minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[core] Don't force seq_lens_cpu publication under piecewise CUDA graph (#28633)

Browse Source 
Co-authored-by: jonnykong <jonnykong@fb.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
This commit is contained in:
Jonny Kong
2026-06-19 15:12:07 -07:00
committed by GitHub
co-authored by jonnykong Liangsheng Yin hnyls2002
parent d271de64fe
commit 3ed46f599f
2 changed files with 7 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/unit/server_args/test_server_args.py
-13
View File
@@ -1009,19 +1009,6 @@ class TestPrefillOnlyDisableKvCache(unittest.TestCase):
class TestCudaGraphConfigDataclassAccess(CustomTestCase):
def test_overlap_force_cpu_seq_lens_with_tc_piecewise_prefill(self):
from sglang.srt.managers.overlap_utils import decide_needs_cpu_seq_lens
server_args = SimpleNamespace(
enable_two_batch_overlap=False,
cuda_graph_config=CudaGraphConfig(
prefill=PhaseConfig(backend=Backend.TC_PIECEWISE)
),
)
attn_backend = SimpleNamespace(needs_cpu_seq_lens=False)
self.assertTrue(decide_needs_cpu_seq_lens(server_args, [attn_backend]))
@patch(
"sglang.srt.model_executor.runner_backend."
"tc_piecewise_cuda_graph_backend.get_moe_a2a_backend"