From cc69ac9e7a7d94fac0b94b8d4d19bf62af69df15 Mon Sep 17 00:00:00 2001 From: Xiaowei Wang <100599594+xiaoweiw-nv@users.noreply.github.com> Date: Tue, 3 Feb 2026 17:45:23 +0800 Subject: [PATCH] Warmup before profiling prefill latency for dynamic chunk sizing (#17198) Co-authored-by: Shangming Cai --- python/sglang/srt/managers/scheduler_pp_mixin.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/managers/scheduler_pp_mixin.py b/python/sglang/srt/managers/scheduler_pp_mixin.py index 17c0fb8c3..3d199b230 100644 --- a/python/sglang/srt/managers/scheduler_pp_mixin.py +++ b/python/sglang/srt/managers/scheduler_pp_mixin.py @@ -1232,8 +1232,9 @@ class ChunkSizePredictor: def fit(self, seq_lens: List[int], latencies: List[float]): """Fit quadratic coefficients f(l) = al^2 + bl + c from data points.""" - L = np.array(seq_lens, dtype=np.float64) - T = np.array(latencies, dtype=np.float64) + # Skip the first data point to reduce fitting bias, as the first run is slower without warmup + L = np.array(seq_lens[1:], dtype=np.float64) + T = np.array(latencies[1:], dtype=np.float64) if len(L) < 8: raise ValueError(