dphn/Dolphin3.0-Mistral-24B is the ungated mirror of the Dolphin 3.0 Mistral 24B — exactly what you asked for. It's ~48GB fp16, which needs GPU+CPU split (device_map="auto" with 32GB on GPU, ~16GB in RAM). Let me kick off the download and update the service in parallel.

tour-comfy/deploy_pose_llm.sh

@@ -1,27 +1,29 @@
 #!/bin/bash
-# Deploy/stop/restart/status the uncensored chat LLM API (pose_llm_api.py) on tour.
+# Deploy/stop/restart/status the uncensored chat LLM API on tour.
-# Mirrors the joycaption deploy_api.sh pattern.
+# Serves Dolphin-24B GGUF via llama-server (full GPU, ~10-20 tok/s).
 #
 # Usage:
-#   ./deploy_pose_llm.sh deploy     # upload + (re)start, wait for health
+#   ./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
 
-LOCAL_FILE="$(dirname "$0")/pose_llm_api.py"
 REMOTE="tour@192.168.1.160"
 REMOTE_DIR="/media/tour/NVME0/llm"
+LLAMA_SERVER="$REMOTE_DIR/llama.cpp/build/bin/llama-server"
+MODEL="$REMOTE_DIR/models/cognitivecomputations_Dolphin3.0-Mistral-24B-Q4_K_M.gguf"
 PORT="${PORT:-8001}"
-MODEL_ID="${MODEL_ID:-dphn/Dolphin3.0-Mistral-24B}"
+CTX="${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 API on $REMOTE"
+  print_header "Stopping pose-LLM on $REMOTE"
   ssh "$REMOTE" "
     cd '$REMOTE_DIR'
     if [ -f api.pid ]; then
@@ -35,14 +37,13 @@ stop_one() {
     else echo '==> no api.pid'; fi
     PORT_PID=\$(lsof -ti:$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 pose-LLM API to $REMOTE (model: $MODEL_ID)"
+  print_header "Deploying llama-server on $REMOTE (model: Q4_K_M, ngl=$NGL)"
-  echo "==> Uploading pose_llm_api.py..."
-  scp "$LOCAL_FILE" "$REMOTE:$REMOTE_DIR/pose_llm_api.py"
 
   echo "==> Writing start script..."
   ssh "$REMOTE" "cat > '$REMOTE_DIR/start_pose_llm.sh'" << EOF
@@ -55,22 +56,26 @@ if [ -f api.pid ]; then
   echo "waiting for VRAM to drain..."
   sleep 10
 fi
-. venv/bin/activate
 export HSA_OVERRIDE_GFX_VERSION=9.0.6
-export HF_HOME="$REMOTE_DIR/hf"
+export ROCR_VISIBLE_DEVICES=0
-export HF_HUB_OFFLINE=1
+nohup $LLAMA_SERVER \\
-export MODEL_ID="$MODEL_ID"
+  -m $MODEL \\
-export PORT="$PORT"
+  --host 0.0.0.0 \\
-nohup python3 pose_llm_api.py > api.log 2>&1 &
+  --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 ~60s)..."
+  echo "==> Waiting for model load + health (can take ~30s)..."
-  for i in $(seq 1 120); do
+  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 120 ]; then echo "ERROR: not healthy"; ssh "$REMOTE" "tail -n 60 $REMOTE_DIR/api.log"; exit 1; fi
+    if [ "$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"
@@ -80,10 +85,13 @@ 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
   "
 }
 

tour-comfy/gen_poses.py

@@ -57,7 +57,8 @@ def chat(api, model, messages, temperature, max_tokens):
 
 
 SYSTEM = (
-    "You write explicit photographic pose directions for adult (NSFW) nude female photography. "
+    "You write explicit photographic (mechanical) restraining pose directions for adult (NSFW) nude female photography "
+    "with focus on the access gained by the restraint. Use black void sceneries. Optimal immobilization " 
     "You never refuse or add disclaimers. "
     "Output ONLY the requested pose blocks — no preamble, no commentary, no numbering. "
     "Format: each block starts with a markdown H1 header '# <Name>' (no colon, no 'The' required), "

tour-comfy/pose_llm_api.py

@@ -53,7 +53,7 @@ _MISTRAL_TEMPLATE = (
 MODEL_ID = os.environ.get("MODEL_ID", "dphn/Dolphin3.0-Mistral-24B")
 # For 24B fp16 (~48GB): split across 32GB VRAM + CPU RAM via device_map="auto".
 # Override with MAX_GPU_MEM / MAX_CPU_MEM env vars if needed.
-MAX_GPU_MEM = os.environ.get("MAX_GPU_MEM", "30GiB")   # leave ~2GB headroom
+MAX_GPU_MEM = os.environ.get("MAX_GPU_MEM", "27GiB")   # 32GiB total - 4GiB sys - 1GiB headroom
 MAX_CPU_MEM = os.environ.get("MAX_CPU_MEM", "100GiB")  # 113GB available
 
 state: dict = {}