Studio: add option to disable the in-memory API monitor (#7156)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>

studio/backend/core/inference/api_monitor.py

@@ -5,6 +5,7 @@
 
 from __future__ import annotations
 
+import os
 import threading
 import time
 import uuid
@@ -18,6 +19,14 @@ _MAX_PROMPT_CHARS = 12000
 _MAX_REPLY_CHARS = 12000
 _PREVIEW_CHARS = 360
 
+# Opt-in startup kill switch for Studio's in-memory API monitor.
+_DISABLE_ENV = "UNSLOTH_STUDIO_DISABLE_API_MONITOR"
+_TRUE_VALUES = frozenset({"1", "true", "yes", "on"})
+
+
+def _api_monitor_disabled() -> bool:
+    return os.environ.get(_DISABLE_ENV, "").strip().lower() in _TRUE_VALUES
+
 
 def _trim(text: Optional[str], limit: int) -> str:
     if not text:
@@ -104,10 +113,16 @@ class ApiMonitorEntry:
 
 
 class ApiMonitor:
-    def __init__(self, max_entries: int = _MAX_ENTRIES):
+    def __init__(
+        self,
+        max_entries: int = _MAX_ENTRIES,
+        *,
+        enabled: bool = True,
+    ):
         self._entries: deque[ApiMonitorEntry] = deque()
         self._max_entries = max(0, max_entries)
         self._lock = threading.Lock()
+        self._enabled = enabled
 
     def start(
         self,
@@ -119,6 +134,8 @@ class ApiMonitor:
         context_length: Optional[int] = None,
         subject: Optional[str] = None,
     ) -> str:
+        if not self._enabled:
+            return ""
         now = time.time()
         entry = ApiMonitorEntry(
             id = f"apireq_{uuid.uuid4().hex[:12]}",
@@ -152,6 +169,8 @@ class ApiMonitor:
         :meth:`fail`; an unload is terminal on arrival. Rows are shared (visible to
         every subject) and share the request retention budget.
         """
+        if not self._enabled:
+            return ""
         now = time.time()
         entry = ApiMonitorEntry(
             id = f"apievt_{uuid.uuid4().hex[:12]}",
@@ -392,4 +411,4 @@ class ApiMonitor:
         self._entries = kept
 
 
-api_monitor = ApiMonitor()
+api_monitor = ApiMonitor(enabled = not _api_monitor_disabled())

studio/backend/tests/test_api_monitor.py

@@ -260,6 +260,63 @@ def test_api_monitor_append_reply_exact_cap_then_more_marks_truncated():
     assert len(reply) == m._MAX_REPLY_CHARS and reply.endswith("...")
 
 
+def test_api_monitor_disabled_is_noop():
+    monitor = ApiMonitor(max_entries = 3, enabled = False)
+
+    request_id = monitor.start(
+        endpoint = "/v1/chat/completions",
+        method = "POST",
+        model = "local-model",
+        prompt = "user: hello",
+        context_length = 100,
+    )
+    load_id = monitor.record_lifecycle(
+        event = "load",
+        model = "local-model",
+        running = True,
+    )
+    unload_id = monitor.record_lifecycle(
+        event = "unload",
+        model = "local-model",
+    )
+    assert request_id == load_id == unload_id == ""
+
+    # Every mutator must be a safe no-op on the falsy id.
+    monitor.append_reply(request_id, "hi")
+    monitor.set_reply(request_id, "hi")
+    monitor.set_usage(request_id, prompt_tokens = 4, completion_tokens = 6)
+    monitor.relabel(load_id, "renamed-model")
+    monitor.set_progress(load_id, 50)
+    monitor.finish(load_id)
+    monitor.fail_open(load_id, "boom")
+    monitor.fail(request_id, "boom")
+    monitor.discard(unload_id)
+
+    assert monitor.snapshot() == []
+    assert monitor.active_count() == 0
+    assert monitor.get(request_id) is None
+
+
+def test_api_monitor_disable_env_var_truthy(monkeypatch):
+    import core.inference.api_monitor as m
+    for value in ("1", "true", "yes", "on", "TRUE", "On", " yes "):
+        monkeypatch.setenv(m._DISABLE_ENV, value)
+        assert m._api_monitor_disabled() is True, value
+
+
+def test_api_monitor_disable_env_var_falsy(monkeypatch):
+    import core.inference.api_monitor as m
+    for value in ("", "0", "false", "no", "off", "disabled"):
+        monkeypatch.setenv(m._DISABLE_ENV, value)
+        assert m._api_monitor_disabled() is False, value
+
+
+def test_api_monitor_disable_env_var_unset(monkeypatch):
+    import core.inference.api_monitor as m
+    monkeypatch.delenv(m._DISABLE_ENV, raising = False)
+    assert m._api_monitor_disabled() is False
+
+
 # ── model lifecycle rows (load / unload) ────────────────────────────