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: exit the inference worker instead of waiting to be killed (#16338)

Yakshith53624604ef83
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 91b2e4bf7e4..a5cd65fb2b1 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -255,6 +255,26 @@ def _send_response(resp_queue: Any, response: dict) -> None:
logger.error("Failed to send response: %s", exc)
+def _exit_after_shutdown(resp_queue: Any) -> None:
+ """Models are unloaded and acked; interpreter shutdown can stall 10s+ on Windows CUDA (#16330)."""
+ try:
+ resp_queue.close()
+ resp_queue.join_thread()
+ except Exception:
+ pass
+ for stream in (sys.stdout, sys.stderr):
+ try:
+ stream.flush()
+ except Exception:
+ pass
+ try:
+ from utils.worker_stderr import stop_worker_stderr_mirror
+ stop_worker_stderr_mirror()
+ except Exception:
+ pass
+ os._exit(0)
+
+
def _encode_share_object(obj: Any) -> bytes:
data = json.dumps(obj, separators = (",", ":"), ensure_ascii = False).encode("utf-8")
if len(data) > _SHARE_OBJECT_MAX_BYTES:
@@ -2665,7 +2685,7 @@ def run_inference_process(
"type": "shutdown_ack",
},
)
- return
+ _exit_after_shutdown(resp_queue)
else:
logger.warning("Unknown command type: %s", cmd_type)
diff --git a/studio/backend/tests/test_inference_worker_shutdown_exit.py b/studio/backend/tests/test_inference_worker_shutdown_exit.py
new file mode 100644
index 00000000000..a84f53f4d10
--- /dev/null
+++ b/studio/backend/tests/test_inference_worker_shutdown_exit.py
@@ -0,0 +1,86 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+ sys.path.insert(0, str(_BACKEND_ROOT))
+
+from studio.backend.core.inference import worker # noqa: E402
+from utils.worker_stderr import WorkerStderrCapture # noqa: E402
+
+
+class _Queue:
+ def __init__(self, raises: bool = False):
+ self.raises = raises
+ self.closed = False
+ self.joined = False
+
+ def close(self):
+ self.closed = True
+ if self.raises:
+ raise OSError("already closed")
+
+ def join_thread(self):
+ self.joined = True
+
+
+def _capture_exit(monkeypatch):
+ codes: list[int] = []
+ monkeypatch.setattr(worker.os, "_exit", lambda code: codes.append(code))
+ return codes
+
+
+def test_shutdown_exits_the_process_after_flushing(monkeypatch):
+ codes = _capture_exit(monkeypatch)
+ queue = _Queue()
+
+ worker._exit_after_shutdown(queue)
+
+ assert codes == [0]
+ assert queue.closed and queue.joined
+
+
+def test_shutdown_still_exits_when_the_queue_is_already_closed(monkeypatch):
+ codes = _capture_exit(monkeypatch)
+
+ worker._exit_after_shutdown(_Queue(raises = True))
+
+ assert codes == [0]
+
+
+_MIRRORED_CHILD = r"""
+import sys
+sys.path.insert(0, %(backend)r)
+from core.inference import worker
+from utils.worker_stderr import install_worker_stderr_mirror
+
+class Q:
+ def close(self): pass
+ def join_thread(self): pass
+
+assert install_worker_stderr_mirror(%(path)r) is True
+for i in range(200):
+ sys.stderr.write(f"bulk {i}\n")
+sys.stderr.write("the last line before shutdown\n")
+worker._exit_after_shutdown(Q())
+"""
+
+
+def test_shutdown_exit_still_relays_the_mirrored_stderr_tail(tmp_path):
+ # os._exit skips the atexit drain, which dropped every line the mirror pump had not relayed yet.
+ capture = WorkerStderrCapture(directory = str(tmp_path), prefix = "unsloth-test-")
+ script = _MIRRORED_CHILD % {"backend": str(_BACKEND_ROOT), "path": capture.path}
+ result = subprocess.run(
+ [sys.executable, "-c", script], stderr = subprocess.PIPE, text = True, timeout = 120
+ )
+ capture.close()
+
+ assert result.returncode == 0, result.stderr
+ assert "the last line before shutdown" in result.stderr
+ assert result.stderr.count("bulk ") == 200
diff --git a/studio/backend/utils/worker_stderr.py b/studio/backend/utils/worker_stderr.py
index f35878a17c5..999c601707e 100644
--- a/studio/backend/utils/worker_stderr.py
+++ b/studio/backend/utils/worker_stderr.py
@@ -7,6 +7,7 @@ from __future__ import annotations
import atexit
import collections
+import functools
import os
import re
import sys
@@ -25,6 +26,7 @@ __all__ = [
"install_worker_stderr_mirror",
"mark_log_record_continuations",
"stderr_tail_from_bytes",
+ "stop_worker_stderr_mirror",
"unexpected_exit_message",
]
@@ -583,6 +585,9 @@ def _stop_mirror(
pass
+_MIRROR_STOPPER = None
+
+
def install_worker_stderr_mirror(
path: "str | None", cap_bytes: int = MIRROR_FILE_CAP_BYTES
) -> bool:
@@ -636,10 +641,21 @@ def install_worker_stderr_mirror(
daemon = True,
)
pump.start()
- atexit.register(_stop_mirror, inherited, pump, stop, relay, rebound)
+ global _MIRROR_STOPPER
+ _MIRROR_STOPPER = functools.partial(_stop_mirror, inherited, pump, stop, relay, rebound)
+ atexit.register(_MIRROR_STOPPER)
return True
+def stop_worker_stderr_mirror() -> None:
+ """The atexit drain, for a worker leaving through os._exit."""
+ global _MIRROR_STOPPER
+ stopper, _MIRROR_STOPPER = _MIRROR_STOPPER, None
+ if stopper is not None:
+ atexit.unregister(stopper)
+ stopper()
+
+
_PREFIX_FORMATTER_CLASS = None
# Marked on the formatter, not via a class: a first-call race builds two classes and double-wraps.
_MARKS_CONTINUATIONS = "_unsloth_marks_continuations"