/api/ps and /api/tags report size and size_vram as a hardcoded 0

LocalAI version:
v4.9.0 (f7ad3f70eb5d8a0ddf80e08557f0d7df28cf032e), image localai/localai:v4.9.0-gpu-nvidia-cuda-12

Environment, CPU architecture, OS, and Version:

x86_64, RTX 3060 12GB, Docker on WSL2

Describe the bug
GET /api/ps reports "size": 0 and "size_vram": 0 for every loaded model. The
values are literals rather than unpopulated fields:

https://github.com/mudler/LocalAI/blob/v4.9.0/core/http/endpoints/ollama/models.go#L106-L110

GET /api/tags has the same literal Size: 0 at line 40, and ExpiresAt on the ps
entry is synthesised as time.Now().Add(24 * time.Hour).

This is worse than omitting the fields. /api/ps is an Ollama-compatibility surface,
and Ollama reports real byte counts there, so a client written against Ollama reads
0 as authoritative and concludes the models cost nothing. Anything scheduling GPU
work from that will over-commit the card. An absent field or a null is safe, because
a consumer can distinguish "unknown" from "nothing"; a 0 cannot be distinguished.

To Reproduce

1. Load any model (granite-4.1-8b, llama-cpp backend) so it is resident.
2. curl localhost:8080/api/ps
3. Compare against nvidia-smi.

Expected behavior
Either the real resident sizes, or the fields omitted/null when the backend cannot
report them. Not 0.

Logs

{"models":[{"name":"granite-4.1-8b:latest","size":0,"size_vram":0,
  "details":{"format":"gguf","family":"llama-cpp","parameter_size":"8B",
  "quantization_level":"Q4_K_M"}}]}

Card at the same moment: 10849 / 12288 MiB used, two models resident (second row
elided, also all-zero).

Additional context
I gave up on /api/ps for this engine entirely and treat the size as unknown unconditionally, because I can't tell its zeros from real ones.

Filed separately as a feature request: there is no endpoint that reports per-model memory at all.

I'm not in a position to take the PR, but I'm happy to test a fix against this setup.

Written by my beloved Claude Code

Start in core/http/endpoints/ollama/models.go at the /api/ps and /api/tags mappings referenced by the issue. Trace how loaded llama-cpp models are represented and whether resident sizes are available; then verify both endpoints with a loaded model. Done means the responses no longer present unavailable size values as authoritative zeroes, while retaining real byte counts when the backend can report them.