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

[CI] Restore the full prefill CUDA graph capture range in test launches (#33847)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-08-06 01:40:10 -07:00
committed by GitHub
parent 0e584529f5
commit efc99a86ff
1 changed files with 0 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/test/test_utils.py
-9
View File
@@ -952,15 +952,6 @@ def popen_launch_server(
other_args = list(other_args) other_args = list(other_args)
other_args += ["--device", str(device)] other_args += ["--device", str(device)]
# Prefill dominates capture time: the bucket list runs to chunked_prefill_size
# (8192 on H100-class GPUs) and its largest buckets cost seconds each, while
# 97% of CI prefill batches are under 1024 tokens -- a server that serves one
# test file captures the rest and never replays it. Decode is left alone: its
# capture cost is per-phase, not per-bucket. Pass the flag to opt out.
prefill_flag = "--cuda-graph-max-bs-prefill"
if not any(str(arg).startswith(prefill_flag) for arg in other_args):
other_args = list(other_args) + [prefill_flag, "1024"]
# CI-specific: Validate cache and enable offline mode if complete # CI-specific: Validate cache and enable offline mode if complete
if env is None: if env is None:
env = os.environ.copy() env = os.environ.copy()