From efc99a86ff39643fd0280cbfea1249587d8f8907 Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Thu, 6 Aug 2026 01:40:10 -0700 Subject: [PATCH] [CI] Restore the full prefill CUDA graph capture range in test launches (#33847) --- python/sglang/test/test_utils.py | 9 --------- 1 file changed, 9 deletions(-) diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py index ae7ac405d..2a9c1a68d 100644 --- a/python/sglang/test/test_utils.py +++ b/python/sglang/test/test_utils.py @@ -952,15 +952,6 @@ def popen_launch_server( other_args = list(other_args) 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 if env is None: env = os.environ.copy()