[CI] Remove crashing Kimi K2.5 EAGLE3/MTP variants, keep TP8 and TP8+DP8 (#21898)
This commit is contained in:
@@ -10,19 +10,13 @@ from sglang.test.test_utils import ModelLaunchSettings
|
|||||||
register_cuda_ci(est_time=3600, suite="nightly-8-gpu-common", nightly=True)
|
register_cuda_ci(est_time=3600, suite="nightly-8-gpu-common", nightly=True)
|
||||||
|
|
||||||
KIMI_K25_MODEL_PATH = "moonshotai/Kimi-K2.5"
|
KIMI_K25_MODEL_PATH = "moonshotai/Kimi-K2.5"
|
||||||
EAGLE3_DRAFT_MODEL_PATH = "AQ-MedAI/Kimi-K25-eagle3"
|
|
||||||
|
|
||||||
|
|
||||||
class TestKimiK25(unittest.TestCase):
|
class TestKimiK25(unittest.TestCase):
|
||||||
"""Unified test class for Kimi-K2.5 performance and accuracy.
|
"""Unified test class for Kimi-K2.5 performance and accuracy.
|
||||||
|
|
||||||
Two variants:
|
Runs TP=8 with tool/reasoning parsers.
|
||||||
- basic: TP=8 + tool/reasoning parsers
|
Runs BOTH performance test and accuracy test (gsm8k).
|
||||||
- eagle3: TP=8 + EAGLE3 speculative decoding with draft model
|
|
||||||
|
|
||||||
Each variant runs BOTH:
|
|
||||||
- Performance test (using NightlyBenchmarkRunner)
|
|
||||||
- Accuracy test (using run_eval with gsm8k)
|
|
||||||
"""
|
"""
|
||||||
|
|
||||||
def test_kimi_k25(self):
|
def test_kimi_k25(self):
|
||||||
@@ -31,13 +25,6 @@ class TestKimiK25(unittest.TestCase):
|
|||||||
"--trust-remote-code",
|
"--trust-remote-code",
|
||||||
"--tool-call-parser=kimi_k2",
|
"--tool-call-parser=kimi_k2",
|
||||||
"--reasoning-parser=kimi_k2",
|
"--reasoning-parser=kimi_k2",
|
||||||
]
|
|
||||||
eagle3_args = [
|
|
||||||
"--speculative-algorithm=EAGLE3",
|
|
||||||
f"--speculative-draft-model-path={EAGLE3_DRAFT_MODEL_PATH}",
|
|
||||||
"--speculative-num-steps=3",
|
|
||||||
"--speculative-eagle-topk=1",
|
|
||||||
"--speculative-num-draft-tokens=4",
|
|
||||||
"--model-loader-extra-config",
|
"--model-loader-extra-config",
|
||||||
'{"enable_multithread_load": true, "num_threads": 64}',
|
'{"enable_multithread_load": true, "num_threads": 64}',
|
||||||
]
|
]
|
||||||
@@ -57,14 +44,8 @@ class TestKimiK25(unittest.TestCase):
|
|||||||
ModelLaunchSettings(
|
ModelLaunchSettings(
|
||||||
KIMI_K25_MODEL_PATH,
|
KIMI_K25_MODEL_PATH,
|
||||||
tp_size=8,
|
tp_size=8,
|
||||||
extra_args=base_args + eagle3_args,
|
extra_args=base_args + dp_attn_args,
|
||||||
variant="TP8+MTP",
|
variant="TP8+DP8",
|
||||||
),
|
|
||||||
ModelLaunchSettings(
|
|
||||||
KIMI_K25_MODEL_PATH,
|
|
||||||
tp_size=8,
|
|
||||||
extra_args=base_args + dp_attn_args + eagle3_args,
|
|
||||||
variant="TP8+DP8+MTP",
|
|
||||||
),
|
),
|
||||||
]
|
]
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user