
def setUpClass(cls):
cls.model = KIMI_K25_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--tp",
str(TP_SIZE),
"--decode-attention-backend",
"triton",
"--prefill-attention-backend",
"aiter",
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
]
env = os.environ.copy()
env["SGLANG_USE_AITER"] = "1"
env["SGLANG_ROCM_FUSED_DECODE_MLA"] = "0"
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=SERVER_LAUNCH_TIMEOUT,
other_args=other_args,
env=env,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_kimi_k25_gsm8k_accuracy(self):
"""Test Kimi-K2.5 with GSM8K few-shot completion benchmark."""
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
num_shots=8,
data_path=None,
num_questions=1319,
parallel=1319,
max_new_tokens=512,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
acc = metrics["accuracy"]
passed = acc >= ACCURACY_THRESHOLD
status = "✅ PASS" if passed else "❌ FAIL"
print(f" accuracy={acc:.3f} threshold={ACCURACY_THRESHOLD} {status}")
if is_in_ci():
summary = "### Kimi-K2.5 Model (MI325)\n\n"
summary += "| Model | TP | Accuracy | Threshold | Status |\n"
summary += "| ----- | -- | -------- | --------- | ------ |\n"
summary += f"| {KIMI_K25_MODEL_PATH} | {TP_SIZE} | {acc:.3f} | {ACCURACY_THRESHOLD} | {status} |\n"
write_github_step_summary(summary)
self.assertGreaterEqual(
acc,
ACCURACY_THRESHOLD,
f"Kimi-K2.5 accuracy {acc:.3f} below threshold {ACCURACY_THRESHOLD}",
)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,106 @@
"""MI35x Kimi-K2.5 GSM8K Completion Evaluation Test (8-GPU)
Tests moonshotai/Kimi-K2.5 with GSM8K few-shot benchmark on MI35x.
Registry: nightly-amd-accuracy-8-gpu-mi35x-kimi-k25 suite
"""
import os
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
# Register for AMD CI - Kimi K2.5 accuracy test on MI35x (~60 min)
register_amd_ci(
est_time=3600, suite="nightly-amd-accuracy-8-gpu-mi35x-kimi-k25", nightly=True
)
KIMI_K25_MODEL_PATH = "moonshotai/Kimi-K2.5"
SERVER_LAUNCH_TIMEOUT = 3600
ACCURACY_THRESHOLD = 0.92
TP_SIZE = 8
class TestKimiK25EvalMI35x(CustomTestCase):
"""Kimi-K2.5 GSM8K Completion Evaluation Test for AMD MI35x."""
@classmethod
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
def test_kimi_k25_gsm8k_accuracy(self):
"""Test Kimi-K2.5 with GSM8K few-shot completion benchmark."""
other_args = [
"--tp",
str(TP_SIZE),
"--decode-attention-backend",
"triton",
"--prefill-attention-backend",
"aiter",
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--watchdog-timeout",
"1200",
]
env = os.environ.copy()
env["SGLANG_USE_AITER"] = "1"
env["SGLANG_ROCM_FUSED_DECODE_MLA"] = "0"
process = popen_launch_server(
KIMI_K25_MODEL_PATH,
self.base_url,
timeout=SERVER_LAUNCH_TIMEOUT,
other_args=other_args,
env=env,
)
try:
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
num_shots=8,
data_path=None,
num_questions=1319,
parallel=1319,
max_new_tokens=512,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
acc = metrics["accuracy"]
passed = acc >= ACCURACY_THRESHOLD
status = "✅ PASS" if passed else "❌ FAIL"
print(f" accuracy={acc:.3f} threshold={ACCURACY_THRESHOLD} {status}")
if is_in_ci():
summary = "### Kimi-K2.5 Model (MI35x)\n\n"
summary += "| Model | TP | Accuracy | Threshold | Status |\n"
summary += "| ----- | -- | -------- | --------- | ------ |\n"
summary += f"| {KIMI_K25_MODEL_PATH} | {TP_SIZE} | {acc:.3f} | {ACCURACY_THRESHOLD} | {status} |\n"
write_github_step_summary(summary)
self.assertGreaterEqual(
acc,
ACCURACY_THRESHOLD,
f"Kimi-K2.5 accuracy {acc:.3f} below threshold {ACCURACY_THRESHOLD}",
)
finally:
kill_process_tree(process.pid)
if __name__ == "__main__":
unittest.main()