reorder

pose_llm/deploy_pose_llm.sh

@@ -0,0 +1,104 @@
+#!/bin/bash
+# Deploy/stop/restart/status the uncensored chat LLM API on tour.
+# Serves Dolphin-24B GGUF via llama-server (full GPU, ~10-20 tok/s).
+#
+# Usage:
+#   ./deploy_pose_llm.sh deploy     # (re)start, wait for health
+#   ./deploy_pose_llm.sh stop
+#   ./deploy_pose_llm.sh restart
+#   ./deploy_pose_llm.sh status
+
+set -euo pipefail
+
+REMOTE="tour@192.168.1.160"
+REMOTE_DIR="/media/tour/NVME0/llm"
+LLAMA_SERVER="$REMOTE_DIR/llama.cpp/build/bin/llama-server"
+MODEL="$REMOTE_DIR/models/cognitivecomputations_Dolphin3.0-Mistral-24B-Q4_K_M.gguf"
+PORT="${PORT:-8001}"
+CTX="${CTX:-4096}"
+NGL="${NGL:-99}"  # GPU layers: 99 = all on GPU
+
+ACTION="${1:-deploy}"
+
+print_header() { echo; echo "============================================================"; echo "$1"; echo "============================================================"; }
+
+stop_one() {
+  print_header "Stopping pose-LLM on $REMOTE"
+  ssh "$REMOTE" "
+    cd '$REMOTE_DIR'
+    if [ -f api.pid ]; then
+      PID=\$(cat api.pid)
+      if kill -0 \"\$PID\" 2>/dev/null; then
+        echo '==> Stopping PID' \"\$PID\"; kill \"\$PID\"
+        for i in \$(seq 1 30); do kill -0 \"\$PID\" 2>/dev/null || { echo '==> stopped'; rm -f api.pid; exit 0; }; sleep 1; done
+        echo '==> hard kill'; kill -9 \"\$PID\" 2>/dev/null || true
+      fi
+      rm -f api.pid
+    else echo '==> no api.pid'; fi
+    PORT_PID=\$(lsof -ti:$PORT 2>/dev/null || true)
+    [ -n \"\$PORT_PID\" ] && { echo '==> killing port $PORT:' \"\$PORT_PID\"; kill \$PORT_PID 2>/dev/null || true; }
+    echo '==> waiting 8s for VRAM drain...'; sleep 8
+    true
+  "
+}
+
+deploy_one() {
+  print_header "Deploying llama-server on $REMOTE (model: Q4_K_M, ngl=$NGL)"
+
+  echo "==> Writing start script..."
+  ssh "$REMOTE" "cat > '$REMOTE_DIR/start_pose_llm.sh'" << EOF
+#!/usr/bin/env bash
+set -euo pipefail
+cd "$REMOTE_DIR"
+if [ -f api.pid ]; then
+  kill \$(cat api.pid) 2>/dev/null || true
+  rm -f api.pid
+  echo "waiting for VRAM to drain..."
+  sleep 10
+fi
+export HSA_OVERRIDE_GFX_VERSION=9.0.6
+export ROCR_VISIBLE_DEVICES=0
+nohup $LLAMA_SERVER \\
+  -m $MODEL \\
+  --host 0.0.0.0 \\
+  --port $PORT \\
+  -ngl $NGL \\
+  -c $CTX \\
+  --chat-template chatml \\
+  -np 1 \\
+  > api.log 2>&1 &
+echo \$! > api.pid
+echo "started PID \$(cat api.pid)"
+EOF
+  ssh "$REMOTE" "chmod +x '$REMOTE_DIR/start_pose_llm.sh' && '$REMOTE_DIR/start_pose_llm.sh'"
+
+  echo "==> Waiting for model load + health (can take ~30s)..."
+  for i in $(seq 1 90); do
+    if ssh "$REMOTE" "curl -fsS http://localhost:$PORT/health >/dev/null 2>&1"; then echo "==> API ready"; break; fi
+    if [ "$i" -eq 90 ]; then echo "ERROR: not healthy after 180s"; ssh "$REMOTE" "tail -n 60 $REMOTE_DIR/api.log"; exit 1; fi
+    sleep 2
+  done
+  echo "==> Health:"; ssh "$REMOTE" "curl -s http://localhost:$PORT/health | python3 -m json.tool"
+}
+
+status_one() {
+  print_header "Status on $REMOTE"
+  ssh "$REMOTE" "
+    cd '$REMOTE_DIR'
+    echo '==> Binary:'; ls -lh $LLAMA_SERVER 2>/dev/null || echo 'NOT BUILT'
+    echo '==> Model:'; ls -lh $MODEL 2>/dev/null || echo 'NOT DOWNLOADED'
+    echo '==> PID:'; cat api.pid 2>/dev/null || echo none
+    echo '==> Port $PORT:'; lsof -i:$PORT 2>/dev/null || echo 'nothing listening'
+    echo '==> Health:'; curl -fsS http://localhost:$PORT/health 2>/dev/null | python3 -m json.tool || echo 'not healthy'
+    echo '==> Last 20 log lines:'; tail -n 20 api.log 2>/dev/null || true
+    echo '==> VRAM:'; /opt/rocm/bin/rocm-smi --showmeminfo vram 2>/dev/null | grep -i 'used\|total' || true
+  "
+}
+
+case "$ACTION" in
+  deploy)  deploy_one ;;
+  stop)    stop_one ;;
+  restart) stop_one; deploy_one ;;
+  status)  status_one ;;
+  *) echo "Usage: $0 [deploy|stop|restart|status]"; exit 1 ;;
+esac