Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: treat unknown parent liveness as alive in the llama-server reaper (#15921)

Daniel Hanec7aba0c6cb9
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 480818103b8..f371a2ae56a 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -33865,7 +33865,7 @@ class LlamaCppBackend:
@staticmethod
def _pid_start_identity(pid: int) -> str:
- """True when the parent still runs, so the server is not orphaned; uncertainty reads as alive so a live server is never mistakenly reaped"""
+ """Process start time as a string ('' when unknown); a recycled PID will not match it."""
try:
import psutil
try:
@@ -33886,7 +33886,7 @@ class LlamaCppBackend:
@staticmethod
def _pid_parent_is_alive(pid: int) -> bool:
- """Reaps only true orphans so a second backend cannot kill a live server; recycled PIDs are rejected by start time and name; SIGKILL becomes SIGTERM on Windows"""
+ """True when the parent still runs, so the server is not orphaned; uncertainty reads as alive."""
try:
import psutil
@@ -33911,7 +33911,7 @@ class LlamaCppBackend:
if ppid <= 1:
return False
return Path(f"/proc/{ppid}").exists()
- return False
+ return True # no psutil and no /proc: unknown
@staticmethod
def _unlink_pidfile(path: Path) -> None:
@@ -33944,7 +33944,6 @@ class LlamaCppBackend:
return 0 # never our own pid; leave the record alone
if cls._pid_parent_is_alive(pid):
- # Kills only Unsloth-root or env-override servers, gathering all candidates first so a failed scan kills none; a positive count arms the VRAM settle wait
return 0
# Parent is gone: candidate orphan. Reject a PID recycled to something else.
@@ -34160,7 +34159,6 @@ class LlamaCppBackend:
# -- Ownership check, orphan check, kill ---------------------------
for pid, binary, kill in candidates:
- # atexit: log streams may be closed, so log nothing; stdlib loggers need raiseExceptions off, while the except covers structlog; neither alone is enough
if not _is_signalable_pid(pid):
continue
# Ownership: exact match OR binary under a known root.
diff --git a/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py b/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py
index 7a79659e07b..b05aef81982 100644
--- a/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py
+++ b/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py
@@ -596,6 +596,29 @@ def test_reap_recorded_pid_spares_live_server(tmp_path):
proc.wait(timeout = 5)
[email protected]("platform", ["win32", "darwin"])
+def test_reap_recorded_pid_spares_live_server_without_psutil(tmp_path, monkeypatch, platform):
+ """Unknown parent liveness (no psutil, no /proc) must never reap a live server."""
+ proc = subprocess.Popen([sys.executable, "-c", "import time; time.sleep(30)"])
+ pidfile = tmp_path / "llama-server.pid"
+ pidfile.write_text(str(proc.pid))
+ monkeypatch.setitem(sys.modules, "psutil", None)
+ monkeypatch.setattr(sys, "platform", platform)
+ try:
+ assert LlamaCppBackend._pid_parent_is_alive(proc.pid) is True
+ with (
+ patch.object(LlamaCppBackend, "_server_pidfile_path", staticmethod(lambda: pidfile)),
+ patch.object(LlamaCppBackend, "_pid_is_llama_server", staticmethod(lambda pid: True)),
+ ):
+ n = LlamaCppBackend._reap_recorded_pid()
+ assert n == 0, "unknown parent liveness must not reap a server"
+ assert proc.poll() is None
+ assert pidfile.exists()
+ finally:
+ proc.kill()
+ proc.wait(timeout = 5)
+
+
def test_reap_recorded_pid_skips_pid_reuse(tmp_path):
"""A recorded PID recycled to a non-llama-server must NOT be killed (only the
stale pidfile is cleaned), so the user's vllm/games are never touched."""