Make agent unload resident Ollama models before threshold restarts
This commit is contained in:
@@ -6,6 +6,8 @@
|
||||
"poll_interval_seconds": 60,
|
||||
"swap_threshold_percent": null,
|
||||
"swap_threshold_mb": 3048,
|
||||
"unload_models_before_restart": true,
|
||||
"unload_wait_seconds": 3,
|
||||
"cooldown_seconds": 600,
|
||||
"max_restarts_per_hour": 2,
|
||||
"restart_force_after_seconds": 300,
|
||||
|
||||
@@ -32,6 +32,8 @@ case "${OS}" in
|
||||
"poll_interval_seconds": 60,
|
||||
"swap_threshold_percent": null,
|
||||
"swap_threshold_mb": 3048,
|
||||
"unload_models_before_restart": true,
|
||||
"unload_wait_seconds": 3,
|
||||
"cooldown_seconds": 600,
|
||||
"max_restarts_per_hour": 2,
|
||||
"restart_force_after_seconds": 300,
|
||||
@@ -97,6 +99,8 @@ PLIST
|
||||
"poll_interval_seconds": 60,
|
||||
"swap_threshold_percent": null,
|
||||
"swap_threshold_mb": 3048,
|
||||
"unload_models_before_restart": true,
|
||||
"unload_wait_seconds": 3,
|
||||
"cooldown_seconds": 600,
|
||||
"max_restarts_per_hour": 2,
|
||||
"restart_force_after_seconds": 300,
|
||||
|
||||
@@ -8,6 +8,7 @@ import http.server
|
||||
import json
|
||||
import os
|
||||
import platform
|
||||
import shutil
|
||||
import socketserver
|
||||
import subprocess
|
||||
import threading
|
||||
@@ -44,6 +45,7 @@ class StateStore:
|
||||
loaded = json.load(handle)
|
||||
if isinstance(loaded, dict):
|
||||
self.data.update(loaded)
|
||||
self.data["restart_in_progress"] = False
|
||||
except (OSError, json.JSONDecodeError):
|
||||
pass
|
||||
|
||||
@@ -168,13 +170,71 @@ class RemoteOllama:
|
||||
return data
|
||||
|
||||
|
||||
|
||||
|
||||
def resolve_ollama_binary(config: Dict[str, Any]) -> Optional[str]:
|
||||
configured = config.get("ollama_binary")
|
||||
if isinstance(configured, str) and configured:
|
||||
if Path(configured).is_file():
|
||||
return configured
|
||||
binary = shutil.which("ollama")
|
||||
if binary:
|
||||
return binary
|
||||
candidates = (
|
||||
"/Applications/Ollama.app/Contents/Resources/ollama",
|
||||
"/usr/local/bin/ollama",
|
||||
"/opt/ollama/ollama",
|
||||
"/usr/bin/ollama",
|
||||
)
|
||||
for candidate in candidates:
|
||||
if Path(candidate).is_file():
|
||||
return candidate
|
||||
return None
|
||||
|
||||
|
||||
def stop_ollama_models(config: Dict[str, Any]) -> List[Dict[str, Any]]:
|
||||
results = []
|
||||
if not bool(config.get("unload_models_before_restart", True)):
|
||||
return results
|
||||
binary = resolve_ollama_binary(config)
|
||||
if binary is None:
|
||||
return [{"model": None, "returncode": -1, "error": "ollama binary not found"}]
|
||||
try:
|
||||
models = RemoteOllama(config).running_models()
|
||||
except Exception as exc:
|
||||
return [{"model": None, "returncode": -1, "error": "cannot list models: %s" % exc}]
|
||||
for model in models:
|
||||
name = model.get("name") or model.get("model")
|
||||
if not name:
|
||||
continue
|
||||
try:
|
||||
completed = subprocess.run(
|
||||
[binary, "stop", str(name)],
|
||||
check=False,
|
||||
capture_output=True,
|
||||
text=True,
|
||||
timeout=30,
|
||||
)
|
||||
results.append({
|
||||
"model": name,
|
||||
"returncode": completed.returncode,
|
||||
"stdout": completed.stdout.strip(),
|
||||
"stderr": completed.stderr.strip(),
|
||||
})
|
||||
except (OSError, subprocess.SubprocessError, TimeoutError) as exc:
|
||||
results.append({"model": name, "returncode": -1, "error": str(exc)})
|
||||
wait_seconds = int(config.get("unload_wait_seconds", 0))
|
||||
if wait_seconds > 0 and (models or results):
|
||||
time.sleep(wait_seconds)
|
||||
return results
|
||||
|
||||
|
||||
class ServiceController:
|
||||
def __init__(self, config: Dict[str, Any], state: StateStore):
|
||||
self.config = config
|
||||
self.state = state
|
||||
|
||||
def restart(self, force: bool = False, reason: str = "threshold") -> Dict[str, Any]:
|
||||
stats_before = None
|
||||
busy = False
|
||||
try:
|
||||
busy = RemoteOllama(self.config).is_busy()
|
||||
@@ -184,6 +244,10 @@ class ServiceController:
|
||||
if busy and not force:
|
||||
raise BusyError("ollama has running models; restart deferred")
|
||||
|
||||
# Stop only resident Ollama models before restarting the service. This
|
||||
# releases loaded resources without broad shell usage.
|
||||
models_stopped = stop_ollama_models(self.config)
|
||||
|
||||
started = common.now_utc()
|
||||
with self.state.lock:
|
||||
if self.state.data.get("restart_in_progress"):
|
||||
@@ -192,6 +256,7 @@ class ServiceController:
|
||||
self.state._save()
|
||||
|
||||
try:
|
||||
service_started = common.now_utc()
|
||||
if SYSTEM == "darwin":
|
||||
plist = str(self.config["launch_agent_plist"])
|
||||
subprocess.run(
|
||||
@@ -236,7 +301,8 @@ class ServiceController:
|
||||
"reason": reason,
|
||||
"busy_before": busy,
|
||||
"health_ok": health_ok,
|
||||
"duration_seconds": round((finished - started).total_seconds(), 2),
|
||||
"models_stopped": models_stopped,
|
||||
"duration_seconds": round((finished - service_started).total_seconds(), 2),
|
||||
}
|
||||
except Exception:
|
||||
self.state.mutate(
|
||||
|
||||
@@ -23,6 +23,9 @@ DEFAULT_CONFIG = {
|
||||
"cooldown_seconds": 600,
|
||||
"max_restarts_per_hour": 2,
|
||||
"restart_force_after_seconds": 300,
|
||||
"unload_models_before_restart": True,
|
||||
"unload_wait_seconds": 3,
|
||||
"ollama_binary": None,
|
||||
"busy_check_timeout_seconds": 3,
|
||||
"health_timeout_seconds": 5,
|
||||
"restart_health_timeout_seconds": 180,
|
||||
@@ -121,6 +124,8 @@ def validate_config(config: dict[str, Any]) -> None:
|
||||
absolute = config.get("swap_threshold_mb")
|
||||
if absolute is not None and float(absolute) < 0:
|
||||
raise ValueError("swap_threshold_mb must be non-negative")
|
||||
if config.get("unload_wait_seconds", 0) < 0:
|
||||
raise ValueError("unload_wait_seconds must be non-negative")
|
||||
|
||||
|
||||
def _suffix_to_mb(value: float, suffix: str) -> float:
|
||||
|
||||
Reference in New Issue
Block a user