#!/usr/bin/env bash
# Launch the two surrogate VLM servers on klaus-3 via llama.cpp (OpenAI-compatible).
#
# Prereqs on klaus-3:
#   - a CUDA build of llama.cpp with the server binary (llama-server)
#   - GGUF weights for qwen-3.5-4b and gemma-4-4b (multimodal / vision projector)
#
# VRAM budget (8GB): two Q4_K_M 4B models (~2.6GB each) + CLIP (~1GB) fit under 8GB.
# Adjust --n-gpu-layers / --ctx-size if you hit OOM.
set -euo pipefail

LLAMA_SERVER="${LLAMA_SERVER:-$HOME/llama.cpp/build/bin/llama-server}"
MODEL_DIR="${MODEL_DIR:-$HOME/workspace/veil-pgd/models}"

QWEN_GGUF="${QWEN_GGUF:-$MODEL_DIR/qwen-3.5-4b-instruct-q4_k_m.gguf}"
QWEN_MMPROJ="${QWEN_MMPROJ:-$MODEL_DIR/qwen-3.5-4b-mmproj.gguf}"
GEMMA_GGUF="${GEMMA_GGUF:-$MODEL_DIR/gemma-4-4b-it-q4_k_m.gguf}"
GEMMA_MMPROJ="${GEMMA_MMPROJ:-$MODEL_DIR/gemma-4-4b-mmproj.gguf}"

mkdir -p "$HOME/workspace/veil-pgd/logs"
LOG="$HOME/workspace/veil-pgd/logs"

echo "Starting qwen-3.5-4b on :8081"
nohup "$LLAMA_SERVER" \
  --model "$QWEN_GGUF" --mmproj "$QWEN_MMPROJ" \
  --host 0.0.0.0 --port 8081 --ctx-size 4096 --n-gpu-layers 99 \
  --alias qwen-3.5-4b > "$LOG/qwen.log" 2>&1 &

echo "Starting gemma-4-4b on :8082"
nohup "$LLAMA_SERVER" \
  --model "$GEMMA_GGUF" --mmproj "$GEMMA_MMPROJ" \
  --host 0.0.0.0 --port 8082 --ctx-size 4096 --n-gpu-layers 99 \
  --alias gemma-4-4b > "$LOG/gemma.log" 2>&1 &

echo "Launched. Tail logs in $LOG. Verify: curl localhost:8081/v1/models"