From b7eccd642f59b4313183cae9b803de07e789407f Mon Sep 17 00:00:00 2001 From: Lianmin Zheng Date: Sun, 16 Aug 2026 15:31:36 -0700 Subject: [PATCH] Increase post-capture decode memory reserve (#34996) --- python/sglang/srt/server_args.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index bebbf3194..18d16fa28 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -4913,7 +4913,7 @@ class ServerArgs: if self.post_capture_kv_sizing_planned(): # Post-capture sizing measures free memory after graph capture, so # skip the graph/activation reserve; keep only the floor + parallel slack. - reserved_mem = 512 + reserved_mem = 1536 reserved_mem += self.tp_size * self.pp_size / 8 * 1024 else: # Tokens the activation working set scales with (per serving mode).