#!/usr/bin/env bash # Benchmark setup — ensure toolboxes and test models are ready set -euo pipefail SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" source "$SCRIPT_DIR/../../lib/common.sh" source "$SCRIPT_DIR/../../lib/detect.sh" TOOLBOXES_REPO="/data/workspace/projects/HomeLab/strix-halo-toolboxes/amd-strix-halo-llamacpp-toolboxes" MODEL_DIR="$(data_dir models)" log_header "Benchmark Setup" # ── 1. Check toolbox containers ────────────────────────── log_info "Checking toolbox containers..." # Minimum required: vulkan-radv (most stable) REQUIRED_TOOLBOXES=("llama-vulkan-radv") OPTIONAL_TOOLBOXES=("llama-rocm-6.4.4" "llama-rocm-7.2" "llama-vulkan-amdvlk") existing=$(detect_toolbox_names 2>/dev/null || true) missing=() for tb in "${REQUIRED_TOOLBOXES[@]}"; do if echo "$existing" | grep -q "^${tb}$"; then log_success "Toolbox: $tb" else missing+=("$tb") log_warn "Toolbox missing: $tb" fi done for tb in "${OPTIONAL_TOOLBOXES[@]}"; do if echo "$existing" | grep -q "^${tb}$"; then log_success "Toolbox: $tb (optional)" else log_info "Toolbox not present: $tb (optional)" fi done if (( ${#missing[@]} > 0 )); then log_info "Need to create required toolboxes." if [[ -d "$TOOLBOXES_REPO" ]]; then log_info "Found toolboxes repo at: $TOOLBOXES_REPO" if confirm "Create missing toolboxes using refresh-toolboxes.sh?"; then for tb in "${missing[@]}"; do log_info "Creating $tb..." bash "$TOOLBOXES_REPO/refresh-toolboxes.sh" "$tb" done fi else log_error "Toolboxes repo not found at: $TOOLBOXES_REPO" log_info "Clone it: git clone https://github.com/kyuz0/amd-strix-halo-toolboxes" log_info "Then re-run this setup." exit 1 fi fi # ── 2. Verify GPU access inside toolboxes ──────────────── log_info "Verifying GPU access in toolboxes..." for tb in "${REQUIRED_TOOLBOXES[@]}"; do if echo "$existing" | grep -qF "$tb"; then if toolbox run -c "$tb" -- llama-cli --list-devices 2>&1 | grep -qi "gpu\|vulkan\|rocm"; then log_success "GPU accessible in $tb" else log_warn "GPU may not be accessible in $tb — check device mappings" fi fi done # ── 3. Check for test models ──────────────────────────── log_info "Checking for test models in $MODEL_DIR..." model_count=$(find "$MODEL_DIR" -name "*.gguf" 2>/dev/null | wc -l) if (( model_count > 0 )); then log_success "Found $model_count model(s):" find "$MODEL_DIR" -name "*.gguf" | while read -r f; do size=$(du -h "$f" | cut -f1) printf " %s (%s)\n" "$(basename "$f")" "$size" done else log_warn "No GGUF models found in $MODEL_DIR" log_info "Download a test model. Example:" echo "" echo " # Small (4B, ~3 GB):" echo " huggingface-cli download Qwen/Qwen3-4B-GGUF Qwen3-4B-Q4_K_M.gguf \\" echo " --local-dir $MODEL_DIR" echo "" echo " # Medium (14B, ~9 GB):" echo " huggingface-cli download Qwen/Qwen3-14B-GGUF Qwen3-14B-Q4_K_M.gguf \\" echo " --local-dir $MODEL_DIR" echo "" if is_cmd huggingface-cli; then if confirm "Download Qwen3-4B Q4_K_M (~3 GB) as test model?"; then huggingface-cli download Qwen/Qwen3-4B-GGUF Qwen3-4B-Q4_K_M.gguf \ --local-dir "$MODEL_DIR" log_success "Model downloaded" fi else log_info "Install huggingface-cli: pip install huggingface_hub[cli]" fi fi log_header "Setup Complete" log_info "Run 'make benchmark-baseline' to capture your baseline."