From 19f4859b30843f2df3a9c0db1859f10591c2a8ec Mon Sep 17 00:00:00 2001
From: Liangsheng Yin <hnyls2002@gmail.com>
Date: Fri, 17 Jul 2026 02:11:41 -0700
Subject: [PATCH] [CI] Exclude current process memory from GPU idle check
 (#31571)

---
 python/sglang/test/test_utils.py | 27 ++++++++++++++++++++++++---
 1 file changed, 24 insertions(+), 3 deletions(-)

diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py
index cf95676d2..773b1c888 100644
--- a/python/sglang/test/test_utils.py
+++ b/python/sglang/test/test_utils.py
@@ -768,6 +768,11 @@ def _subprocess_popen_with_outputs(
     env: Optional[dict],
     return_stdout_stderr: Optional[tuple],
 ) -> subprocess.Popen:
+    # Release allocator-cached GPU memory to the driver before spawning a
+    # server: cached blocks stay cudaMalloc'd and shrink the child's memory.
+    if torch.cuda.is_initialized():
+        torch.cuda.empty_cache()
+
     if not return_stdout_stderr:
         return subprocess.Popen(command, stdout=None, stderr=None, env=env)
 
@@ -2201,6 +2206,7 @@ def _visible_gpu_indices(pynvml) -> List[int]:
 
 
 def _collect_busy_gpu_reports(pynvml, gpu_indices: List[int]) -> List[str]:
+    self_pid = os.getpid()
     reports = []
     for index in gpu_indices:
         handle = pynvml.nvmlDeviceGetHandleByIndex(index)
@@ -2209,10 +2215,25 @@ def _collect_busy_gpu_reports(pynvml, gpu_indices: List[int]) -> List[str]:
             continue
         try:
             procs = pynvml.nvmlDeviceGetComputeRunningProcesses(handle)
-            proc_info = ", ".join(
-                f"pid={proc.pid} {_format_gib(proc.usedGpuMemory)}" for proc in procs
-            )
         except pynvml.NVMLError:
+            procs = None
+        if procs is not None:
+            # Discount our own usage: the caching allocator retains memory
+            # across test classes, and waiting on ourselves never succeeds.
+            self_used = sum(
+                proc.usedGpuMemory or 0 for proc in procs if proc.pid == self_pid
+            )
+            if used_bytes - self_used < _GPU_IDLE_USED_MEMORY_THRESHOLD:
+                continue
+            proc_info = ", ".join(
+                f"pid={proc.pid} {_format_gib(proc.usedGpuMemory)}"
+                for proc in procs
+                if proc.pid != self_pid
+            ) or (
+                f"no other compute processes;"
+                f" self pid={self_pid} holds {_format_gib(self_used)}"
+            )
+        else:
             proc_info = ""
         reports.append(
             f"GPU {index} uses {_format_gib(used_bytes)}"