Add RTX 4060 / i7-13700H optimized inference example for Lenovo Legion Slim 5

[Copilot]

Adds a ready-to-use inference script tuned for the Lenovo Legion Slim 5 (RTX 4060 8 GB VRAM, i7-13700H, 16 GB DDR5 RAM).

Changes

• examples/high_level_api/legion_slim5_rtx4060.py — new example with hardware-aware defaults:
  • n_gpu_layers=-1 + offload_kqv=True — full GPU offload with KV-cache on VRAM
  • n_threads=6 — P-cores only for best throughput on the hybrid i7-13700H topology
  • use_mmap=True / use_mlock=False — fast NVMe model loading without pinning the constrained 16 GB RAM
  • n_ctx=4096, n_batch=512 — safe defaults within 8 GB VRAM budget
  • Inline quantisation guide (Q5_K_M / Q6_K recommended; Q8_0 flagged as marginal)

Usage

# Install with CUDA support
CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --force-reinstall --no-cache-dir

# Run inference
python examples/high_level_api/legion_slim5_rtx4060.py \
    -m ./mistral-7b-Q5_K_M.gguf \
    -p "Your prompt here" \
    --max-tokens 256

# Reduce GPU layers if VRAM OOM
python examples/high_level_api/legion_slim5_rtx4060.py \
    -m ./model.gguf --n-gpu-layers 28