[bugfix] Fix prefill tbo disabled when --deepep-mode=auto (#14333)
Co-authored-by: Cheng Wan <wan4ch@gmail.com>
This commit is contained in:
@@ -131,6 +131,8 @@ def prepare_mlp_sync_batch_raw(
|
|||||||
) and not disable_cuda_graph
|
) and not disable_cuda_graph
|
||||||
|
|
||||||
is_extend_in_batch = local_batch.forward_mode.is_extend() if local_batch else False
|
is_extend_in_batch = local_batch.forward_mode.is_extend() if local_batch else False
|
||||||
|
if local_batch is not None:
|
||||||
|
local_batch.is_extend_in_batch = is_extend_in_batch
|
||||||
|
|
||||||
tbo_preparer = TboDPAttentionPreparer()
|
tbo_preparer = TboDPAttentionPreparer()
|
||||||
if len(offload_tags) == 0 and disable_overlap_schedule:
|
if len(offload_tags) == 0 and disable_overlap_schedule:
|
||||||
|
|||||||
@@ -1,3 +1,4 @@
|
|||||||
|
import os
|
||||||
import unittest
|
import unittest
|
||||||
from types import SimpleNamespace
|
from types import SimpleNamespace
|
||||||
|
|
||||||
@@ -224,6 +225,10 @@ class TestTBO(CustomTestCase):
|
|||||||
"--max-running-requests",
|
"--max-running-requests",
|
||||||
"512",
|
"512",
|
||||||
],
|
],
|
||||||
|
env={
|
||||||
|
**os.environ,
|
||||||
|
"SGLANG_TBO_DEBUG": "1",
|
||||||
|
},
|
||||||
)
|
)
|
||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
@@ -351,6 +356,10 @@ class TestMTPWithTBO(CustomTestCase):
|
|||||||
"--max-running-requests",
|
"--max-running-requests",
|
||||||
"128",
|
"128",
|
||||||
],
|
],
|
||||||
|
env={
|
||||||
|
**os.environ,
|
||||||
|
"SGLANG_TBO_DEBUG": "1",
|
||||||
|
},
|
||||||
)
|
)
|
||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
|
|||||||
Reference in New Issue
Block a user