noonghunnaandClaude Opus 4.8
3fc26857f2
Replace the separate `image-studio` / `video-studio` / `comfyui` scenes with
a single `ai-studio` scene: ComfyUI holds both GPUs always, and image / video /
audio are picked as *lanes* in OWUI rather than gpu-mode switches. Removes the
gemma-12b chat brain from the studio bundle (kept on disk, just unwired) and
points OWUI's DEFAULT_MODELS at the qwen director.
- gpu-mode.sh: `mode_ai_studio` replaces the two studio modes; drop the gemma
start/stop + the GPU0-pin split; preflight now reads the shared manifest.
- scripts/lib/studio-models.tsv: single source-of-truth model manifest (modality,
label, root, rel_path, size, installer) consumed by both gpu-mode preflight and
the c3 cockpit, so the two can't drift.
- download_{director,video,audio,ace_step,stable_audio,studio}_models.sh: idempotent
fetchers; download_studio_models.sh orchestrates "grab everything".
- setup-video-studio.sh + repointed setup-image-studio.sh → `gpu-mode ai-studio`.
- litellm: drop the gemma-4-12b :8069 route; comfyui compose/entrypoint GPU-pin
comments corrected (no scene sets the pin now).
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm
916 lines
42 KiB
Bash
Executable File
916 lines
42 KiB
Bash
Executable File
#!/bin/bash
|
||
# GPU/RAM Mode Switcher for AI Inference Stack
|
||
# Manages Docker containers to avoid GPU/RAM contention on dual-3090 setup
|
||
# Location: club-3090/scripts/gpu-mode.sh (symlinked to /usr/local/bin/gpu-mode)
|
||
|
||
set -e
|
||
|
||
# club-3090 is the canonical repo (qwen36-dual-3090 + /opt/ai/compose/<svc>
|
||
# both deprecated 2026-05-10 — supporting services moved into services/).
|
||
CLUB3090_DIR="/opt/ai/github/club-3090"
|
||
COMPOSE_BASE="$CLUB3090_DIR/services"
|
||
# Post-PR-A (<quant>/ layer): dual composes live under <topology>/<quant>/.
|
||
# Point each var at the quant dir so `compose_at` cd's into it — mount-safe,
|
||
# the same invocation switch.sh uses (project dir = compose-file dir).
|
||
DUAL_27B_DIR="$CLUB3090_DIR/models/qwen3.6-27b/vllm/compose/dual/autoround-int4"
|
||
GEMMA_DUAL_DIR="$CLUB3090_DIR/models/gemma-4-31b/vllm/compose/dual/autoround-int4"
|
||
GEMMA_DUAL_AWQ_DIR="$CLUB3090_DIR/models/gemma-4-31b/vllm/compose/dual/awq"
|
||
# Qwen3.6-40B-Deckard: uncensored dense 40B, Q6_K GGUF + embedded MTP head,
|
||
# layer-split across both cards (llama.cpp). Dual-only — see `gpu-mode deckard`.
|
||
DECKARD_DIR="$CLUB3090_DIR/models/qwen3.6-40b-deckard/llama-cpp/compose/dual/piehsoft-q6k"
|
||
# DiffusionGemma 26B-A4B — vLLM's first dLLM (TP=2, official :gemma image + 3 fix-mounts).
|
||
# Dual-only (both cards); served via the catalog slug vllm/diffusiongemma-dual (the dgemma scene was removed). 🧪
|
||
DGEMMA_DIR="$CLUB3090_DIR/models/diffusiongemma-26b-a4b/vllm/compose/dual/fp8"
|
||
# (gemma-4-12b chat brain retired from the studio 2026-06-23 — ai-studio uses the qwen
|
||
# director only. Weights kept on disk; gemma stays a core model via `gpu-mode gemma`.)
|
||
|
||
# Estate planner state file (v0.7.0+). Instances booted via launch.sh --estate
|
||
# or --estate-file are tracked here and persist via Docker `restart:
|
||
# unless-stopped`, so they DO survive a plain mode-switch unless explicitly
|
||
# torn down via launch.sh --down-estate. mode_off uses this path to clean
|
||
# them up alongside the older vLLM/Gemma/ComfyUI services.
|
||
ESTATE_YAML="${HOME}/.club3090/estate.yml"
|
||
|
||
GREEN='\033[0;32m'
|
||
YELLOW='\033[1;33m'
|
||
RED='\033[0;31m'
|
||
CYAN='\033[0;36m'
|
||
NC='\033[0m' # No Color
|
||
|
||
# Standard supporting services living under $CLUB3090_DIR/services.
|
||
# Ollama removed 2026-06-22 (dropped from serving 2026-05-10 — Qwen/Gemma route
|
||
# through LiteLLM directly; the unused services/ollama/ compose dir was deleted).
|
||
SERVICES=(openwebui litellm qdrant searxng)
|
||
|
||
# Run a docker compose command in any directory, with optional -f override.
|
||
# Args: <dir> <action> [compose_file]
|
||
#
|
||
# Always passes --env-file $CLUB3090_DIR/.env when that file exists, so
|
||
# ${MODEL_DIR} (and other repo-level vars) resolve correctly regardless of
|
||
# which compose dir we're cd'd into. Without this, docker compose only
|
||
# auto-loads .env from the compose file's own directory and falls back to
|
||
# the relative-path default `../../../../../models-cache` (mostly empty).
|
||
#
|
||
# stderr is preserved (no 2>/dev/null) so real errors surface.
|
||
compose_at() {
|
||
local dir=$1
|
||
local action=$2
|
||
local file=${3:-docker-compose.yml}
|
||
if [ -f "$dir/$file" ]; then
|
||
local env_args=()
|
||
if [ -f "$CLUB3090_DIR/.env" ]; then
|
||
env_args=(--env-file "$CLUB3090_DIR/.env")
|
||
fi
|
||
(cd "$dir" && sudo docker compose "${env_args[@]}" -f "$file" $action)
|
||
fi
|
||
}
|
||
|
||
# Like compose_at, but injects per-invocation env assignments that survive `sudo`.
|
||
# `sudo docker compose` sanitizes the caller's environment, so vars set in the shell
|
||
# don't reach compose interpolation — pass them as leading `VAR=val` args to the
|
||
# command instead (`sudo VAR=val docker compose ...`).
|
||
# Args: <dir> <action> <file> [VAR=val ...]
|
||
compose_at_env() {
|
||
local dir=$1 action=$2 file=$3; shift 3
|
||
local envs=("$@")
|
||
if [ -f "$dir/$file" ]; then
|
||
local env_args=()
|
||
if [ -f "$CLUB3090_DIR/.env" ]; then
|
||
env_args=(--env-file "$CLUB3090_DIR/.env")
|
||
fi
|
||
(cd "$dir" && sudo "${envs[@]}" docker compose "${env_args[@]}" -f "$file" $action)
|
||
fi
|
||
}
|
||
|
||
# Standard service helpers (look in $COMPOSE_BASE/<service>)
|
||
compose_cmd() {
|
||
compose_at "$COMPOSE_BASE/$1" "$2"
|
||
}
|
||
|
||
start_service() {
|
||
printf " ${GREEN}▲${NC} Starting %-12s" "$1..."
|
||
compose_cmd "$1" "up -d" && echo "done" || echo "failed"
|
||
}
|
||
|
||
stop_service() {
|
||
printf " ${RED}▼${NC} Stopping %-12s" "$1..."
|
||
compose_cmd "$1" "down" && echo "done" || echo "skipped"
|
||
}
|
||
|
||
# Project-specific helpers
|
||
start_27b_dual_mtp() {
|
||
printf " ${GREEN}▲${NC} Starting 27b-dual-mtp..."
|
||
compose_at "$DUAL_27B_DIR" "up -d" fp8-mtp.yml && echo "done" || echo "failed"
|
||
}
|
||
stop_27b_dual_mtp() {
|
||
printf " ${RED}▼${NC} Stopping 27b-dual-mtp..."
|
||
compose_at "$DUAL_27B_DIR" "down" fp8-mtp.yml && echo "done" || echo "skipped"
|
||
}
|
||
|
||
start_27b_dual_dflash() {
|
||
printf " ${GREEN}▲${NC} Starting 27b-dual-dflash..."
|
||
compose_at "$DUAL_27B_DIR" "up -d" dflash.yml && echo "done" || echo "failed"
|
||
}
|
||
stop_27b_dual_dflash() {
|
||
printf " ${RED}▼${NC} Stopping 27b-dual-dflash..."
|
||
compose_at "$DUAL_27B_DIR" "down" dflash.yml && echo "done" || echo "skipped"
|
||
}
|
||
|
||
start_27b_dual_dflash_noviz() {
|
||
printf " ${GREEN}▲${NC} Starting 27b-dflash-noviz..."
|
||
compose_at "$DUAL_27B_DIR" "up -d" dflash-noviz.yml && echo "done" || echo "failed"
|
||
}
|
||
stop_27b_dual_dflash_noviz() {
|
||
printf " ${RED}▼${NC} Stopping 27b-dflash-noviz..."
|
||
compose_at "$DUAL_27B_DIR" "down" dflash-noviz.yml && echo "done" || echo "skipped"
|
||
}
|
||
|
||
start_27b_dual_turbo() {
|
||
printf " ${GREEN}▲${NC} Starting 27b-dual-turbo..."
|
||
compose_at "$DUAL_27B_DIR" "up -d" turbo.yml && echo "done" || echo "failed"
|
||
}
|
||
stop_27b_dual_turbo() {
|
||
printf " ${RED}▼${NC} Stopping 27b-dual-turbo..."
|
||
compose_at "$DUAL_27B_DIR" "down" turbo.yml && echo "done" || echo "skipped"
|
||
}
|
||
|
||
# Stop every 27b serving variant before starting a new one
|
||
stop_all_27b() {
|
||
stop_27b_dual_mtp
|
||
stop_27b_dual_dflash
|
||
stop_27b_dual_dflash_noviz
|
||
stop_27b_dual_turbo
|
||
}
|
||
|
||
# --- ComfyUI (image / video generation) -------------------------------------
|
||
# GPU-bound — mutex with all vLLM / SGLang / llama-server LLM serving.
|
||
start_comfyui() {
|
||
printf " ${GREEN}▲${NC} Starting comfyui..."
|
||
compose_at "$COMPOSE_BASE/comfyui" "up -d" && echo "done" || echo "failed"
|
||
}
|
||
stop_comfyui() {
|
||
printf " ${RED}▼${NC} Stopping comfyui..."
|
||
compose_at "$COMPOSE_BASE/comfyui" "down" && echo "done" || echo "skipped"
|
||
}
|
||
|
||
# Video-studio sidecars: the always-on media gallery (:8189) and the prompt
|
||
# "director" LLM (:8090, GPU0). See services/studio/ + docs/ai-studio/video.md.
|
||
start_studio_gallery() {
|
||
printf " ${GREEN}▲${NC} Starting studio-gallery (:8189)..."
|
||
compose_at "$COMPOSE_BASE/studio/gallery" "up -d" && echo "done" || echo "failed"
|
||
}
|
||
start_studio_director() {
|
||
printf " ${GREEN}▲${NC} Starting studio-director (:8090, GPU0)..."
|
||
compose_at "$COMPOSE_BASE/studio/enhancer" "up -d" && echo "done" || echo "failed"
|
||
}
|
||
start_studio_orchestrator() {
|
||
printf " ${GREEN}▲${NC} Starting studio-orchestrator (:8190, long-clip chaining)..."
|
||
compose_at "$COMPOSE_BASE/studio/orchestrator" "up -d --build" && echo "done" || echo "failed"
|
||
}
|
||
# Native-button image shim (:8191): ComfyUI reverse-proxy that crafts Ideogram-4 JSON
|
||
# captions (via the director) so OWUI's native 🖼️ image button renders instead of hitting
|
||
# the "blocked by safety filter" placeholder. Needs the director (:8090) up.
|
||
start_studio_image_shim() {
|
||
printf " ${GREEN}▲${NC} Starting studio-image-shim (:8191, native-button Ideogram captions)..."
|
||
compose_at "$COMPOSE_BASE/studio/image-shim" "up -d --build" && echo "done" || echo "failed"
|
||
}
|
||
# Integrated-voices TTS + audio mixdown (:8192, Kokoro on CPU). The pipe POSTs /narrate after a
|
||
# video render to mix a voiceover over the clip's native audio (ducked + normalized). No GPU.
|
||
start_studio_tts() {
|
||
printf " ${GREEN}▲${NC} Starting studio-tts (:8192, Kokoro voices + mixdown, CPU)..."
|
||
compose_at "$COMPOSE_BASE/studio/tts" "up -d --build" && echo "done" || echo "failed"
|
||
}
|
||
|
||
# (start_comfyui_gpu0 + the gemma-4-12b chat brain were retired with the
|
||
# image-studio→ai-studio consolidation 2026-06-23 — ComfyUI now always spans both GPUs.)
|
||
|
||
# --- Gemma 4 31B dual-card serving variants ---------------------------------
|
||
# (start_gemma_mtp + the gemma-mtp/gemma-int8 scenes were removed — only 'gemma'
|
||
# (INT8 PTH KV, 262K, :8032) remains. stop_gemma_mtp is KEPT: stop_all_gemma
|
||
# (→ mode_off) defensively tears down a stray bf16 gemma container.)
|
||
stop_gemma_mtp() {
|
||
printf " ${RED}▼${NC} Stopping gemma-mtp..."
|
||
compose_at "$GEMMA_DUAL_DIR" "down" bf16-mtp.yml && echo "done" || echo "skipped"
|
||
}
|
||
|
||
start_gemma_int8() {
|
||
printf " ${GREEN}▲${NC} Starting gemma-int8..."
|
||
compose_at "$GEMMA_DUAL_DIR" "up -d" int8.yml && echo "done" || echo "failed"
|
||
}
|
||
stop_gemma_int8() {
|
||
printf " ${RED}▼${NC} Stopping gemma-int8..."
|
||
compose_at "$GEMMA_DUAL_DIR" "down" int8.yml && echo "done" || echo "skipped"
|
||
}
|
||
|
||
# Stop every Gemma serving variant before starting a new one
|
||
stop_all_gemma() {
|
||
stop_gemma_mtp
|
||
stop_gemma_int8
|
||
}
|
||
|
||
start_deckard() {
|
||
printf " ${GREEN}▲${NC} Starting deckard-40b..."
|
||
compose_at "$DECKARD_DIR" "up -d" mtp.yml && echo "done" || echo "failed"
|
||
}
|
||
stop_deckard() {
|
||
printf " ${RED}▼${NC} Stopping deckard-40b..."
|
||
compose_at "$DECKARD_DIR" "down" mtp.yml && echo "done" || echo "skipped"
|
||
}
|
||
# (start_diffusiongemma + the diffusiongemma SCENE were removed — the model stays
|
||
# serveable via the catalog slug vllm/diffusiongemma-dual. stop_diffusiongemma is
|
||
# KEPT: mode_off + the studio modes call it to defensively clear a stray dgemma
|
||
# container off the cards.)
|
||
stop_diffusiongemma() {
|
||
printf " ${RED}▼${NC} Stopping diffusiongemma-26b-a4b..."
|
||
compose_at_env "$DGEMMA_DIR" "down" base.yml PORT=8199 && echo "done" || echo "skipped"
|
||
}
|
||
|
||
show_status() {
|
||
echo ""
|
||
echo -e "${CYAN}═══ Service Status ═══${NC}"
|
||
sudo docker ps --format "table {{.Names}}\t{{.Status}}\t{{.Ports}}" 2>/dev/null
|
||
echo ""
|
||
echo -e "${CYAN}═══ Active Model(s) ═══${NC}"
|
||
# Check ports in priority order: 8010, 8012, 8020, 11434, 4000
|
||
if curl -sf -m 2 http://localhost:8010/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 http://localhost:8010/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} 27b-dual-mtp @ :8010 → ${m:-unknown} (MTP n=3 + fp8 + 262K + vision)"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8012/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 http://localhost:8012/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} 27b-dflash @ :8012 → ${m:-unknown} (DFlash N=5 + 185K + vision)"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8013/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 http://localhost:8013/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} 27b-dflash-noviz @ :8013 → ${m:-unknown} (DFlash N=5 + 200K, no vision)"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8011/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 http://localhost:8011/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} 27b-turbo @ :8011 → ${m:-unknown} (TurboQuant_3bit_nc + MTP n=3 + v7.14, 4-stream concurrency)"
|
||
fi
|
||
# :8020 = llama.cpp single-card. llamacpp/default + llamacpp/mtp share the
|
||
# base container llama-cpp-qwen36-27b (same compose, collapsed 2026-05-22);
|
||
# llamacpp/mtp-vision now defaults to llama-cpp-qwen36-27b-vision (#169).
|
||
# All still match the llama-cpp-* prefix used for detection below.
|
||
if curl -sf -m 2 http://localhost:8020/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 http://localhost:8020/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} llamacpp/single @ :8020 → ${m:-unknown} (llama.cpp single-card)"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8030/v1/models >/dev/null 2>&1; then
|
||
local m container engine_tag
|
||
m=$(curl -sf -m 2 http://localhost:8030/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
# Detect engine via container name on the port (was hardcoded to "gemma-mtp"
|
||
# / Gemma description; post-v0.8.3, llamacpp/mtp-vision also lands on :8030).
|
||
container=$(sudo docker ps --format '{{.Names}} {{.Ports}}' 2>/dev/null | awk '/:8030->/ {print $1; exit}')
|
||
if [[ "$container" == llama-cpp-* ]]; then
|
||
engine_tag="llamacpp/mtp-vision @ :8030 → ${m:-unknown} (Q4_K_M + MTP + vision, 49K)"
|
||
else
|
||
engine_tag="gemma-mtp @ :8030 → ${m:-unknown} (Gemma 4 31B + MTP n=3 + bf16 KV + 32K)"
|
||
fi
|
||
echo -e " ${GREEN}▶${NC} $engine_tag"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8032/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 http://localhost:8032/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} gemma-int8 @ :8032 → ${m:-unknown} (INT8 PTH KV)"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8090/v1/models >/dev/null 2>&1; then
|
||
echo -e " ${GREEN}▶${NC} studio director @ :8090 → qwen3.5-4b-uncensored (prompt crafter, GPU0, llama.cpp)"
|
||
fi
|
||
if curl -sf -m 2 http://localhost:8188/ >/dev/null 2>&1; then
|
||
echo -e " ${GREEN}▶${NC} ComfyUI @ :8188 → image/video generation (GPU-bound, mutex with LLM)"
|
||
fi
|
||
if curl -sf -m 2 -H "Authorization: Bearer sk-litellm-master-key" http://localhost:4000/v1/models >/dev/null 2>&1; then
|
||
local m
|
||
m=$(curl -sf -m 2 -H "Authorization: Bearer sk-litellm-master-key" http://localhost:4000/v1/models | python3 -c "import sys,json;d=json.load(sys.stdin);print(', '.join(x['id'] for x in d.get('data',[])))" 2>/dev/null)
|
||
echo -e " ${GREEN}▶${NC} LiteLLM @ :4000 → ${m:-unknown}"
|
||
fi
|
||
if ! curl -sf -m 2 http://localhost:8010/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:8011/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:8012/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:8013/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:8030/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:8032/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:8033/v1/models >/dev/null 2>&1 \
|
||
&& ! curl -sf -m 2 http://localhost:11434/api/tags >/dev/null 2>&1; then
|
||
echo -e " ${YELLOW}(no inference endpoint responding)${NC}"
|
||
fi
|
||
echo ""
|
||
echo -e "${CYAN}═══ GPU Status ═══${NC}"
|
||
nvidia-smi --query-gpu=index,memory.used,memory.total,memory.free,utilization.gpu --format=csv,noheader 2>/dev/null || echo "nvidia-smi not available"
|
||
# One-line power-cap state: enforced vs default per card (see 'gpu-mode power-cap').
|
||
nvidia-smi --query-gpu=index,power.limit,power.default_limit --format=csv,noheader,nounits 2>/dev/null \
|
||
| while IFS=',' read -r gi lim def; do
|
||
gi="${gi// /}"; lim="${lim// /}"; def="${def// /}"
|
||
if awk "BEGIN{exit !($lim < $def)}"; then
|
||
echo -e " power cap: GPU ${gi} ${YELLOW}${lim}W${NC} (capped; default ${def}W)"
|
||
else
|
||
echo -e " power cap: GPU ${gi} ${GREEN}${lim}W${NC} (uncapped; default ${def}W)"
|
||
fi
|
||
done || true
|
||
echo ""
|
||
echo -e "${CYAN}═══ RAM Status ═══${NC}"
|
||
free -h | head -2
|
||
echo ""
|
||
echo -e "${CYAN}═══ Disk Status ═══${NC}"
|
||
df -h / /mnt/models 2>/dev/null | tail -2
|
||
echo ""
|
||
echo -e "${CYAN}═══ Docker Disk ═══${NC}"
|
||
sudo docker system df 2>/dev/null | head -5 || echo "(docker not running)"
|
||
local docker_dir_size tmp_size
|
||
docker_dir_size=$(sudo du -sh /var/lib/docker 2>/dev/null | cut -f1)
|
||
tmp_size=$(sudo du -sh /tmp 2>/dev/null | cut -f1)
|
||
echo ""
|
||
echo " /var/lib/docker (on /): ${docker_dir_size:-?}"
|
||
echo " /tmp (on /): ${tmp_size:-?}"
|
||
echo ""
|
||
}
|
||
|
||
mode_prune() {
|
||
echo -e "${CYAN}═══ Docker prune (safe) ═══${NC}"
|
||
echo "Removes images not referenced by any container (running OR stopped)."
|
||
echo "Does NOT touch build cache or volumes — use 'prune-all' for those."
|
||
echo ""
|
||
echo "${CYAN}── Before ──${NC}"
|
||
sudo docker system df 2>/dev/null | head -5
|
||
echo ""
|
||
sudo docker image prune -a -f 2>&1 | tail -10
|
||
echo ""
|
||
echo "${CYAN}── After ──${NC}"
|
||
sudo docker system df 2>/dev/null | head -5
|
||
}
|
||
|
||
mode_prune_all() {
|
||
echo -e "${CYAN}═══ Docker prune (aggressive) ═══${NC}"
|
||
echo "Removes:"
|
||
echo " - images not referenced by any container"
|
||
echo " - all build cache (kept ≤5 GB)"
|
||
echo " - dangling networks"
|
||
echo "Does NOT remove volumes (qdrant-data, openwebui-data are safe)."
|
||
echo ""
|
||
echo "${CYAN}── Before ──${NC}"
|
||
sudo docker system df 2>/dev/null | head -5
|
||
echo ""
|
||
echo "${YELLOW}Pruning images...${NC}"
|
||
sudo docker image prune -a -f 2>&1 | tail -3
|
||
echo ""
|
||
echo "${YELLOW}Pruning networks...${NC}"
|
||
sudo docker network prune -f 2>&1 | tail -3
|
||
echo ""
|
||
echo "${YELLOW}Pruning build cache (keeping 5 GB)...${NC}"
|
||
sudo docker buildx prune -f --keep-storage 5GB 2>&1 | tail -3
|
||
echo ""
|
||
echo "${CYAN}── After ──${NC}"
|
||
sudo docker system df 2>/dev/null | head -5
|
||
}
|
||
|
||
mode_chat() {
|
||
echo -e "${CYAN}═══ Switching to CHAT mode ═══${NC}"
|
||
echo "Starting: Open WebUI, LiteLLM, Qdrant, SearXNG"
|
||
echo "Stopping: all GPU-served model containers (Qwen + Gemma)"
|
||
echo ""
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_all_gemma
|
||
stop_comfyui
|
||
start_service openwebui
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}Chat mode active.${NC} Open WebUI: http://192.168.86.33:8080"
|
||
}
|
||
|
||
mode_27b() {
|
||
echo -e "${CYAN}═══ Switching to 27B dual-card MTP mode (default) ═══${NC}"
|
||
echo "Starting: Qwen3.6-27B MTP n=3 + fp8 KV + 262K + vision + 2 streams (TP=2)"
|
||
echo "Port: 8010 | Container: vllm-qwen36-27b-dual"
|
||
echo "Stopping: other 27B variants"
|
||
echo ""
|
||
stop_all_gemma
|
||
stop_comfyui
|
||
stop_27b_dual_dflash
|
||
stop_27b_dual_dflash_noviz
|
||
stop_27b_dual_turbo
|
||
start_27b_dual_mtp
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service openwebui
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}27B dual-card MTP mode active.${NC} API: http://192.168.86.33:8010"
|
||
echo -e "${YELLOW}Per-stream: 68 narr / 89 code TPS short, 36 TPS @ 100K, 28 TPS @ 200K warm.${NC}"
|
||
echo -e "${YELLOW}2 concurrent streams. KV pool 168K, max concurrency 2.36× at full 262K.${NC}"
|
||
echo -e "${YELLOW}Vision + tools + thinking + 262K ctx all working. Boot ~3-4 min.${NC}"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f vllm-qwen36-27b-dual${NC}"
|
||
}
|
||
|
||
# (mode_gemma — the bf16 'gemma-mtp' fallback scene — was removed; only the INT8
|
||
# 'gemma' scene (mode_gemma_int8) remains. The bf16 compose is still serveable
|
||
# via its catalog slug if needed.)
|
||
|
||
mode_gemma_dflash() {
|
||
echo -e "${CYAN}═══ Switching to Gemma 4 31B DFlash mode ═══${NC}"
|
||
echo "Starting: Gemma 4 31B + z-lab DFlash drafter (TP=2, :8032)"
|
||
echo ""
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_gemma_mtp
|
||
stop_gemma_int8
|
||
stop_gemma_dflash_int8
|
||
stop_gemma_awq
|
||
start_gemma_dflash
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service openwebui
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}Gemma 4 31B DFlash mode active.${NC} API: http://192.168.86.33:8032"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f vllm-gemma-4-31b-dflash${NC}"
|
||
}
|
||
|
||
mode_gemma_int8() {
|
||
echo -e "${CYAN}═══ Switching to Gemma 4 31B INT8-PTH mode (dual default, long ctx) ═══${NC}"
|
||
echo "Starting: Gemma 4 31B + INT8 PTH KV + 262K ctx (TP=2, :8032)"
|
||
echo ""
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_gemma_mtp
|
||
stop_gemma_dflash
|
||
stop_gemma_dflash_int8
|
||
stop_gemma_awq
|
||
start_gemma_int8
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service openwebui
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}Gemma 4 31B INT8 PTH mode active.${NC} API: http://192.168.86.33:8032"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f vllm-gemma-4-31b-mtp-int8${NC}"
|
||
}
|
||
mode_deckard() {
|
||
echo -e "${CYAN}═══ Switching to DECKARD-40B mode (uncensored, dual-card) ═══${NC}"
|
||
echo "Starting: Qwen3.6-40B-Deckard Q6_K + MTP n=2 + q8_0 KV + 128K ctx (llama.cpp, :8199)"
|
||
echo "Stopping: all other GPU models (Deckard layer-splits across both cards)"
|
||
echo ""
|
||
stop_all_27b
|
||
stop_all_gemma
|
||
stop_comfyui
|
||
start_deckard
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service openwebui
|
||
start_service searxng
|
||
# Deckard isn't in the LiteLLM gateway config, so wire it into Open WebUI
|
||
# directly as an OpenAI connection (reuses switch.sh --owui's helper).
|
||
# Best-effort: a no-op if OWUI isn't running / not ready yet.
|
||
if [ -x "$CLUB3090_DIR/scripts/lib/owui-register.sh" ]; then
|
||
bash "$CLUB3090_DIR/scripts/lib/owui-register.sh" 8199 || true
|
||
fi
|
||
echo ""
|
||
echo -e "${GREEN}Deckard-40B mode active.${NC} API: http://192.168.86.33:8199 (model: deckard-40b)"
|
||
echo -e "${YELLOW}MTP n=2: ~36 narr / 46 code TPS · 128K ctx @ q8_0 KV · uncensored, text-only.${NC}"
|
||
echo -e "${YELLOW}First boot ~1-2 min (31 GB GGUF load + 128K KV alloc across both cards).${NC}"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f llama-cpp-deckard-40b${NC}"
|
||
}
|
||
|
||
mode_gemma_dflash_int8() {
|
||
echo -e "${CYAN}═══ Switching to Gemma 4 31B DFlash + INT8 PTH mode ═══${NC}"
|
||
echo "Starting: Gemma 4 31B + DFlash + INT8 PTH KV (TP=2, :8032). Requires vllm#42102."
|
||
echo ""
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_gemma_mtp
|
||
stop_gemma_dflash
|
||
stop_gemma_int8
|
||
stop_gemma_awq
|
||
start_gemma_dflash_int8
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service openwebui
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}Gemma 4 31B DFlash + INT8 mode active.${NC} API: http://192.168.86.33:8032"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f vllm-gemma-4-31b-dflash-int8${NC}"
|
||
}
|
||
|
||
mode_gemma_awq() {
|
||
echo -e "${CYAN}═══ Switching to Gemma 4 31B AWQ-4bit mode ═══${NC}"
|
||
echo "Starting: Gemma 4 31B AWQ-4bit (TP=2, :8033)"
|
||
echo ""
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_gemma_mtp
|
||
stop_gemma_dflash
|
||
stop_gemma_int8
|
||
stop_gemma_dflash_int8
|
||
start_gemma_awq
|
||
start_service litellm
|
||
start_service qdrant
|
||
start_service openwebui
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}Gemma 4 31B AWQ mode active.${NC} API: http://192.168.86.33:8033"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f vllm-gemma-4-31b-awq${NC}"
|
||
}
|
||
|
||
# (mode_diffusiongemma removed — DiffusionGemma is a niche dLLM; its gpu-mode scene
|
||
# was redundant with the catalog slug vllm/diffusiongemma-dual, which is the way
|
||
# to serve it. 'off' still tears down any running dgemma container.)
|
||
|
||
# Verify a studio scene's MODELS are on disk BEFORE starting its containers —
|
||
# else they boot with no weights (silent failure: the director llama.cpp has no
|
||
# GGUF, ComfyUI has no checkpoint). gpu-mode only STARTS the bundle; the models
|
||
# are fetched by setup-image-studio.sh. This mirrors preflight_compose_deps for
|
||
# the LLM composes — it CHECKS + points at the installer, never auto-downloads.
|
||
# Both the CLI and the cockpit reach the studio scenes through here, so one gate
|
||
# protects both. Skip: STUDIO_NO_PREFLIGHT=1.
|
||
preflight_studio_models() {
|
||
if [ "${STUDIO_NO_PREFLIGHT:-0}" = "1" ]; then return 0; fi
|
||
# Drive the check off the SHARED manifest (scripts/lib/studio-models.tsv) — the SAME
|
||
# file c3 reads — so "what ai-studio needs" is defined once (director · image · video
|
||
# · audio) and the two surfaces can't drift. director = HARD requirement (no lane can
|
||
# craft without it); every other lane's models missing is a WARNING (studio still boots).
|
||
local manifest="$CLUB3090_DIR/scripts/lib/studio-models.tsv"
|
||
if [ ! -f "$manifest" ]; then
|
||
echo -e "${YELLOW}[preflight] studio manifest missing ($manifest) — skipping model check.${NC}" >&2
|
||
return 0
|
||
fi
|
||
# Roots: weights (director, MODEL_DIR from .env) · comfy (image/video/audio tree).
|
||
local model_dir comfy_models
|
||
model_dir="$(grep -E '^MODEL_DIR=' "$CLUB3090_DIR/.env" 2>/dev/null | tail -1 | cut -d= -f2-)"
|
||
model_dir="${model_dir:-/mnt/models/huggingface}"
|
||
comfy_models="${COMFYUI_MODELS_DIR:-/mnt/models/comfyui/models}"
|
||
local director_missing=0 warns=() modality label root rel size installer base
|
||
while IFS=$'\t' read -r modality label root rel size installer; do
|
||
case "$modality" in ''|\#*) continue ;; esac # skip blank / comment rows
|
||
[ -n "$rel" ] || continue
|
||
if [ "$root" = "weights" ]; then base="$model_dir"; else base="$comfy_models"; fi
|
||
[ -e "$base/$rel" ] && continue # model present
|
||
if [ "$modality" = "director" ]; then
|
||
director_missing=1
|
||
echo -e "${RED}[preflight] ai-studio director GGUF not on disk — no lane can craft a prompt:${NC}" >&2
|
||
echo " - $label → $base/$rel" >&2
|
||
echo -e "${YELLOW}[preflight] Fix: bash $installer${NC}" >&2
|
||
else
|
||
warns+=("$modality: $label → bash $installer")
|
||
fi
|
||
done < "$manifest"
|
||
if [ "$director_missing" = "1" ]; then
|
||
echo -e "${YELLOW}[preflight] Skip: STUDIO_NO_PREFLIGHT=1 gpu-mode ai-studio${NC}" >&2
|
||
return 1
|
||
fi
|
||
local w
|
||
for w in "${warns[@]}"; do
|
||
echo -e "${YELLOW}[preflight] lane models absent — $w${NC}" >&2
|
||
done
|
||
return 0
|
||
}
|
||
|
||
mode_ai_studio() {
|
||
preflight_studio_models ai-studio || return 1
|
||
echo -e "${CYAN}═══ Switching to AI-STUDIO mode (image · video · audio · voice) ═══${NC}"
|
||
echo "Starting: ComfyUI :8188 (both GPUs) + director :8090 + gallery/orchestrator/image-shim/tts + Open WebUI"
|
||
echo "Stopping: all GPU-bound LLM serving (Qwen + Gemma + DiffusionGemma)"
|
||
echo ""
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_all_gemma
|
||
stop_diffusiongemma
|
||
start_comfyui
|
||
start_studio_director
|
||
start_studio_gallery
|
||
start_studio_orchestrator
|
||
start_studio_image_shim
|
||
start_studio_tts
|
||
start_service openwebui
|
||
start_service litellm
|
||
start_service searxng
|
||
echo ""
|
||
echo -e "${GREEN}AI-studio mode active.${NC} — one scene; pick the lane in Open WebUI."
|
||
echo -e " Open WebUI: http://192.168.86.33:8080 (image · video · music · SFX · voice lanes)"
|
||
echo -e " Gallery: http://192.168.86.33:8189 (all generated media; survives ComfyUI down)"
|
||
echo -e " ComfyUI: http://192.168.86.33:8188 (full node graph / control)"
|
||
echo -e "${YELLOW}First ComfyUI boot can take a few min (clones + node deps). Video DiT splits across both 3090s (DisTorch); image/audio lanes run on GPU0 beside the director.${NC}"
|
||
echo -e "${YELLOW}GPU-mutex with the dual-card LLMs. Premium voice (step-audio-editx) is on-demand on GPU1 — mutually exclusive with an active video render.${NC}"
|
||
echo -e "${YELLOW}Tail: sudo docker logs -f comfyui${NC}"
|
||
}
|
||
|
||
# ('bigmodel' was removed — it was ~redundant with 'off' (both stop everything).
|
||
# Its only useful extra, the freed VRAM/RAM readout, is folded into mode_off
|
||
# below; the cache-drop was marginal (Linux auto-reclaims page cache, and it does
|
||
# nothing for VRAM) and the llama-server hint lives in the docs.)
|
||
|
||
stop_estate() {
|
||
# Tear down any estate-managed instances (launch.sh --estate-file or --estate
|
||
# bookings persist via Docker `restart: unless-stopped`). No-op if no estate
|
||
# plan exists or launch.sh is unavailable.
|
||
if [[ ! -f "$ESTATE_YAML" ]]; then
|
||
return 0
|
||
fi
|
||
if ! command -v bash >/dev/null 2>&1 || [[ ! -x "$CLUB3090_DIR/scripts/launch.sh" ]]; then
|
||
return 0
|
||
fi
|
||
if ! python3 -c "import yaml; d=yaml.safe_load(open('$ESTATE_YAML')); raise SystemExit(0 if d and d.get('estate') else 1)" 2>/dev/null; then
|
||
return 0 # empty/missing estate list
|
||
fi
|
||
printf " ${RED}▼${NC} Stopping estate-managed instances..."
|
||
if bash "$CLUB3090_DIR/scripts/launch.sh" --down-estate "$ESTATE_YAML" >/dev/null 2>&1; then
|
||
echo "done"
|
||
else
|
||
echo "skipped (no instances or already down)"
|
||
fi
|
||
}
|
||
|
||
# --- GPU power-cap controls -------------------------------------------------
|
||
# The rig normally runs both 3090s capped at 230W (quieter / cooler — see the
|
||
# systemd unit below). The cap suppresses benchmark TPS, so maintainers need a
|
||
# quick way to uncap to the hardware default for a true-TPS bench, then re-cap.
|
||
#
|
||
# `nvidia-power-cap.service` is the single source of truth for the 230W value
|
||
# AND re-applies it on every boot (Type=oneshot, RemainAfterExit=yes, enabled).
|
||
# So `power-cap on` *restarts* that unit — `restart` (not `start`) is required:
|
||
# the unit is already `active` from boot, and `systemctl start` on an
|
||
# already-active RemainAfterExit oneshot is a no-op (it won't re-run ExecStart,
|
||
# so the cap wouldn't actually re-apply after a `power-cap off`). `restart`
|
||
# stops it (clearing RemainAfterExit) then re-runs both `-pl 230` ExecStart
|
||
# lines. `power-cap off` reads each card's Default Power Limit from nvidia-smi
|
||
# (370W on GPU 0, 420W on GPU 1 here — they differ, so we never hardcode) and
|
||
# applies it. `off` is session-scoped: a reboot OR a driver reload re-applies
|
||
# 230W via the service. We never disable the service.
|
||
POWER_CAP_SERVICE="nvidia-power-cap.service"
|
||
|
||
# Print per-GPU enforced / default / min / max power limits (one row per card).
|
||
powercap_show() {
|
||
echo -e "${CYAN}═══ GPU Power Limits ═══${NC}"
|
||
if ! command -v nvidia-smi >/dev/null 2>&1; then
|
||
echo -e " ${RED}✗ nvidia-smi not found${NC} — cannot read power limits."
|
||
return 1
|
||
fi
|
||
if ! nvidia-smi \
|
||
--query-gpu=index,power.limit,power.default_limit,power.min_limit,power.max_limit \
|
||
--format=csv 2>/dev/null; then
|
||
echo -e " ${RED}✗ nvidia-smi query failed${NC} — driver loaded?"
|
||
return 1
|
||
fi
|
||
}
|
||
|
||
# Echo the current enforced limit per GPU (used after on/off to confirm effect).
|
||
powercap_echo_enforced() {
|
||
local line
|
||
while IFS= read -r line; do
|
||
echo -e " ${GREEN}▶${NC} GPU ${line%%,*} enforced limit:${line#*,} W"
|
||
done < <(nvidia-smi --query-gpu=index,power.limit --format=csv,noheader,nounits 2>/dev/null)
|
||
}
|
||
|
||
mode_powercap() {
|
||
local action="${1:-status}"
|
||
if ! command -v nvidia-smi >/dev/null 2>&1; then
|
||
echo -e "${RED}✗ nvidia-smi not found.${NC} Install the NVIDIA driver / utils first." >&2
|
||
exit 1
|
||
fi
|
||
# A numeric action = an explicit CUSTOM wattage applied to both cards (the
|
||
# serve-cockpit power-cap menu's "custom" option). Validated against each
|
||
# card's [min,max] range; session-scoped like `off` (the boot service still
|
||
# re-applies 230W on reboot/reload).
|
||
if [[ "$action" =~ ^[0-9]+$ ]]; then
|
||
echo -e "${CYAN}═══ Setting custom GPU power cap (${action}W) ═══${NC}"
|
||
local cidx cmin cmax crc=0 capplied=0
|
||
while IFS=',' read -r cidx cmin cmax; do
|
||
cidx="${cidx// /}"; cmin="${cmin%%.*}"; cmin="${cmin// /}"
|
||
cmax="${cmax%%.*}"; cmax="${cmax// /}"
|
||
[ -z "$cidx" ] && continue
|
||
if [ -n "$cmin" ] && [ -n "$cmax" ] && { [ "$action" -lt "$cmin" ] || [ "$action" -gt "$cmax" ]; }; then
|
||
echo -e " ${RED}✗ GPU ${cidx}: ${action}W out of range [${cmin},${cmax}]W${NC}" >&2
|
||
crc=1; continue
|
||
fi
|
||
echo " Setting GPU ${cidx} → ${action} W..."
|
||
if ! sudo nvidia-smi -i "$cidx" -pl "$action" >/dev/null 2>&1; then
|
||
echo -e " ${RED}✗ Failed to set GPU ${cidx} to ${action} W${NC} (sudo? driver?)." >&2
|
||
crc=1
|
||
else
|
||
capplied=1
|
||
fi
|
||
done < <(nvidia-smi --query-gpu=index,power.min_limit,power.max_limit \
|
||
--format=csv,noheader,nounits 2>/dev/null)
|
||
if [ "$capplied" -eq 0 ]; then
|
||
echo -e "${RED}✗ No GPUs updated.${NC} Check the value is within range: nvidia-smi -q -d POWER" >&2
|
||
exit 1
|
||
fi
|
||
echo -e "${GREEN}Custom cap ${action}W applied.${NC} ${YELLOW}Session-scoped — a reboot or driver"
|
||
echo -e "reload re-applies 230W via ${POWER_CAP_SERVICE}.${NC}"
|
||
powercap_echo_enforced
|
||
[ "$crc" -eq 0 ] || exit 1
|
||
return
|
||
fi
|
||
case "$action" in
|
||
on)
|
||
echo -e "${CYAN}═══ Re-applying GPU power cap (230W) ═══${NC}"
|
||
echo "Restarting ${POWER_CAP_SERVICE} (the boot-time 230W enforcer)."
|
||
# restart, not start — the unit is already active from boot, so
|
||
# `start` is a no-op on a RemainAfterExit oneshot (won't re-run -pl).
|
||
if sudo systemctl restart "$POWER_CAP_SERVICE" 2>/dev/null; then
|
||
echo -e "${GREEN}Power cap re-applied via systemd.${NC}"
|
||
else
|
||
# Fallback: service missing/disabled — apply 230W directly.
|
||
echo -e "${YELLOW}systemctl restart failed; falling back to direct nvidia-smi -pl 230.${NC}" >&2
|
||
if ! { sudo nvidia-smi -i 0 -pl 230 && sudo nvidia-smi -i 1 -pl 230; }; then
|
||
echo -e "${RED}✗ Failed to set 230W cap.${NC} Check sudo + driver state with: nvidia-smi -q -d POWER" >&2
|
||
exit 1
|
||
fi
|
||
fi
|
||
powercap_echo_enforced
|
||
;;
|
||
off)
|
||
echo -e "${CYAN}═══ Uncapping GPUs to hardware default ═══${NC}"
|
||
# Read each card's Default Power Limit — they can differ (370 vs 420
|
||
# here), and nvidia-smi has no "reset" flag, so we pass the value.
|
||
local idx def rc=0 applied=0
|
||
while IFS=',' read -r idx def; do
|
||
idx="${idx// /}"
|
||
def="${def// /}"
|
||
[ -z "$idx" ] && continue
|
||
echo " Setting GPU ${idx} → ${def} W (default)..."
|
||
if ! sudo nvidia-smi -i "$idx" -pl "$def" >/dev/null 2>&1; then
|
||
echo -e " ${RED}✗ Failed to set GPU ${idx} to ${def} W${NC} (sudo? driver?)." >&2
|
||
rc=1
|
||
else
|
||
applied=1
|
||
fi
|
||
done < <(nvidia-smi --query-gpu=index,power.default_limit \
|
||
--format=csv,noheader,nounits 2>/dev/null)
|
||
if [ "$applied" -eq 0 ]; then
|
||
echo -e "${RED}✗ No GPUs updated.${NC} Check: nvidia-smi -q -d POWER" >&2
|
||
exit 1
|
||
fi
|
||
echo -e "${GREEN}Uncapped to default.${NC} ${YELLOW}Session-scoped — a reboot or driver"
|
||
echo -e "reload re-applies 230W via ${POWER_CAP_SERVICE}. Run 'gpu-mode power-cap on' to re-cap now.${NC}"
|
||
powercap_echo_enforced
|
||
[ "$rc" -eq 0 ] || exit 1
|
||
;;
|
||
status)
|
||
powercap_show
|
||
;;
|
||
*)
|
||
echo -e "${RED}Unknown power-cap action:${NC} $action" >&2
|
||
echo "Usage: gpu-mode power-cap <on|off|status|WATTS>" >&2
|
||
exit 1
|
||
;;
|
||
esac
|
||
}
|
||
|
||
mode_off() {
|
||
echo -e "${CYAN}═══ Stopping ALL services ═══${NC}"
|
||
stop_all_27b
|
||
stop_deckard
|
||
stop_diffusiongemma
|
||
stop_all_gemma
|
||
stop_comfyui
|
||
stop_estate
|
||
for svc in "${SERVICES[@]}"; do
|
||
stop_service "$svc"
|
||
done
|
||
echo ""
|
||
echo -e "${GREEN}All services stopped.${NC}"
|
||
# Freed-resource readout (folded in from the removed 'bigmodel' scene) — handy
|
||
# before a one-off manual workload (a custom llama-server GGUF, an experiment).
|
||
echo ""
|
||
echo -e "${CYAN}═══ Free resources ═══${NC}"
|
||
echo -e "VRAM:"
|
||
nvidia-smi --query-gpu=memory.free,memory.total --format=csv,noheader 2>/dev/null
|
||
echo -e "RAM:"
|
||
free -h | grep Mem | awk '{print " Free: "$4" / Total: "$2}'
|
||
}
|
||
|
||
# --- Scene catalog (`--list-modes [--json]`) --------------------------------
|
||
# Machine-readable mirror of the dispatch case + what each mode_* function
|
||
# actually starts. Each row: name (the dispatch keyword), group, description,
|
||
# services (logical service/container names the mode brings up), ports, gpus.
|
||
# Groups follow the contract: serving / studio / ops. This is hand-maintained
|
||
# alongside the dispatch case below — keep them in lockstep when adding a mode.
|
||
#
|
||
# Format is a TSV heredoc (name \t group \t description \t services \t ports \t
|
||
# gpus) so the data lives in one readable place; we render it to a plain table
|
||
# or to JSON via python3 (already a hard dep of show_status). The plain render
|
||
# is the default; --json emits the [{name,group,description,services,ports,gpus}]
|
||
# array the contract specifies. services/ports are comma-joined in the TSV and
|
||
# split into JSON arrays; gpus is the human GPU-usage note.
|
||
list_modes_data() {
|
||
# name<TAB>group<TAB>description<TAB>services<TAB>ports<TAB>gpus
|
||
cat <<'TSV'
|
||
chat ops Open WebUI + LiteLLM + Qdrant + SearXNG (browser chat, no GPU model) openwebui,litellm,qdrant,searxng 8080,4000 none
|
||
27b models Qwen3.6-27B MTP n=3 + fp8 KV + 262K + vision (TP=2) — default vllm-qwen36-27b-dual,litellm,qdrant,openwebui,searxng 8010,8080,4000 both
|
||
gemma models Gemma 4 31B INT8 PTH KV + 262K + vision (TP=2) — dual default vllm-gemma-4-31b-mtp-int8,litellm,qdrant,openwebui,searxng 8032,8080,4000 both
|
||
deckard models Qwen3.6-40B-Deckard Q6_K + MTP n=2 + q8_0 KV + 128K (llama.cpp, dual) llama-cpp-deckard-40b,litellm,qdrant,openwebui,searxng 8199,8080,4000 both
|
||
ai-studio studio image · video · audio · voice — ComfyUI both GPUs + qwen director + sidecars + Open WebUI (pick the lane in OWUI) comfyui,studio-director,studio-gallery,studio-orchestrator,studio-image-shim,studio-tts,openwebui,litellm,searxng 8188,8090,8189,8190,8191,8192,8080,4000 both
|
||
off ops Stop all services all-stopped none
|
||
power-cap ops GPU power-cap controls (on/off/status; both 3090s, 230W default cap) both
|
||
prune ops docker image prune -a (safe — only unreferenced images) none
|
||
prune-all ops + build cache (keep 5 GB) + dangling networks (volumes safe) none
|
||
TSV
|
||
}
|
||
|
||
list_modes() {
|
||
local as_json=0
|
||
[[ "${1:-}" == "--json" ]] && as_json=1
|
||
if [[ "$as_json" -eq 1 ]]; then
|
||
list_modes_data | python3 -c '
|
||
import sys, json
|
||
rows = []
|
||
for line in sys.stdin:
|
||
line = line.rstrip("\n")
|
||
if not line:
|
||
continue
|
||
parts = line.split("\t")
|
||
# pad to 6 fields so trailing-empty columns survive split
|
||
while len(parts) < 6:
|
||
parts.append("")
|
||
name, group, desc, services, ports, gpus = parts[:6]
|
||
rows.append({
|
||
"name": name,
|
||
"group": group,
|
||
"description": desc,
|
||
"services": [s for s in services.split(",") if s],
|
||
"ports": [p for p in ports.split(",") if p],
|
||
"gpus": gpus,
|
||
})
|
||
print(json.dumps(rows, indent=2))
|
||
'
|
||
else
|
||
echo ""
|
||
echo -e "${CYAN}═══ Scene Catalog ═══${NC}"
|
||
local cur_group=""
|
||
while IFS=$'\t' read -r name group desc services ports gpus; do
|
||
[[ -z "$name" ]] && continue
|
||
if [[ "$group" != "$cur_group" ]]; then
|
||
cur_group="$group"
|
||
echo ""
|
||
echo -e "${CYAN}[$group]${NC}"
|
||
fi
|
||
printf " %-16s %s\n" "$name" "$desc"
|
||
done < <(list_modes_data)
|
||
echo ""
|
||
echo -e "${YELLOW}--list-modes --json for the machine-readable catalog.${NC}"
|
||
fi
|
||
}
|
||
|
||
usage() {
|
||
echo ""
|
||
echo -e "${CYAN}GPU Mode Switcher${NC} — AI Inference Stack Manager"
|
||
echo ""
|
||
echo "Usage: gpu-mode <mode>"
|
||
echo ""
|
||
echo "Modes:"
|
||
echo " chat Open WebUI + LiteLLM + Qdrant + SearXNG (browser chat, no GPU model)"
|
||
echo ""
|
||
echo " Qwen 3.6 27B (dual 3090, TP=2):"
|
||
echo " 27b ⭐ DEFAULT — Qwen3.6-27B MTP + fp8 + 262K + vision + 2 streams (:8010)"
|
||
echo ""
|
||
echo " Gemma 4 31B (dual 3090, TP=2):"
|
||
echo " gemma ⭐ DEFAULT — Gemma 4 31B INT8 PTH KV + 262K + vision (:8032)"
|
||
echo ""
|
||
echo " Qwen 3.6 40B Deckard (uncensored, dual 3090, llama.cpp):"
|
||
echo " deckard Q6_K + MTP n=2 + q8_0 KV + 128K ctx (:8199) — text-only, both cards"
|
||
echo ""
|
||
echo " AI Studio (image · video · audio · voice — one scene, pick the lane in OWUI):"
|
||
echo " ai-studio ⭐ ComfyUI both GPUs + qwen director (:8090) + gallery/orchestrator/tts"
|
||
echo " + Open WebUI (:8080) — image/video/music/SFX/voice lanes (alias: aistudio)"
|
||
echo ""
|
||
echo " off Stop all services"
|
||
echo " status Show running services, GPU, RAM, disk, Docker disk"
|
||
echo ""
|
||
echo " GPU power cap (both 3090s; normally capped at 230W for quiet/cool operation):"
|
||
echo " power-cap on Re-apply the 230W cap (via nvidia-power-cap.service)"
|
||
echo " power-cap off Uncap to hardware default for a true-TPS bench"
|
||
echo " (session-scoped — a reboot / driver reload re-caps at 230W)"
|
||
echo " power-cap <WATTS> Apply a custom cap to both cards (e.g. 'power-cap 280';"
|
||
echo " validated against each card's [min,max]; session-scoped)"
|
||
echo " power-cap status Show per-GPU enforced / default / min / max power limits"
|
||
echo " (alias: powercap)"
|
||
echo ""
|
||
echo " Maintenance:"
|
||
echo " prune docker image prune -a (safe — only unreferenced images)"
|
||
echo " prune-all + build cache (keep 5 GB) + dangling networks (volumes safe)"
|
||
echo ""
|
||
}
|
||
|
||
case "${1:-}" in
|
||
chat) mode_chat ;;
|
||
27b) mode_27b ;;
|
||
gemma) mode_gemma_int8 ;;
|
||
deckard) mode_deckard ;;
|
||
ai-studio|aistudio) mode_ai_studio ;;
|
||
off) mode_off ;;
|
||
status) show_status ;;
|
||
power-cap|powercap) mode_powercap "${2:-status}" ;;
|
||
prune) mode_prune ;;
|
||
prune-all) mode_prune_all ;;
|
||
--list-modes) list_modes "${2:-}" ;;
|
||
*) usage ;;
|
||
esac
|