minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph (#18013)

Browse Source
This commit is contained in:
Minglei Zhu
2026-02-01 14:26:26 +08:00
committed by GitHub
parent 11892599f1
commit 38d275a9fd
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/schedule_batch.py
+9 -1
View File
@@ -2283,6 +2283,15 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
def maybe_evict_swa(self): def maybe_evict_swa(self):
if self.tree_cache.supports_swa(): if self.tree_cache.supports_swa():
sliding_window_size = self.tree_cache.sliding_window_size sliding_window_size = self.tree_cache.sliding_window_size
server_args = get_global_server_args()
if (
self.forward_mode.is_decode()
and server_args.enable_piecewise_cuda_graph
and not self.tree_cache.is_chunk_cache()
):
return
for idx, req in enumerate(self.reqs): for idx, req in enumerate(self.reqs):
if self.forward_mode.is_decode(): if self.forward_mode.is_decode():
# We set evict_swa condition here with two reasons: # We set evict_swa condition here with two reasons:
@@ -2297,7 +2306,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
if req.extend_batch_idx < 2: if req.extend_batch_idx < 2:
continue continue
else: else:
server_args = get_global_server_args()
pre_len = ( pre_len = (
pre_len - server_args.chunked_prefill_size pre_len - server_args.chunked_prefill_size
if server_args.chunked_prefill_size > 0 if server_args.chunked_prefill_size > 0