`mlock: true` does not lock model in RAM — no mlock() syscall observed

[ilsawa]

LocalAI version:
v3.0.0

Environment, CPU architecture, OS, and Version:
Linux 6.5.0-41-generic #41~22.04.1-Ubuntu SMP PREEMPT_DYNAMIC
x86_64 architecture, AMD CPU (16 threads), 64 GB RAM, 50 GB RAM free
Manual FHS-compliant install (not Docker), user localai

Describe the bug
Setting mlock: true in model configuration does not result in any memory being locked in RAM. No mlock() or mlockall() syscall is ever invoked.
The model remains swappable even though mlock is explicitly requested.

To Reproduce
Model config (llama3-8b-instruct:Q6_K.yaml):

name: "llama3-8b-instruct:Q6_K"
context_size: 32768
mmap: true
mlock: true
parameters:
  model: Meta-Llama-3-8B-Instruct.Q6_K.gguf
  n_threads: 16

Systemd service:

[Service]
User=localai
...
LimitMEMLOCK=infinity

The localai user has unlimited memlock:

sudo -u localai bash -c 'ulimit -l'
# → unlimited

Started manually with:

sudo -u localai strace -f -e mlock,mlockall /usr/local/bin/local-ai --models-path /mnt/prg/localai/models --threads 16 --debug

Expected behavior
Expected to see a mlock() or mlockall() syscall when model is loaded, and a non-zero value in /proc/$(pidof local-ai)/status under VmLck:.

Logs
Debug output confirms mlock: true was parsed:

LLMConfig:{... MMap:0xc056efd4a9 MMlock:0xc056efd6d1 ...}

But VmLck remains:

VmLck:         0 kB

And strace shows no mlock() or mlockall() calls were made.

Additional context

• Running mlock() manually as localai works (verified with small C test)
• Happens regardless of mmap: true or mmap: false
• Behavior confirmed for quantized llama3-8b Q6_K model
• Possibly a regression or unimplemented backend feature?