Increase post-capture decode memory reserve (#34996)

2026-08-16 15:31:36 -07:00
parent f61f584347
commit b7eccd642f
1 changed files with 1 additions and 1 deletions
@@ -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).