feat: refactor GMS client memory manager with tiered API (#6549)

a6d970e9 · Schwinn Saereesitthipitak · GitHub · 651ef5b5 · a6d970e9 · a6d970e9
Unverified Commit a6d970e9 authored Feb 25, 2026 by Schwinn Saereesitthipitak Committed by GitHub Feb 25, 2026
2 changed files
--- a/tests/fault_tolerance/gpu_memory_service/utils/sglang.py
+++ b/tests/fault_tolerance/gpu_memory_service/utils/sglang.py
@@ -44,7 +44,7 @@ class SGLangWithGMSProcess(ManagedProcess):
                "gms",
                "--enable-memory-saver",
                "--mem-fraction-static",
-                "0.8",
+                "0.9",
                "--port",
                str(sglang_port),
            ],

--- a/tests/fault_tolerance/gpu_memory_service/utils/vllm.py
+++ b/tests/fault_tolerance/gpu_memory_service/utils/vllm.py
@@ -54,7 +54,7 @@ class VLLMWithGMSProcess(ManagedProcess):
                "gms",
                "--enable-sleep-mode",
                "--gpu-memory-utilization",
-                "0.8",
+                "0.9",
                "--kv-events-config",
                kv_events_cfg,
            ],