Fix Kimi K2.5 dp attention+ spec decoding launch crash (#21391)
This commit is contained in:
@@ -38,11 +38,15 @@ class TestKimiK25(unittest.TestCase):
|
||||
"--speculative-num-steps=3",
|
||||
"--speculative-eagle-topk=1",
|
||||
"--speculative-num-draft-tokens=4",
|
||||
"--mem-frac=0.85",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true, "num_threads": 64}',
|
||||
]
|
||||
|
||||
dp_attn_args = [
|
||||
"--dp=8",
|
||||
"--enable-dp-attention",
|
||||
]
|
||||
|
||||
variants = [
|
||||
ModelLaunchSettings(
|
||||
KIMI_K25_MODEL_PATH,
|
||||
@@ -56,6 +60,12 @@ class TestKimiK25(unittest.TestCase):
|
||||
extra_args=base_args + eagle3_args,
|
||||
variant="TP8+MTP",
|
||||
),
|
||||
ModelLaunchSettings(
|
||||
KIMI_K25_MODEL_PATH,
|
||||
tp_size=8,
|
||||
extra_args=base_args + dp_attn_args + eagle3_args,
|
||||
variant="TP8+DP8+MTP",
|
||||
),
|
||||
]
|
||||
|
||||
run_combined_tests(
|
||||
|
||||
Reference in New Issue
Block a user