Bug: fits_resource_constraints compares model memory against system RAM, not GPU VRAM

[mrveiss]

Problem

ModelInfo.fits_resource_constraints() in autobot-backend/utils/model_optimization/types.py compares the estimated model memory requirement against SystemResources.available_memory_gb (system RAM). However, LLM models are loaded into GPU VRAM, not system RAM.

This means:

• A system with 64GB RAM but 8GB VRAM will allow loading a 30GB model that will OOM on the GPU
• A system with 8GB RAM but 24GB VRAM will reject models that would fit perfectly on the GPU

SystemResources has a gpu_vram_gb field (added in #1966) but it is never populated by callers, and fits_resource_constraints() doesn't use it.

Discovered During

Working on #1966 (model memory estimation).

Expected Behavior

• When GPU is available, compare estimated model memory against GPU VRAM
• When CPU-only inference, compare against system RAM
• SystemResources callers should populate gpu_vram_gb from GPU detection (Bug: GPU detection only recognizes NVIDIA RTX cards, fails on AMD/Intel/non-RTX NVIDIA #1959)

Impact

High — Models may be recommended that OOM on GPU, or rejected despite fitting in VRAM.

[github-actions]

PR #2063 (merged to Dev_new_gui) references this issue with a close keyword.

fix(model): compare model memory against GPU VRAM instead of system RAM (#2015)

If this issue is fully resolved, close it manually. If work remains, no action is needed.

[mrveiss]

Merged via PR. Closes #2015.