# Start llama.cpp server for E19 (4 parallel slots x 3072 context, q8_0 KV cache, 40 of 48 layers on the GPU), then run the sampler.
# Set $model to your Qwen2.5-14B-Instruct Q4_K_M (imatrix) GGUF and $server to llama-server.exe.
$model = ".\models\qwen2.5-14b-instruct-Q4_K_M-imat.gguf"; $server = ".\llama.cpp\llama-server.exe"
$p = Start-Process $server -ArgumentList "--port 8095 -c 12288 -np 4 -t 16 -fa on -ctk q8_0 -ctv q8_0 -m $model -ngl 40" -PassThru -WindowStyle Hidden
do { Start-Sleep 5; try { $ok = (Invoke-RestMethod http://127.0.0.1:8095/health).status -eq "ok" } catch { $ok = $false } } until ($ok -or $p.HasExited)
python e19_math500_sc.py
$p.Kill()