titaniumtown/nixos
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

llama-cpp: xmrig + grafana hooks

Browse Source
This commit is contained in:
Simon Gardling
2026-04-02 22:47:53 -04:00
parent ab9c12cb97
commit 096ffeb943
8 changed files with 369 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files

99
tests/llama-cpp-annotations.nix
View File

@@ -1,5 +1,4 @@
{
lib,
pkgs,
...
}:
@@ -8,49 +7,7 @@ let
script = ../services/llama-cpp-annotations.py;
python = pkgs.python3;
mockLlamaCpp = pkgs.writeText "mock-llama-cpp-server.py" ''
import http.server, json, sys, os
PORT = int(sys.argv[1])
STATE_FILE = sys.argv[2]
if not os.path.exists(STATE_FILE):
with open(STATE_FILE, "w") as f:
json.dump([{"id": 0, "is_processing": False, "next_token": {"n_decoded": 0}}], f)
class Handler(http.server.BaseHTTPRequestHandler):
def log_message(self, fmt, *args):
pass
def _json(self, code, body):
data = json.dumps(body).encode()
self.send_response(code)
self.send_header("Content-Type", "application/json")
self.end_headers()
self.wfile.write(data)
def do_GET(self):
if self.path == "/slots":
with open(STATE_FILE) as f:
slots = json.load(f)
self._json(200, slots)
else:
self.send_response(404)
self.end_headers()
def do_POST(self):
if self.path == "/test/set-slots":
length = int(self.headers.get("Content-Length", 0))
body = json.loads(self.rfile.read(length)) if length else []
with open(STATE_FILE, "w") as f:
json.dump(body, f)
self._json(200, {"ok": True})
else:
self.send_response(404)
self.end_headers()
http.server.HTTPServer(("127.0.0.1", PORT), Handler).serve_forever()
'';
mockLlamaProcess = ./mock-llama-server-proc.py;
in
pkgs.testers.runNixOSTest {
name = "llama-cpp-annotations";
@@ -61,6 +18,7 @@ pkgs.testers.runNixOSTest {
environment.systemPackages = [
pkgs.python3
pkgs.curl
pkgs.procps
];
};
@@ -69,25 +27,23 @@ pkgs.testers.runNixOSTest {
import time
GRAFANA_PORT = 13000
LLAMA_PORT = 16688
ANNOTS_FILE = "/tmp/annotations.json"
SLOTS_FILE = "/tmp/llama-slots.json"
LLAMA_STATE = "/tmp/llama-state.txt"
STATE_FILE = "/tmp/llama-annot-state.json"
PYTHON = "${python}/bin/python3"
MOCK_GRAFANA = "${mockGrafana}"
MOCK_LLAMA = "${mockLlamaCpp}"
MOCK_LLAMA = "${mockLlamaProcess}"
SCRIPT = "${script}"
def read_annotations():
out = machine.succeed(f"cat {ANNOTS_FILE} 2>/dev/null || echo '[]'")
return json.loads(out.strip())
def set_slots(slots):
machine.succeed(
f"curl -sf -X POST http://127.0.0.1:{LLAMA_PORT}/test/set-slots "
f"-H 'Content-Type: application/json' "
f"-d '{json.dumps(slots)}'"
)
def set_busy():
machine.succeed(f"echo busy > {LLAMA_STATE}")
def set_idle():
machine.succeed(f"echo idle > {LLAMA_STATE}")
start_all()
machine.wait_for_unit("multi-user.target")
@@ -98,10 +54,7 @@ pkgs.testers.runNixOSTest {
f"systemd-run --unit=mock-grafana {PYTHON} {MOCK_GRAFANA} {GRAFANA_PORT} {ANNOTS_FILE}"
)
machine.succeed(
f"echo '[{{\"id\": 0, \"is_processing\": false, \"next_token\": {{\"n_decoded\": 0}}}}]' > {SLOTS_FILE}"
)
machine.succeed(
f"systemd-run --unit=mock-llama {PYTHON} {MOCK_LLAMA} {LLAMA_PORT} {SLOTS_FILE}"
f"systemd-run --unit=mock-llama {PYTHON} {MOCK_LLAMA} {LLAMA_STATE}"
)
machine.wait_until_succeeds(
f"curl -sf http://127.0.0.1:{GRAFANA_PORT}/api/annotations -X POST "
@@ -109,7 +62,7 @@ pkgs.testers.runNixOSTest {
timeout=10,
)
machine.wait_until_succeeds(
f"curl -sf http://127.0.0.1:{LLAMA_PORT}/slots | grep -q is_processing",
"pgrep -x llama-server",
timeout=10,
)
machine.succeed(f"echo '[]' > {ANNOTS_FILE}")
@@ -117,62 +70,62 @@ pkgs.testers.runNixOSTest {
with subtest("Start annotation service"):
machine.succeed(
f"systemd-run --unit=llama-annot "
f"--setenv=LLAMA_CPP_URL=http://127.0.0.1:{LLAMA_PORT} "
f"--setenv=GRAFANA_URL=http://127.0.0.1:{GRAFANA_PORT} "
f"--setenv=STATE_FILE={STATE_FILE} "
f"--setenv=POLL_INTERVAL=2 "
f"--setenv=CPU_THRESHOLD=10 "
f"{PYTHON} {SCRIPT}"
)
time.sleep(3)
time.sleep(5)
with subtest("No annotations when slots are idle"):
with subtest("No annotations when idle"):
annots = read_annotations()
assert annots == [], f"Expected no annotations, got: {annots}"
with subtest("Annotation created when slot starts processing"):
set_slots([{"id": 0, "is_processing": True, "next_token": {"n_decoded": 0}}])
with subtest("Annotation created when llama-server becomes busy"):
set_busy()
machine.wait_until_succeeds(
f"cat {ANNOTS_FILE} | {PYTHON} -c "
f"\"import sys,json; a=json.load(sys.stdin); exit(0 if a else 1)\"",
timeout=15,
timeout=20,
)
annots = read_annotations()
assert len(annots) == 1, f"Expected 1 annotation, got: {annots}"
assert "llama-cpp" in annots[0].get("tags", []), f"Missing tag: {annots[0]}"
assert "slot 0" in annots[0]["text"], f"Missing slot info: {annots[0]['text']}"
assert "LLM request" in annots[0]["text"], f"Missing text: {annots[0]['text']}"
assert "timeEnd" not in annots[0], f"timeEnd should not be set: {annots[0]}"
with subtest("Annotation closed when slot stops processing"):
set_slots([{"id": 0, "is_processing": False, "next_token": {"n_decoded": 42}}])
with subtest("Annotation closed when llama-server becomes idle"):
set_idle()
machine.wait_until_succeeds(
f"cat {ANNOTS_FILE} | {PYTHON} -c "
f"\"import sys,json; a=json.load(sys.stdin); exit(0 if a and 'timeEnd' in a[0] else 1)\"",
timeout=15,
timeout=20,
)
annots = read_annotations()
assert len(annots) == 1, f"Expected 1, got: {annots}"
assert "timeEnd" in annots[0], f"timeEnd missing: {annots[0]}"
assert annots[0]["timeEnd"] > annots[0]["time"], "timeEnd should be after time"
assert "42 tokens" in annots[0].get("text", ""), f"Token count missing: {annots[0]}"
assert "s)" in annots[0].get("text", ""), f"Duration missing: {annots[0]}"
with subtest("State survives restart"):
set_slots([{"id": 0, "is_processing": True, "next_token": {"n_decoded": 0}}])
set_busy()
machine.wait_until_succeeds(
f"cat {ANNOTS_FILE} | {PYTHON} -c "
f"\"import sys,json; a=json.load(sys.stdin); exit(0 if len(a)==2 else 1)\"",
timeout=15,
timeout=20,
)
machine.succeed("systemctl stop llama-annot || true")
time.sleep(1)
machine.succeed(
f"systemd-run --unit=llama-annot-2 "
f"--setenv=LLAMA_CPP_URL=http://127.0.0.1:{LLAMA_PORT} "
f"--setenv=GRAFANA_URL=http://127.0.0.1:{GRAFANA_PORT} "
f"--setenv=STATE_FILE={STATE_FILE} "
f"--setenv=POLL_INTERVAL=2 "
f"--setenv=CPU_THRESHOLD=10 "
f"{PYTHON} {SCRIPT}"
)
time.sleep(4)
time.sleep(6)
annots = read_annotations()
assert len(annots) == 2, f"Restart should not duplicate, got: {annots}"
'';

162
tests/llama-cpp-xmrig-pause.nix Normal file
View File

@@ -0,0 +1,162 @@
{
pkgs,
...
}:
let
script = ../services/llama-cpp-xmrig-pause.py;
python = pkgs.python3;
# SmolLM-135M Q2_K: 85MB, modern GGUFv3, generates ~30 tok/s on one CPU
# thread — slow enough that a 200-token request keeps the process busy for
# several seconds, fast enough that tests don't crawl.
tinyModel = pkgs.fetchurl {
url = "https://huggingface.co/QuantFactory/SmolLM-135M-GGUF/resolve/main/SmolLM-135M.Q2_K.gguf";
hash = "sha256-DX46drPNJILNba21xfY2tyE0/yPWgOhz43gJdeSYKh4=";
};
in
pkgs.testers.runNixOSTest {
name = "llama-cpp-xmrig-pause";
nodes.machine =
{ pkgs, ... }:
{
environment.systemPackages = [
pkgs.python3
pkgs.procps
pkgs.curl
pkgs.llama-cpp
];
# Mock xmrig as a simple sleep process that can be stopped/started.
systemd.services.xmrig = {
description = "Mock xmrig miner";
serviceConfig = {
ExecStart = "${pkgs.coreutils}/bin/sleep infinity";
Type = "simple";
};
wantedBy = [ "multi-user.target" ];
};
};
testScript = ''
import time
PORT = 18088
MODEL = "${tinyModel}"
PYTHON = "${python}/bin/python3"
SCRIPT = "${script}"
# Tuned for test speed while remaining realistic.
# POLL_INTERVAL=1 keeps detection latency low.
# GRACE_PERIOD=5 is long enough to verify "stays stopped" but short enough
# that the full test completes in ~2 minutes.
# CPU_THRESHOLD=10 is low because the VM has limited cores and the model
# is small — but any active inference still saturates a core.
POLL_INTERVAL = "1"
GRACE_PERIOD = "5"
CPU_THRESHOLD = "10"
infer_counter = 0
def send_completion(n_predict=200):
"""Fire a completion request in the background via a transient systemd unit."""
global infer_counter
infer_counter += 1
name = f"infer-{infer_counter}"
machine.succeed(
f"systemd-run --unit={name} --property=Type=exec "
f"curl -sf -X POST http://127.0.0.1:{PORT}/completion "
f"-H 'Content-Type: application/json' "
f"-d '{{\"prompt\": \"Once upon a time in a land far away there lived\", \"n_predict\": {n_predict}}}'"
)
return name
def wait_inference_done(unit_name, timeout=60):
"""Wait for a background inference request to finish."""
machine.wait_until_fails(
f"systemctl is-active {unit_name}",
timeout=timeout,
)
start_all()
machine.wait_for_unit("multi-user.target")
machine.wait_for_unit("xmrig.service")
with subtest("Start llama-server"):
machine.succeed(
f"systemd-run --unit=llama-server "
# Single inference thread to maximise per-core CPU%, which is
# what the monitor measures. Keeps token generation slow enough
# (~30 tok/s) that a 200-token request sustains load for seconds.
f"llama-server --model {MODEL} --port {PORT} --ctx-size 512 -t 1 -np 1"
)
machine.wait_until_succeeds(
f"curl -sf http://127.0.0.1:{PORT}/health",
timeout=30,
)
machine.succeed("pgrep -x llama-server")
with subtest("Start pause monitor"):
machine.succeed(
f"systemd-run --unit=llama-xmrig-pause "
f"--setenv=POLL_INTERVAL={POLL_INTERVAL} "
f"--setenv=GRACE_PERIOD={GRACE_PERIOD} "
f"--setenv=CPU_THRESHOLD={CPU_THRESHOLD} "
f"{PYTHON} {SCRIPT}"
)
# The monitor needs two consecutive polls to compute a CPU delta.
# Wait for baseline to stabilise.
time.sleep(3)
with subtest("xmrig stays running while llama-server is idle"):
machine.succeed("systemctl is-active xmrig")
with subtest("xmrig stopped during prompt processing"):
unit = send_completion(n_predict=200)
machine.wait_until_fails("systemctl is-active xmrig", timeout=20)
with subtest("xmrig remains stopped during grace period after inference ends"):
wait_inference_done(unit)
# Inference just finished. The monitor will need 1-2 polls to detect
# idle, then the grace period starts. Checking 2s after completion
# is well within the 5s grace window.
time.sleep(2)
machine.fail("systemctl is-active xmrig")
with subtest("xmrig resumes after grace period expires"):
# Already idle since previous subtest. Grace period (5s) plus
# detection delay (~2 polls) means xmrig should restart within ~8s.
machine.wait_until_succeeds("systemctl is-active xmrig", timeout=15)
with subtest("Sequential prompts do not cause xmrig flapping"):
# First prompt — stop xmrig
unit1 = send_completion(n_predict=200)
machine.wait_until_fails("systemctl is-active xmrig", timeout=20)
wait_inference_done(unit1)
# Brief idle gap — shorter than grace period
time.sleep(2)
# Second prompt arrives before grace period expires, resetting it
unit2 = send_completion(n_predict=200)
time.sleep(3)
# xmrig must still be stopped
machine.fail("systemctl is-active xmrig")
wait_inference_done(unit2)
machine.wait_until_succeeds("systemctl is-active xmrig", timeout=15)
with subtest("xmrig stays stopped during sustained inference"):
unit = send_completion(n_predict=500)
machine.wait_until_fails("systemctl is-active xmrig", timeout=20)
# Stay busy longer than the grace period to prove continuous
# activity keeps xmrig stopped indefinitely.
time.sleep(8)
machine.fail("systemctl is-active xmrig")
wait_inference_done(unit)
machine.wait_until_succeeds("systemctl is-active xmrig", timeout=15)
'';
}

42
tests/mock-llama-server-proc.py Normal file
View File

@@ -0,0 +1,42 @@
#!/usr/bin/env python3
"""
Mock llama-server process for NixOS VM tests.
Sets /proc/self/comm to "llama-server" via prctl so that monitoring scripts
(llama-cpp-annotations, llama-cpp-xmrig-pause) can discover this process
the same way they discover the real one.
Usage: python3 mock-llama-server-proc.py <state-file>
The state file controls behavior:
"busy" -> burn CPU in a tight loop (simulates prompt processing / inference)
"idle" -> sleep (simulates waiting for requests)
"""
import ctypes
import ctypes.util
import sys
import time
STATE_FILE = sys.argv[1]
# PR_SET_NAME = 15, sets /proc/self/comm
libc = ctypes.CDLL(ctypes.util.find_library("c"), use_errno=True)
libc.prctl(15, b"llama-server", 0, 0, 0)
with open(STATE_FILE, "w") as f:
f.write("idle")
while True:
try:
with open(STATE_FILE) as f:
state = f.read().strip()
except Exception:
state = "idle"
if state == "busy":
end = time.monotonic() + 0.1
while time.monotonic() < end:
_ = sum(range(10000))
else:
time.sleep(0.5)

4
tests/tests.nix
View File

@@ -28,9 +28,9 @@ in
# zfs scrub annotations test
zfsScrubAnnotationsTest = handleTest ./zfs-scrub-annotations.nix;
# llama-cpp annotation service test
# llama-cpp tests
llamaCppAnnotationsTest = handleTest ./llama-cpp-annotations.nix;
llamaCppXmrigPauseTest = handleTest ./llama-cpp-xmrig-pause.nix;
# ntfy alerts test
ntfyAlertsTest = handleTest ./ntfy-alerts.nix;