llama-cpp: xmrig + grafana hooks

tests/mock-llama-server-proc.py

@@ -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)