feat(orchestrator): prefix all LLM runner env vars with LLM_

README.md

@@ -33,7 +33,8 @@ uv run bin/llm-tui
 ```
 
 Inside the TUI:
-- `F5` / **Launch**: Starts the server(s) in the background.
+- **Tied Configurations**: GGUF model weights, vision multimodal projectors (`mmproj`), chat templates, and inference presets are automatically bound bi-directionally. Selecting an inference preset auto-selects its matching GGUF model and mmproj; selecting a GGUF model auto-selects its paired mmproj and appropriate preset tuning.
+- `F5` / **Launch**: Starts the server(s) in the background with selected weights, projector, and parameters.
 - `F8` / **Stop**: Stops server on the selected port.
 - `F9` / **Stop All**: Kills all running `llama-server` instances.
 - `F2` / **Refresh**: Refreshes health status.
@@ -50,6 +51,12 @@ Inside the TUI:
 # SuperQwen 3.8 27B Solo (256K Context on port 11434)
 ./bin/llm-runner superqwen
 
+# Qwen 3.5 4B Solo (port 11434)
+./bin/llm-runner qwen4b
+
+# Custom Model Instance (with custom GGUF and vision projector)
+MODEL=models/muse-glimmer/Muse-Glimmer-30B-UD-Q4_K_XL.gguf MMPROJ=models/muse-glimmer/mmproj-Muse-Glimmer-30B-Q8_0.gguf ./bin/llm-runner custom
+
 # Check status of running servers
 ./bin/llm-runner status
 

bin/llm-runner

@@ -6,8 +6,19 @@
 
 set -euo pipefail
 
+# Sanitize environment:
+# TEMP/TMP are standard POSIX variables for temporary file directories.
+# If set to a sampling temperature (float/number), ROCm/libstdc++ crashes during initialization!
+if [ -n "${TEMP:-}" ] && [ ! -d "${TEMP}" ]; then
+    TEMPERATURE="${TEMPERATURE:-${TEMP}}"
+    unset TEMP
+fi
+if [ -n "${TMP:-}" ] && [ ! -d "${TMP}" ]; then
+    unset TMP
+fi
+
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
-LOG_DIR="${LOG_DIR:-${SCRIPT_DIR}/logs}"
+LOG_DIR="${LLM_LOG_DIR:-${LOG_DIR:-${SCRIPT_DIR}/logs}}"
 mkdir -p "${LOG_DIR}"
 
 # ------------------------------------------------------------------------------
@@ -45,7 +56,7 @@ LLAMA_SERVER="$(find_llama_server)"
 check_health() {
     local port="${1:-11434}"
     local host="${2:-127.0.0.1}"
-    if curl -s "http://${host}:${port}/health" 2>/dev/null | grep -q '"status":\s*"ok"\|"status":\s*"loading model"\|"ok"'; then
+    if curl -s "http://${host}:${port}/health" 2>/dev/null | grep -q '"status":\s*"ok"\|"ok"'; then
         return 0
     fi
     return 1
@@ -84,14 +95,8 @@ cmd_stop() {
         pkill -f "llama-server" 2>/dev/null || true
         log_success "All llama-server processes terminated."
     else
-        local pid
+        fuser -k -9 "$target/tcp" 2>/dev/null || true
-        pid=$(lsof -ti :"$target" 2>/dev/null || fuser "$target/tcp" 2>/dev/null || true)
+        log_success "Killed server on port ${target}."
-        if [ -n "$pid" ]; then
-            kill "$pid" 2>/dev/null || true
-            log_success "Killed server on port ${target} (PID: ${pid})."
-        else
-            log_warn "No active server found on port ${target}."
-        fi
     fi
 }
 
@@ -122,20 +127,19 @@ launch_instance() {
         exit 1
     fi
 
-    # Check if port is already in use
+    # Ensure target port is free
-    local existing_pid
+    if fuser "$port/tcp" >/dev/null 2>&1; then
-    existing_pid=$(lsof -ti :"$port" 2>/dev/null || fuser "$port/tcp" 2>/dev/null || true)
+        log_warn "Port ${port} is in use. Stopping existing process on port ${port}..."
-    if [ -n "$existing_pid" ]; then
+        fuser -k -9 "$port/tcp" 2>/dev/null || true
-        log_warn "Port ${port} is already in use by PID ${existing_pid}. Terminating existing process..."
-        kill -9 "$existing_pid" 2>/dev/null || true
         sleep 1
     fi
 
     local flags=(
         -m "${model}"
-        --host "${HOST:-0.0.0.0}"
+        --host "${LLM_HOST:-${HOST:-0.0.0.0}}"
         --port "${port}"
         --ctx-size "${ctx}"
+        -np "${LLM_PARALLEL:-${PARALLEL:-1}}"
         --n-gpu-layers "${ngl}"
         --load-mode "${load_mode}"
         --temp "${temp}"
@@ -145,12 +149,12 @@ launch_instance() {
         -ctk "${ctk}"
         -ctv "${ctv}"
         -fa "${flash_attn}"
-        -b "${BATCH_SIZE:-2048}"
+        -b "${LLM_BATCH_SIZE:-${BATCH_SIZE:-2048}}"
-        -ub "${UBATCH_SIZE:-512}"
+        -ub "${LLM_UBATCH_SIZE:-${UBATCH_SIZE:-512}}"
     )
 
     if [ -n "${mmproj}" ] && [ -f "${mmproj}" ]; then
-        flags+=(--mmproj "${mmproj}" --image-min-tokens "${IMAGE_MIN_TOKENS:-1024}")
+        flags+=(--mmproj "${mmproj}" --image-min-tokens "${LLM_IMAGE_MIN_TOKENS:-${IMAGE_MIN_TOKENS:-1024}}")
     fi
 
     if [ -n "${template}" ] && [ -f "${template}" ]; then
@@ -188,6 +192,8 @@ launch_instance() {
 # ------------------------------------------------------------------------------
 cmd_cluster() {
     log_info "Starting Multi-Agent Cluster (SuperQwen 27B + 2x Qwen 4B)..."
+    pkill -f "llama-server" 2>/dev/null || true
+    sleep 1
 
     # Trap to gracefully stop all background instances on Ctrl+C
     cleanup_cluster() {
@@ -199,41 +205,43 @@ cmd_cluster() {
     trap cleanup_cluster SIGINT SIGTERM EXIT
 
     # SuperQwen 3.8 27B (Planner)
-    local planner_model="${PLANNER_MODEL:-${SCRIPT_DIR}/models/Qwen3.8/Qwen3.8-27B-UD-Q4_K_XL.gguf}"
+    local planner_model="${LLM_PLANNER_MODEL:-${PLANNER_MODEL:-${LLM_MODEL:-${MODEL:-${SCRIPT_DIR}/models/Qwen3.8/Qwen3.8-27B-UD-Q4_K_XL.gguf}}}}"
     if [ ! -f "${planner_model}" ] && [ -f "${SCRIPT_DIR}/models/Qwen3.8/SuperQwen3.8-27b-abliterated-Q4_K_M.gguf" ]; then
         planner_model="${SCRIPT_DIR}/models/Qwen3.8/SuperQwen3.8-27b-abliterated-Q4_K_M.gguf"
     fi
-    local planner_mmproj="${PLANNER_MMPROJ:-${SCRIPT_DIR}/models/Qwen3.8/mmproj-SuperQwen3.8-27b-abliterated-Q8_0.gguf}"
+    local planner_mmproj="${LLM_PLANNER_MMPROJ:-${PLANNER_MMPROJ:-${SCRIPT_DIR}/models/Qwen3.8/mmproj-SuperQwen3.8-27b-abliterated-Q8_0.gguf}}"
-    local planner_tmpl="${PLANNER_TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.8/chat_template.jinja}"
+    if [ "${planner_mmproj}" = "none" ]; then planner_mmproj=""; fi
+    local planner_tmpl="${LLM_PLANNER_TEMPLATE:-${PLANNER_TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.8/chat_template.jinja}}"
 
     # Qwen 3.5 4B (Workers)
-    local worker_model="${WORKER_MODEL:-${SCRIPT_DIR}/models/Qwen3.5/Qwen3.5-4B-Q5_K_M.gguf}"
+    local worker_model="${LLM_WORKER_MODEL:-${WORKER_MODEL:-${SCRIPT_DIR}/models/Qwen3.5/Qwen3.5-4B-Q5_K_M.gguf}}"
-    local worker_mmproj="${WORKER_MMPROJ:-${SCRIPT_DIR}/models/Qwen3.5/mmproj-BF16.gguf}"
+    local worker_mmproj="${LLM_WORKER_MMPROJ:-${WORKER_MMPROJ:-${SCRIPT_DIR}/models/Qwen3.5/mmproj-BF16.gguf}}"
-    local worker_tmpl="${WORKER_TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.5/chat_template.jinja}"
+    if [ "${worker_mmproj}" = "none" ]; then worker_mmproj=""; fi
+    local worker_tmpl="${LLM_WORKER_TEMPLATE:-${WORKER_TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.5/chat_template.jinja}}"
 
     # 1. Planner on 11434
-    launch_instance "planner_27b" "$planner_model" "${PLANNER_PORT:-11434}" "${PLANNER_CTX:-65536}" \
+    launch_instance "planner_27b" "$planner_model" "${LLM_PLANNER_PORT:-${PLANNER_PORT:-11434}}" "${LLM_PLANNER_CTX:-${PLANNER_CTX:-65536}}" \
-        "$planner_mmproj" "$planner_tmpl" "${PLANNER_TEMP:-0.6}" "${PLANNER_TOP_P:-0.95}" "${PLANNER_TOP_K:-40}" \
+        "$planner_mmproj" "$planner_tmpl" "${LLM_PLANNER_TEMP:-${PLANNER_TEMP:-0.6}}" "${LLM_PLANNER_TOP_P:-${PLANNER_TOP_P:-0.95}}" "${LLM_PLANNER_TOP_K:-${PLANNER_TOP_K:-40}}" \
-        "${PLANNER_MIN_P:-0.05}" "${PLANNER_NGL:-99}" "${PLANNER_LOAD_MODE:-none}" "${FLASH_ATTN:-on}" \
+        "${LLM_PLANNER_MIN_P:-${PLANNER_MIN_P:-0.05}}" "${LLM_PLANNER_NGL:-${PLANNER_NGL:-99}}" "${LLM_PLANNER_LOAD_MODE:-${PLANNER_LOAD_MODE:-none}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
-        "${PLANNER_CTK:-q8_0}" "${PLANNER_CTV:-q8_0}" "false"
+        "${LLM_PLANNER_CTK:-${PLANNER_CTK:-q8_0}}" "${LLM_PLANNER_CTV:-${PLANNER_CTV:-q8_0}}" "false"
 
     # 2. Worker 1 on 11435
-    launch_instance "worker_4b_1" "$worker_model" "${WORKER1_PORT:-11435}" "${WORKER1_CTX:-16384}" \
+    launch_instance "worker_4b_1" "$worker_model" "${LLM_WORKER1_PORT:-${WORKER1_PORT:-11435}}" "${LLM_WORKER1_CTX:-${WORKER1_CTX:-16384}}" \
-        "$worker_mmproj" "$worker_tmpl" "${WORKER1_TEMP:-0.7}" "${WORKER1_TOP_P:-0.8}" "${WORKER1_TOP_K:-20}" \
+        "$worker_mmproj" "$worker_tmpl" "${LLM_WORKER1_TEMP:-${WORKER1_TEMP:-0.7}}" "${LLM_WORKER1_TOP_P:-${WORKER1_TOP_P:-0.8}}" "${LLM_WORKER1_TOP_K:-${WORKER1_TOP_K:-20}}" \
-        "${WORKER1_MIN_P:-0.05}" "${WORKER1_NGL:-99}" "${LOAD_MODE:-mmap}" "${FLASH_ATTN:-on}" \
+        "${LLM_WORKER1_MIN_P:-${WORKER1_MIN_P:-0.05}}" "${LLM_WORKER1_NGL:-${WORKER1_NGL:-99}}" "${LLM_LOAD_MODE:-${LOAD_MODE:-mmap}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
-        "${WORKER1_CTK:-q8_0}" "${WORKER1_CTV:-q8_0}" "false"
+        "${LLM_WORKER1_CTK:-${WORKER1_CTK:-q8_0}}" "${LLM_WORKER1_CTV:-${WORKER1_CTV:-q8_0}}" "false"
 
     # 3. Worker 2 on 11436
-    launch_instance "worker_4b_2" "$worker_model" "${WORKER2_PORT:-11436}" "${WORKER2_CTX:-16384}" \
+    launch_instance "worker_4b_2" "$worker_model" "${LLM_WORKER2_PORT:-${WORKER2_PORT:-11436}}" "${LLM_WORKER2_CTX:-${WORKER2_CTX:-16384}}" \
-        "$worker_mmproj" "$worker_tmpl" "${WORKER2_TEMP:-0.7}" "${WORKER2_TOP_P:-0.8}" "${WORKER2_TOP_K:-20}" \
+        "$worker_mmproj" "$worker_tmpl" "${LLM_WORKER2_TEMP:-${WORKER2_TEMP:-0.7}}" "${LLM_WORKER2_TOP_P:-${WORKER2_TOP_P:-0.8}}" "${LLM_WORKER2_TOP_K:-${WORKER2_TOP_K:-20}}" \
-        "${WORKER2_MIN_P:-0.05}" "${WORKER2_NGL:-99}" "${LOAD_MODE:-mmap}" "${FLASH_ATTN:-on}" \
+        "${LLM_WORKER2_MIN_P:-${WORKER2_MIN_P:-0.05}}" "${LLM_WORKER2_NGL:-${WORKER2_NGL:-99}}" "${LLM_LOAD_MODE:-${LOAD_MODE:-mmap}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
-        "${WORKER2_CTK:-q8_0}" "${WORKER2_CTV:-q8_0}" "false"
+        "${LLM_WORKER2_CTK:-${WORKER2_CTK:-q8_0}}" "${LLM_WORKER2_CTV:-${WORKER2_CTV:-q8_0}}" "false"
 
     echo ""
     log_success "All 3 cluster servers are operational:"
-    echo "  [Planner]     SuperQwen 3.8 27B -> http://localhost:${PLANNER_PORT:-11434}"
+    echo "  [Planner]     SuperQwen 3.8 27B -> http://localhost:${LLM_PLANNER_PORT:-${PLANNER_PORT:-11434}}"
-    echo "  [Implementer] Qwen 3.5 4B #1    -> http://localhost:${WORKER1_PORT:-11435}"
+    echo "  [Implementer] Qwen 3.5 4B #1    -> http://localhost:${LLM_WORKER1_PORT:-${WORKER1_PORT:-11435}}"
-    echo "  [Implementer] Qwen 3.5 4B #2    -> http://localhost:${WORKER2_PORT:-11436}"
+    echo "  [Implementer] Qwen 3.5 4B #2    -> http://localhost:${LLM_WORKER2_PORT:-${WORKER2_PORT:-11436}}"
     echo ""
     echo "Logs directory: ${LOG_DIR}/"
     echo "Press Ctrl+C to stop all instances gracefully."
@@ -242,14 +250,20 @@ cmd_cluster() {
 
 cmd_muse() {
     log_info "Starting Muse-Glimmer 30B Solo Server (Max Context: 128K)..."
-    local model="${MODEL:-${SCRIPT_DIR}/models/muse-glimmer/Muse-Glimmer-30B-UD-Q4_K_XL.gguf}"
+    pkill -f "llama-server" 2>/dev/null || true
-    local mmproj="${MMPROJ:-${SCRIPT_DIR}/models/muse-glimmer/mmproj-Muse-Glimmer-30B-Q8_0.gguf}"
+    sleep 1
-    local tmpl="${TEMPLATE:-${SCRIPT_DIR}/templates/muse-glimmer/chat_template.jinja}"
+    local model="${LLM_MODEL:-${MODEL:-${SCRIPT_DIR}/models/muse-glimmer/Muse-Glimmer-30B-UD-Q4_K_XL.gguf}}"
+    local mmproj="${LLM_MMPROJ:-${MMPROJ:-${SCRIPT_DIR}/models/muse-glimmer/mmproj-Muse-Glimmer-30B-Q8_0.gguf}}"
+    if [ "${mmproj}" = "none" ]; then mmproj=""; fi
+    local tmpl="${LLM_TEMPLATE:-${TEMPLATE:-${SCRIPT_DIR}/templates/muse-glimmer/chat_template.jinja}}"
 
-    launch_instance "muse_solo" "$model" "${PORT:-11434}" "${CTX:-131072}" \
+    launch_instance "muse_solo" "$model" "${LLM_PORT:-${PORT:-11434}}" "${LLM_CTX:-${CTX:-131072}}" \
-        "$mmproj" "$tmpl" "${TEMP:-0.6}" "${TOP_P:-0.95}" "${TOP_K:-40}" \
+        "$mmproj" "$tmpl" "${LLM_TEMPERATURE:-${TEMPERATURE:-${SAMPLING_TEMP:-0.6}}}" \
-        "${MIN_P:-0.05}" "${NGL:-99}" "${LOAD_MODE:-none}" "${FLASH_ATTN:-on}" \
+        "${LLM_TOP_P:-${TOP_P:-0.95}}" "${LLM_TOP_K:-${TOP_K:-40}}" \
-        "${CTK:-q8_0}" "${CTV:-q8_0}" "${REASONING_PRESERVE:-true}"
+        "${LLM_MIN_P:-${MIN_P:-0.05}}" "${LLM_NGL:-${NGL:-99}}" \
+        "${LLM_LOAD_MODE:-${LOAD_MODE:-none}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
+        "${LLM_CTK:-${CTK:-q8_0}}" "${LLM_CTV:-${CTV:-q8_0}}" \
+        "${LLM_REASONING_PRESERVE:-${REASONING_PRESERVE:-true}}"
 
     echo ""
     echo "Streaming logs: tail -f ${LOG_DIR}/muse_solo.log"
@@ -259,17 +273,23 @@ cmd_muse() {
 
 cmd_superqwen() {
     log_info "Starting SuperQwen 3.8 27B Solo Server (Max Context: 256K)..."
-    local model="${MODEL:-${SCRIPT_DIR}/models/Qwen3.8/SuperQwen3.8-27b-abliterated-Q4_K_M.gguf}"
+    pkill -f "llama-server" 2>/dev/null || true
+    sleep 1
+    local model="${LLM_MODEL:-${MODEL:-${SCRIPT_DIR}/models/Qwen3.8/SuperQwen3.8-27b-abliterated-Q4_K_M.gguf}}"
     if [ ! -f "${model}" ] && [ -f "${SCRIPT_DIR}/models/Qwen3.8/Qwen3.8-27B-UD-Q4_K_XL.gguf" ]; then
         model="${SCRIPT_DIR}/models/Qwen3.8/Qwen3.8-27B-UD-Q4_K_XL.gguf"
     fi
-    local mmproj="${MMPROJ:-${SCRIPT_DIR}/models/Qwen3.8/mmproj-SuperQwen3.8-27b-abliterated-Q8_0.gguf}"
+    local mmproj="${LLM_MMPROJ:-${MMPROJ:-${SCRIPT_DIR}/models/Qwen3.8/mmproj-SuperQwen3.8-27b-abliterated-Q8_0.gguf}}"
-    local tmpl="${TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.8/chat_template.jinja}"
+    if [ "${mmproj}" = "none" ]; then mmproj=""; fi
+    local tmpl="${LLM_TEMPLATE:-${TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.8/chat_template.jinja}}"
 
-    launch_instance "superqwen_solo" "$model" "${PORT:-11434}" "${CTX:-262144}" \
+    launch_instance "superqwen_solo" "$model" "${LLM_PORT:-${PORT:-11434}}" "${LLM_CTX:-${CTX:-262144}}" \
-        "$mmproj" "$tmpl" "${TEMP:-0.6}" "${TOP_P:-0.95}" "${TOP_K:-40}" \
+        "$mmproj" "$tmpl" "${LLM_TEMPERATURE:-${TEMPERATURE:-${SAMPLING_TEMP:-0.6}}}" \
-        "${MIN_P:-0.05}" "${NGL:-99}" "${LOAD_MODE:-none}" "${FLASH_ATTN:-on}" \
+        "${LLM_TOP_P:-${TOP_P:-0.95}}" "${LLM_TOP_K:-${TOP_K:-40}}" \
-        "${CTK:-q8_0}" "${CTV:-q8_0}" "${REASONING_PRESERVE:-true}"
+        "${LLM_MIN_P:-${MIN_P:-0.05}}" "${LLM_NGL:-${NGL:-99}}" \
+        "${LLM_LOAD_MODE:-${LOAD_MODE:-none}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
+        "${LLM_CTK:-${CTK:-q8_0}}" "${LLM_CTV:-${CTV:-q8_0}}" \
+        "${LLM_REASONING_PRESERVE:-${REASONING_PRESERVE:-true}}"
 
     echo ""
     echo "Streaming logs: tail -f ${LOG_DIR}/superqwen_solo.log"
@@ -279,14 +299,19 @@ cmd_superqwen() {
 
 cmd_qwen4b() {
     log_info "Starting Qwen 3.5 4B Solo Server..."
-    local model="${MODEL:-${SCRIPT_DIR}/models/Qwen3.5/Qwen3.5-4B-Q5_K_M.gguf}"
+    pkill -f "llama-server" 2>/dev/null || true
-    local mmproj="${MMPROJ:-${SCRIPT_DIR}/models/Qwen3.5/mmproj-BF16.gguf}"
+    sleep 1
-    local tmpl="${TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.5/chat_template.jinja}"
+    local model="${LLM_MODEL:-${MODEL:-${SCRIPT_DIR}/models/Qwen3.5/Qwen3.5-4B-Q5_K_M.gguf}}"
+    local mmproj="${LLM_MMPROJ:-${MMPROJ:-${SCRIPT_DIR}/models/Qwen3.5/mmproj-BF16.gguf}}"
+    if [ "${mmproj}" = "none" ]; then mmproj=""; fi
+    local tmpl="${LLM_TEMPLATE:-${TEMPLATE:-${SCRIPT_DIR}/templates/Qwen3.5/chat_template.jinja}}"
 
-    launch_instance "qwen4b_solo" "$model" "${PORT:-11434}" "${CTX:-32768}" \
+    launch_instance "qwen4b_solo" "$model" "${LLM_PORT:-${PORT:-11434}}" "${LLM_CTX:-${CTX:-32768}}" \
-        "$mmproj" "$tmpl" "${TEMP:-0.7}" "${TOP_P:-0.8}" "${TOP_K:-20}" \
+        "$mmproj" "$tmpl" "${LLM_TEMPERATURE:-${TEMPERATURE:-${SAMPLING_TEMP:-0.7}}}" \
-        "${MIN_P:-0.05}" "${NGL:-99}" "${LOAD_MODE:-mmap}" "${FLASH_ATTN:-on}" \
+        "${LLM_TOP_P:-${TOP_P:-0.8}}" "${LLM_TOP_K:-${TOP_K:-20}}" \
-        "${CTK:-q8_0}" "${CTV:-q8_0}" "false"
+        "${LLM_MIN_P:-${MIN_P:-0.05}}" "${LLM_NGL:-${NGL:-99}}" \
+        "${LLM_LOAD_MODE:-${LOAD_MODE:-mmap}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
+        "${LLM_CTK:-${CTK:-q8_0}}" "${LLM_CTV:-${CTV:-q8_0}}" "false"
 
     echo ""
     echo "Streaming logs: tail -f ${LOG_DIR}/qwen4b_solo.log"
@@ -294,6 +319,34 @@ cmd_qwen4b() {
     wait
 }
 
+cmd_custom() {
+    log_info "Starting Custom Server Instance..."
+    pkill -f "llama-server" 2>/dev/null || true
+    sleep 1
+    local model="${LLM_MODEL:-${MODEL:-}}"
+    if [ -z "${model}" ] || [ ! -f "${model}" ]; then
+        log_error "No valid MODEL specified! Pass LLM_MODEL=/path/to/model.gguf"
+        exit 1
+    fi
+    local mmproj="${LLM_MMPROJ:-${MMPROJ:-}}"
+    if [ "${mmproj}" = "none" ]; then mmproj=""; fi
+    local tmpl="${LLM_TEMPLATE:-${TEMPLATE:-}}"
+    local name="custom_solo"
+
+    launch_instance "${name}" "$model" "${LLM_PORT:-${PORT:-11434}}" "${LLM_CTX:-${CTX:-65536}}" \
+        "$mmproj" "$tmpl" "${LLM_TEMPERATURE:-${TEMPERATURE:-${SAMPLING_TEMP:-0.6}}}" \
+        "${LLM_TOP_P:-${TOP_P:-0.95}}" "${LLM_TOP_K:-${TOP_K:-40}}" \
+        "${LLM_MIN_P:-${MIN_P:-0.05}}" "${LLM_NGL:-${NGL:-99}}" \
+        "${LLM_LOAD_MODE:-${LOAD_MODE:-none}}" "${LLM_FLASH_ATTN:-${FLASH_ATTN:-on}}" \
+        "${LLM_CTK:-${CTK:-q8_0}}" "${LLM_CTV:-${CTV:-q8_0}}" \
+        "${LLM_REASONING_PRESERVE:-${REASONING_PRESERVE:-false}}"
+
+    echo ""
+    echo "Streaming logs: tail -f ${LOG_DIR}/${name}.log"
+    echo "Press Ctrl+C to stop."
+    wait
+}
+
 # ------------------------------------------------------------------------------
 # Help / Usage
 # ------------------------------------------------------------------------------
@@ -306,16 +359,22 @@ Commands:
   muse          Launch Muse-Glimmer 30B Solo (128K context, port 11434)
   superqwen     Launch SuperQwen 3.8 27B Solo (256K context, port 11434)
   qwen4b        Launch Qwen 3.5 4B Solo (port 11434)
+  custom        Launch custom model instance (requires LLM_MODEL=/path/to/model.gguf)
   status        Show current server health and active ports
   stop [target] Stop servers (target: "all" or port number like 11434)
   help          Display this help message
 
 Environment Variables:
-  HOST          Bind address (default: 0.0.0.0)
+  LLM_MODEL         Path to GGUF model file (fallback: MODEL)
-  PORT          Main port override (default: 11434)
+  LLM_MMPROJ        Path to mmproj file, or "none" (fallback: MMPROJ)
-  CTX           Context size override
+  LLM_TEMPLATE      Path to chat template file (fallback: TEMPLATE)
-  FLASH_ATTN    Enable Flash Attention: "on" or "off" (default: on)
+  LLM_HOST          Bind address (default: 0.0.0.0)
-  LOG_DIR       Path to write logs (default: ./logs)
+  LLM_PORT          Main port override (default: 11434)
+  LLM_CTX           Context size override (default: preset-specific)
+  LLM_TEMPERATURE   Sampling temperature (default: preset-specific)
+  LLM_PARALLEL      Parallel slot count (default: 1)
+  LLM_FLASH_ATTN    Enable Flash Attention: "on" or "off" (default: on)
+  LLM_LOG_DIR       Path to write logs (default: ./logs)
 
 EOF
 }
@@ -331,6 +390,7 @@ case "${ACTION}" in
     muse)       cmd_muse "$@" ;;
     superqwen)  cmd_superqwen "$@" ;;
     qwen4b)     cmd_qwen4b "$@" ;;
+    custom)     cmd_custom "$@" ;;
     status)     cmd_status "$@" ;;
     stop)       cmd_stop "$@" ;;
     help|--help|-h) cmd_help ;;

bin/llm-tui

@@ -14,6 +14,7 @@ import signal
 import subprocess
 import urllib.request
 import json
+from dataclasses import dataclass
 from pathlib import Path
 from typing import Optional
 
@@ -45,10 +46,22 @@ RUNNER_BIN = SCRIPT_DIR / "bin" / "llm-runner"
 
 TOTAL_USABLE_VRAM_GB = 30.0
 
+
+@dataclass
+class ModelProfile:
+    name: str
+    path: Path
+    size_gb: float
+    directory: Path
+    paired_mmproj: Optional[Path]
+    paired_template: Optional[Path]
+    preset_key: str
+
+
 PRESETS = {
     "cluster": {
         "name": "🌟 Multi-Agent Cluster (SuperQwen 27B + 2x Qwen 4B)",
-        "desc": "Planner on 11434, Worker 1 on 11435, Worker 2 on 11436. Optimized for Pi & Hermes agents.",
+        "desc": "Planner on 11434 (SuperQwen 27B + mmproj), Worker 1 on 11435, Worker 2 on 11436 (Qwen 4B + mmproj). Optimized for Pi & Hermes agents.",
         "port": 11434,
         "ctx": 65536,
         "temp": 0.6,
@@ -59,10 +72,12 @@ PRESETS = {
         "flash_attn": True,
         "kv_quant": "q8_0",
         "command": "cluster",
+        "target_model_match": "SuperQwen",
+        "is_cluster": True,
     },
     "muse": {
         "name": "🔮 Muse-Glimmer 30B Solo (128K Context)",
-        "desc": "Dedicated deep-reasoning agent with max 128K context on port 11434.",
+        "desc": "Dedicated deep-reasoning agent with max 128K context on port 11434. Auto-pairs Muse-Glimmer Q8_0 vision projector.",
         "port": 11434,
         "ctx": 131072,
         "temp": 0.6,
@@ -73,10 +88,12 @@ PRESETS = {
         "flash_attn": True,
         "kv_quant": "q8_0",
         "command": "muse",
+        "target_model_match": "Muse-Glimmer",
+        "is_cluster": False,
     },
     "superqwen": {
         "name": "🧠 SuperQwen 3.8 27B Solo (256K Context)",
-        "desc": "Dedicated ultra-long reasoning agent with max 256K context on port 11434.",
+        "desc": "Dedicated ultra-long reasoning agent with max 256K context on port 11434. Auto-pairs SuperQwen Q8_0 vision projector.",
         "port": 11434,
         "ctx": 262144,
         "temp": 0.6,
@@ -87,10 +104,12 @@ PRESETS = {
         "flash_attn": True,
         "kv_quant": "q8_0",
         "command": "superqwen",
+        "target_model_match": "SuperQwen",
+        "is_cluster": False,
     },
     "qwen4b": {
         "name": "⚡ Qwen 3.5 4B Solo (Fast Instruct)",
-        "desc": "Lightweight high-speed instruct model on port 11434 with ~3.9GB VRAM footprint.",
+        "desc": "Lightweight high-speed instruct model on port 11434 with ~3.9GB VRAM footprint. Auto-pairs BF16 vision projector.",
         "port": 11434,
         "ctx": 32768,
         "temp": 0.7,
@@ -101,6 +120,8 @@ PRESETS = {
         "flash_attn": True,
         "kv_quant": "q8_0",
         "command": "qwen4b",
+        "target_model_match": "Qwen3.5-4B",
+        "is_cluster": False,
     },
     "custom": {
         "name": "🛠️ Custom Model & Parameters",
@@ -115,28 +136,80 @@ PRESETS = {
         "flash_attn": True,
         "kv_quant": "q8_0",
         "command": "custom",
+        "target_model_match": None,
+        "is_cluster": False,
     },
 }
 
 
-def find_gguf_models() -> list[tuple[str, str]]:
+def scan_models() -> tuple[list[ModelProfile], list[tuple[str, str]], list[tuple[str, str]]]:
-    models: list[tuple[str, str]] = []
+    profiles: list[ModelProfile] = []
-    if MODELS_DIR.exists():
+    model_options: list[tuple[str, str]] = []
-        for path in sorted(MODELS_DIR.rglob("*.gguf")):
+    mmproj_options: list[tuple[str, str]] = [("None (Text Only / Vision Disabled)", "none")]
-            if "mmproj" not in path.name.lower():
-                rel = path.relative_to(MODELS_DIR)
-                models.append((f"{path.name} ({path.stat().st_size / 1e9:.1f} GB)", str(path)))
-    if not models:
-        models.append(("No GGUF models found in models/", ""))
-    return models
 
+    if not MODELS_DIR.exists():
+        return profiles, [("No GGUF models found in models/", "")], mmproj_options
 
-def find_mmproj_files() -> list[tuple[str, str]]:
+    # Gather all mmproj projector files first
-    projectors = [("None (Text Only)", "")]
+    all_mmproj = sorted(MODELS_DIR.rglob("mmproj*.gguf"))
-    if MODELS_DIR.exists():
+    for p in all_mmproj:
-        for path in sorted(MODELS_DIR.rglob("mmproj*.gguf")):
+        try:
-            projectors.append((f"{path.name} ({path.stat().st_size / 1e9:.1f} GB)", str(path)))
+            sz = p.stat().st_size / 1e9
-    return projectors
+        except Exception:
+            sz = 0.0
+        mmproj_options.append((f"{p.name} ({sz:.1f} GB) [{p.parent.name}]", str(p)))
+
+    # Gather base model GGUF files
+    for p in sorted(MODELS_DIR.rglob("*.gguf")):
+        if "mmproj" in p.name.lower():
+            continue
+        try:
+            sz = p.stat().st_size / 1e9
+        except Exception:
+            sz = 0.0
+
+        # Find paired mmproj in same folder
+        paired_mmproj = None
+        for m in all_mmproj:
+            if m.parent == p.parent:
+                paired_mmproj = m
+                break
+
+        # Find paired chat template in directory
+        tmpl = p.parent / "chat_template.jinja"
+        paired_template = tmpl if tmpl.exists() else None
+
+        # Determine preset mapping
+        p_lower = (p.name + " " + p.parent.name).lower()
+        if "muse" in p_lower:
+            preset_key = "muse"
+            tag = "Solo Reasoning (128K)"
+        elif "superqwen" in p_lower or "qwen3.8" in p_lower:
+            preset_key = "superqwen"
+            tag = "Solo 256K / Planner"
+        elif "qwen3.5" in p_lower or "4b" in p_lower:
+            preset_key = "qwen4b"
+            tag = "Fast Instruct / Worker"
+        else:
+            preset_key = "custom"
+            tag = "Custom"
+
+        prof = ModelProfile(
+            name=p.name,
+            path=p,
+            size_gb=sz,
+            directory=p.parent,
+            paired_mmproj=paired_mmproj,
+            paired_template=paired_template,
+            preset_key=preset_key,
+        )
+        profiles.append(prof)
+        model_options.append((f"{p.name} ({sz:.1f} GB) — {tag}", str(p)))
+
+    if not model_options:
+        model_options.append(("No GGUF models found in models/", ""))
+
+    return profiles, model_options, mmproj_options
 
 
 def check_port_status(port: int) -> tuple[bool, str]:
@@ -147,6 +220,15 @@ def check_port_status(port: int) -> tuple[bool, str]:
             data = json.loads(response.read().decode())
             status = data.get("status", "unknown")
             return True, f"Online ({status})"
+    except urllib.error.HTTPError as e:
+        if e.code == 503:
+            try:
+                data = json.loads(e.read().decode())
+                msg = data.get("error", {}).get("message", "Loading model")
+                return True, f"Loading ({msg})"
+            except Exception:
+                return True, "Loading model..."
+        return False, f"HTTP Error {e.code}"
     except Exception:
         return False, "Offline"
 
@@ -173,6 +255,7 @@ class LLMConfigApp(App):
         border: solid #3b4252;
         padding: 1 2;
         margin-bottom: 1;
+        height: auto;
     }
     .title {
         color: #88c0d0;
@@ -224,17 +307,26 @@ class LLMConfigApp(App):
         text-style: bold;
         border: solid #d08770;
     }
+    #binding-indicator {
+        background: #1e2531;
+        border-left: thick #88c0d0;
+        padding: 1 2;
+        margin-bottom: 1;
+        height: auto;
+    }
     #vram-meter {
         background: #232834;
         border: solid #4c566a;
         padding: 1;
         margin-top: 1;
+        height: auto;
     }
     #status-card {
         background: #1e222b;
         border: solid #4c566a;
         padding: 1;
         margin-bottom: 1;
+        height: auto;
     }
     Log {
         background: #0f1117;
@@ -253,13 +345,37 @@ class LLMConfigApp(App):
     ]
 
     selected_preset: str = "cluster"
-    models_list: list[tuple[str, str]] = []
+    profiles: list[ModelProfile] = []
-    mmproj_list: list[tuple[str, str]] = []
+    profiles_by_path: dict[str, ModelProfile] = {}
+    models_options: list[tuple[str, str]] = []
+    mmproj_options: list[tuple[str, str]] = []
+    _sync_lock: bool = False
+    _last_selected_model: str = ""
+    _last_selected_mmproj: str = ""
+    _last_log_size: int = -1
+    _last_log_name: str = ""
 
     def compose(self) -> ComposeResult:
         yield Header(show_clock=True)
-        self.models_list = find_gguf_models()
+        self.profiles, self.models_options, self.mmproj_options = scan_models()
-        self.mmproj_list = find_mmproj_files()
+        self.profiles_by_path = {str(p.path): p for p in self.profiles}
+
+        # Choose default model and mmproj based on initial cluster preset
+        default_model = ""
+        default_mmproj = "none"
+        for p in self.profiles:
+            if "superqwen" in p.name.lower() or "qwen3.8" in p.name.lower():
+                default_model = str(p.path)
+                if p.paired_mmproj:
+                    default_mmproj = str(p.paired_mmproj)
+                break
+        if not default_model and self.profiles:
+            default_model = str(self.profiles[0].path)
+            if self.profiles[0].paired_mmproj:
+                default_mmproj = str(self.profiles[0].paired_mmproj)
+
+        self._last_selected_model = default_model
+        self._last_selected_mmproj = default_mmproj
 
         with TabbedContent():
             with TabPane("🚀 Model Launcher & Form", id="tab-launch"):
@@ -280,23 +396,26 @@ class LLMConfigApp(App):
                             classes="text-muted",
                         )
 
-                    # 2. Parameters Form
+                    # 2. Active Tied Binding Indicator Card
+                    yield Static(id="binding-indicator")
+
+                    # 3. Parameters Form (GGUF, MMProj, and Tuning)
                     with Container(classes="card", id="form-container"):
                         yield Label("Hyperparameters & Hardware Tuning", classes="title")
 
                         with Horizontal(classes="field-row"):
                             yield Label("GGUF Model:", classes="field-label")
                             yield Select(
-                                options=[(lbl, val) for lbl, val in self.models_list],
+                                options=[(lbl, val) for lbl, val in self.models_options],
-                                value=self.models_list[0][1] if self.models_list else None,
+                                value=default_model if default_model else None,
                                 id="select-model",
                             )
 
                         with Horizontal(classes="field-row"):
                             yield Label("Vision MMProj:", classes="field-label")
                             yield Select(
-                                options=[(lbl, val) for lbl, val in self.mmproj_list],
+                                options=[(lbl, val) for lbl, val in self.mmproj_options],
-                                value=self.mmproj_list[0][1] if self.mmproj_list else None,
+                                value=default_mmproj,
                                 id="select-mmproj",
                             )
 
@@ -339,7 +458,7 @@ class LLMConfigApp(App):
                         # VRAM Estimation Banner
                         yield Static(id="vram-meter")
 
-                    # 3. Execution Action Bar
+                    # 4. Execution Action Bar
                     with Horizontal(classes="card"):
                         yield Button("🚀 Launch Server (F5)", variant="primary", id="btn-launch")
                         yield Button("🛑 Stop Server (F8)", variant="error", id="btn-stop")
@@ -355,15 +474,25 @@ class LLMConfigApp(App):
         yield Footer()
 
     def on_mount(self) -> None:
-        self.update_vram_estimation()
+        self.apply_preset(self.selected_preset)
         self.action_refresh_status()
         self.set_interval(3.0, self.action_refresh_status)
         self.set_interval(1.5, self.tail_logs)
 
-    def on_radio_set_changed(self, event: RadioSet.Changed) -> None:
+    def _set_active_radio(self, preset_key: str) -> None:
-        radio_id = event.pressed.id or ""
+        rs = self.query_one("#preset-radios", RadioSet)
-        preset_key = radio_id.replace("preset-", "")
+        target = self.query_one(f"#preset-{preset_key}", RadioButton)
-        if preset_key in PRESETS:
+        with rs.prevent(RadioButton.Changed):
+            if rs._pressed_button and rs._pressed_button != target:
+                rs._pressed_button.value = False
+            target.value = True
+            rs._pressed_button = target
+
+    def apply_preset(self, preset_key: str) -> None:
+        if preset_key not in PRESETS:
+            return
+        self._sync_lock = True
+        try:
             self.selected_preset = preset_key
             preset = PRESETS[preset_key]
             self.query_one("#preset-description", Label).update(preset["desc"])
@@ -375,13 +504,134 @@ class LLMConfigApp(App):
             self.query_one("#input-ngl", Input).value = str(preset["ngl"])
             self.query_one("#check-fa", Checkbox).value = preset["flash_attn"]
             self.query_one("#select-kv", Select).value = preset["kv_quant"]
+
+            # Automatically tie GGUF model and MMProj to preset
+            prof = self._find_profile_for_preset(preset_key)
+            if prof:
+                self._last_selected_model = str(prof.path)
+                self.query_one("#select-model", Select).value = str(prof.path)
+                target_mmproj = str(prof.paired_mmproj) if prof.paired_mmproj else "none"
+                self._last_selected_mmproj = target_mmproj
+                self.query_one("#select-mmproj", Select).value = target_mmproj
+
+            # Sync active RadioButton in RadioSet
+            self._set_active_radio(preset_key)
+
+            self.update_binding_indicator()
+            self.update_vram_estimation()
+        finally:
+            self._sync_lock = False
+
+    def _find_profile_for_preset(self, preset_key: str) -> Optional[ModelProfile]:
+        preset = PRESETS.get(preset_key)
+        if not preset or not preset.get("target_model_match"):
+            return None
+        target = preset["target_model_match"].lower()
+        for prof in self.profiles:
+            if target in prof.name.lower() or target in prof.directory.name.lower():
+                return prof
+        return None
+
+    def on_radio_set_changed(self, event: RadioSet.Changed) -> None:
+        if self._sync_lock:
+            return
+        radio_id = event.pressed.id or ""
+        preset_key = radio_id.replace("preset-", "")
+        self.apply_preset(preset_key)
+
+    def on_select_changed(self, event: Select.Changed) -> None:
+        if self._sync_lock:
+            return
+
+        if event.select.id == "select-model":
+            model_path_str = str(event.value or "")
+            if model_path_str == self._last_selected_model:
+                return
+            self._last_selected_model = model_path_str
+
+            self._sync_lock = True
+            try:
+                prof = self.profiles_by_path.get(model_path_str)
+                if prof:
+                    # 1. Automatically tie matching MMProj
+                    if prof.paired_mmproj:
+                        self._last_selected_mmproj = str(prof.paired_mmproj)
+                        self.query_one("#select-mmproj", Select).value = str(prof.paired_mmproj)
+                    else:
+                        self._last_selected_mmproj = "none"
+                        self.query_one("#select-mmproj", Select).value = "none"
+
+                    # 2. Automatically sync inference preset and sampling defaults
+                    p_key = prof.preset_key
+                    if self.selected_preset == "cluster" and p_key in ("superqwen", "qwen4b"):
+                        # SuperQwen and Qwen 4B are the Planner and Workers of Cluster mode
+                        pass
+                    elif p_key != "custom" and p_key in PRESETS:
+                        self.selected_preset = p_key
+                        self._set_active_radio(p_key)
+                        preset = PRESETS[p_key]
+                        self.query_one("#preset-description", Label).update(preset["desc"])
+                        self.query_one("#input-port", Input).value = str(preset["port"])
+                        self.query_one("#input-ctx", Input).value = str(preset["ctx"])
+                        self.query_one("#input-temp", Input).value = str(preset["temp"])
+                        self.query_one("#input-top-p", Input).value = str(preset["top_p"])
+                        self.query_one("#input-min-p", Input).value = str(preset["min_p"])
+                        self.query_one("#input-ngl", Input).value = str(preset["ngl"])
+                        self.query_one("#check-fa", Checkbox).value = preset["flash_attn"]
+                        self.query_one("#select-kv", Select).value = preset["kv_quant"]
+                    else:
+                        self.selected_preset = "custom"
+                        self._set_active_radio("custom")
+                        self.query_one("#preset-description", Label).update(PRESETS["custom"]["desc"])
+
+                self.update_binding_indicator()
+                self.update_vram_estimation()
+            finally:
+                self._sync_lock = False
+
+        elif event.select.id == "select-mmproj":
+            mmproj_str = str(event.value or "")
+            if mmproj_str == self._last_selected_mmproj:
+                return
+            self._last_selected_mmproj = mmproj_str
+            self.update_binding_indicator()
+            self.update_vram_estimation()
+
+        elif event.select.id == "select-kv":
+            self.update_binding_indicator()
             self.update_vram_estimation()
 
     def on_input_changed(self, event: Input.Changed) -> None:
         self.update_vram_estimation()
 
-    def on_select_changed(self, event: Select.Changed) -> None:
+    def update_binding_indicator(self) -> None:
-        self.update_vram_estimation()
+        preset_key = self.selected_preset
+        preset = PRESETS.get(preset_key, PRESETS["custom"])
+        model_val = str(self.query_one("#select-model", Select).value or "")
+        mmproj_val = str(self.query_one("#select-mmproj", Select).value or "")
+
+        model_name = Path(model_val).name if (model_val and Path(model_val).exists()) else "None"
+        if mmproj_val and mmproj_val != "none" and Path(mmproj_val).exists():
+            mmproj_text = f"[green]✔ {Path(mmproj_val).name} (Vision Enabled)[/]"
+        else:
+            mmproj_text = "[yellow]○ None (Text Only / Vision Disabled)[/]"
+
+        if preset.get("is_cluster"):
+            text = (
+                f"[bold cyan]🔗 Tied Configuration:[/] [bold white]{preset['name']}[/]\n"
+                f"   • [bold]Planner (Port 11434):[/] SuperQwen 3.8 27B + mmproj-SuperQwen (64K Ctx)\n"
+                f"   • [bold]Workers (Ports 11435/11436):[/] 2x Qwen 3.5 4B + mmproj-BF16 (16K Ctx)\n"
+                f"   • [bold green]● GGUF weights, projectors, and inference presets are coordinated across all 3 nodes[/]"
+            )
+        else:
+            text = (
+                f"[bold cyan]🔗 Tied Configuration:[/] [bold white]{preset['name']}[/]\n"
+                f"   • [bold]GGUF Model:[/] [bold magenta]{model_name}[/]\n"
+                f"   • [bold]Vision Projector:[/] {mmproj_text}\n"
+                f"   • [bold]Chat Template:[/] Jinja template auto-paired from model directory"
+            )
+
+        self.query_one("#binding-indicator", Static).update(text)
 
     def update_vram_estimation(self) -> None:
         try:
@@ -390,25 +640,31 @@ class LLMConfigApp(App):
             ctx = 65536
 
         preset = self.selected_preset
+        kv_choice = str(self.query_one("#select-kv", Select).value or "q8_0")
+        kv_mult = 1.0 if kv_choice == "q8_0" else (2.0 if kv_choice == "f16" else 0.5)
+
         if preset == "cluster":
             weights_gb = 25.9
-            kv_gb = 3.2
+            kv_gb = 3.2 * kv_mult
-        elif preset == "muse":
-            weights_gb = 17.5
-            kv_gb = (ctx / 131072.0) * 8.5
-        elif preset == "superqwen":
-            weights_gb = 17.5
-            kv_gb = (ctx / 262144.0) * 10.5
-        elif preset == "qwen4b":
-            weights_gb = 3.9
-            kv_gb = (ctx / 32768.0) * 1.5
         else:
             model_val = str(self.query_one("#select-model", Select).value or "")
             if model_val and Path(model_val).exists():
-                weights_gb = Path(model_val).stat().st_size / 1e9
+                model_gb = Path(model_val).stat().st_size / 1e9
             else:
-                weights_gb = 15.0
+                model_gb = 16.0
-            kv_gb = (ctx / 65536.0) * 4.0
+
+            mmproj_val = str(self.query_one("#select-mmproj", Select).value or "")
+            if mmproj_val and mmproj_val != "none" and Path(mmproj_val).exists():
+                mmproj_gb = Path(mmproj_val).stat().st_size / 1e9
+            else:
+                mmproj_gb = 0.0
+
+            weights_gb = model_gb + mmproj_gb
+
+            if weights_gb > 10.0:
+                kv_gb = (ctx / 131072.0) * 8.5 * kv_mult
+            else:
+                kv_gb = (ctx / 32768.0) * 1.5 * kv_mult
 
         total_est = weights_gb + kv_gb
         remaining = TOTAL_USABLE_VRAM_GB - total_est
@@ -431,7 +687,12 @@ class LLMConfigApp(App):
         lines = []
         for port, label in [(11434, "Planner / Solo"), (11435, "Worker 1"), (11436, "Worker 2")]:
             online, msg = check_port_status(port)
-            badge = "[green]● ONLINE[/]" if online else "[yellow]○ OFFLINE[/]"
+            if not online:
+                badge = "[yellow]○ OFFLINE[/]"
+            elif "loading" in msg.lower():
+                badge = "[cyan]⏳ LOADING[/]"
+            else:
+                badge = "[green]● ONLINE[/]"
             lines.append(f"Port {port} ({label}): {badge} - {msg}")
 
         card = self.query_one("#status-card", Static)
@@ -443,8 +704,13 @@ class LLMConfigApp(App):
         if logs:
             latest = logs[0]
             try:
+                cur_size = latest.stat().st_size
+                if cur_size == self._last_log_size and latest.name == self._last_log_name:
+                    return
+                self._last_log_size = cur_size
+                self._last_log_name = latest.name
                 with open(latest, "r", errors="ignore") as f:
-                    f.seek(max(0, latest.stat().st_size - 8192))
+                    f.seek(max(0, cur_size - 16384))
                     tail = f.read()
                     log_widget.clear()
                     log_widget.write(f"--- Streaming {latest.name} ---\n" + tail)
@@ -454,22 +720,51 @@ class LLMConfigApp(App):
     def action_launch(self) -> None:
         preset = self.selected_preset
         log_widget = self.query_one("#log-stream", Log)
-        log_widget.write(f"\n[bold cyan]>>> Launching preset: {preset}...[/]\n")
+
+        model_val = str(self.query_one("#select-model", Select).value or "")
+        mmproj_val = str(self.query_one("#select-mmproj", Select).value or "")
 
         env = os.environ.copy()
-        env["CTX"] = self.query_one("#input-ctx", Input).value or "65536"
+        # Clean up POSIX temp directory variables that could collide with sampling parameters
-        env["PORT"] = self.query_one("#input-port", Input).value or "11434"
+        env.pop("TEMP", None)
-        env["TEMP"] = self.query_one("#input-temp", Input).value or "0.6"
+        env.pop("TMP", None)
-        env["TOP_P"] = self.query_one("#input-top-p", Input).value or "0.95"
-        env["MIN_P"] = self.query_one("#input-min-p", Input).value or "0.05"
-        env["NGL"] = self.query_one("#input-ngl", Input).value or "99"
-        env["FLASH_ATTN"] = "on" if self.query_one("#check-fa", Checkbox).value else "off"
-        env["CTK"] = str(self.query_one("#select-kv", Select).value or "q8_0")
-        env["CTV"] = env["CTK"]
 
-        cmd = [str(RUNNER_BIN), preset]
+        env["LLM_TEMPERATURE"] = self.query_one("#input-temp", Input).value or "0.6"
-        subprocess.Popen(cmd, env=env, cwd=str(SCRIPT_DIR), start_new_session=True)
+        env["LLM_CTX"] = self.query_one("#input-ctx", Input).value or "65536"
-        self.notify(f"Launching {preset} server in background. Checking ports...", severity="information")
+        env["LLM_PORT"] = self.query_one("#input-port", Input).value or "11434"
+        env["LLM_TOP_P"] = self.query_one("#input-top-p", Input).value or "0.95"
+        env["LLM_MIN_P"] = self.query_one("#input-min-p", Input).value or "0.05"
+        env["LLM_NGL"] = self.query_one("#input-ngl", Input).value or "99"
+        env["LLM_FLASH_ATTN"] = "on" if self.query_one("#check-fa", Checkbox).value else "off"
+        env["LLM_CTK"] = str(self.query_one("#select-kv", Select).value or "q8_0")
+        env["LLM_CTV"] = env["LLM_CTK"]
+        env["LLM_PARALLEL"] = "1"
+
+        if model_val and Path(model_val).exists():
+            env["LLM_MODEL"] = model_val
+            tmpl = Path(model_val).parent / "chat_template.jinja"
+            if tmpl.exists():
+                env["LLM_TEMPLATE"] = str(tmpl)
+
+        if mmproj_val and mmproj_val != "none" and Path(mmproj_val).exists():
+            env["LLM_MMPROJ"] = mmproj_val
+        else:
+            env["LLM_MMPROJ"] = "none"
+
+        runner_cmd = PRESETS.get(preset, {}).get("command", preset)
+        cmd = [str(RUNNER_BIN), runner_cmd]
+
+        log_widget.write(
+            f"\n[bold cyan]>>> Launching preset '{preset}' (runner command: {runner_cmd})...[/]\n"
+            f"    GGUF Model:    {Path(model_val).name if model_val else 'Default'}\n"
+            f"    Vision MMProj: {Path(mmproj_val).name if (mmproj_val and mmproj_val != 'none') else 'None (Text Only)'}\n"
+            f"    Port:          {env['LLM_PORT']} | Context: {env['LLM_CTX']} | KV Quant: {env['LLM_CTK']}\n"
+        )
+
+        runner_log = open(LOGS_DIR / "runner.log", "a")
+        subprocess.Popen(cmd, env=env, cwd=str(SCRIPT_DIR), stdout=runner_log, stderr=runner_log, start_new_session=True)
+        self.notify(f"Launching {preset} server. Switching to Status & Logs...", severity="information")
+        self.query_one(TabbedContent).active = "tab-status"
         self.action_refresh_status()
 
     def action_stop(self) -> None: