전체 자산 매니페스트(이미지 다이제스트·소스 위치·호스트 잔존).
# kube-rtx3 (192.168.0.54) 잔여 Vast.ai 임차인 자산 — 소스 전용 아카이브 매니페스트
# 추출일: 2026-06-24 / 출처 호스트: 192.168.0.54 (격리 dockerd 로 조회)
# 목적: 퍼블릭 베이스/모델 가중치는 제외하고, "임차인이 어떻게 서비스를 구성했는지"(커스텀
# 소스·워커·오케스트레이션·배포 스크립트)만 학습/복구용으로 보존.
# 주의: 전부 제3자 임차인 잔여물. .env/SSH키/터널 자격증명 등 비밀 포함 → 외부 공유 금지.
meta:
source_host: 192.168.0.54
source_node: kube-rtx3
extracted_at: "2026-06-24"
storage_driver_on_host: overlayfs (containerd snapshotter, docker 29.4.2)
weight_exclude_patterns: ["*.pth","*.safetensors","*.ckpt","*.pt","*.bin","*.onnx","*.gguf","*.h5","*.pkl","*.mp4/mov","*.png/jpg/webp","__pycache__",".venv","node_modules","*.log"]
lost:
# 1차 정리(2026-06-24) 때 "공개 ghcr=재pull 가능" 으로 잘못 분류되어 삭제됨.
# ghcr 익명 토큰 401 → 사실상 복구 불가. (남겼어야 할 항목)
- name: horse-bestview-keypoints-segmentation-detection-dino-worker
image: ghcr.io/sanchit2843/horse-bestview-keypoints-segmentation-detection-dino-worker:v6
digest: sha256:3a796a4276168afedb8b18c40e4fb72f3d16445a14c20ca3b078ab4699151b26
workdir: /app
cmd: ["sh","-c","python /app/horse_worker/preflight.py && python -m worker_runtime.main --launch-mode cloud_docker --run-id ${RUN_ID} --executor-id ${EXECUTOR_ID} --config ${RUN_CONFIG}"]
custom_source_that_was_inside:
- /app/worker_runtime (클라우드 워커 진입/실행)
- /app/orchestrator, /app/providers, /app/queues, /app/storage, /app/core, /app/config, /app/inferctl, /app/local_agent
- /app/horse_worker/{infer_one.py,preflight.py}
- /app/inference_bestview_keypoints/{keypoints.py,segdetdino.py,best_view.py,model_server.py,run.py}
- weights(삭제됨): yolo11l-seg.pt, dinov3_vitb16_pretrain*.pth, classifier_epoch_030*.pt, horse22_yolo26x_best.pt
stack: PyTorch 2.8 / Ultralytics YOLO11 / DINOv3 ViT / ffmpeg
recoverable: false
note: ghcr 비공개(401). sanchit2843 GitHub 자격증명 없으면 복구 불가.
deleted_repullable:
# 1차 정리(2026-06-24)에 삭제됨. 전부 공개 레지스트리 재pull 가능(가역). 빌드 히스토리로 역분석.
- name: perch7-img2
image: perch7/img2_v2.0.31/ssh:latest
digest: sha256:15a196af9b19125f850a8f75f1147ae60ac549dea862be55683a9fe9941b4486
size: 15.8GB
created: "2026-06-18"
service: FLUX 기반 img2img·txt2img 생성(port 3000). NVFP4(nunchaku)+SageAttention, 암호화 server.py.enc, lakonlab, taef1/T5/LoRA bake-in
custom: true
doc: perch7-img2-analysis.md
note: design-txt2img-paid-service.md의 실제 원형
- name: frosting-new-onf-server-sd
image: frosting/new-onf-server-sd:3.0.0
digest: sha256:a40ac8a02b7943d5db3348a7ee87425b438fe0e18baf483a935c5e1e1e8cfa9a
size: 12.7GB (보고서 비압축 42.1GB)
created: "2026-03-02"
service: Stable Diffusion FastAPI 생성 서버(uvicorn run:app :3000). onediff/oneflow 가속, Real-ESRGAN 업스케일, e621 임베딩
custom: true
doc: frosting-sd-server-analysis.md
- name: pearl-miner
images: [alphaminetech/pearl-miner:latest, alphaminetech/pearl-miner/ssh:latest]
digests: [sha256:72d8f940a19d8e2ecc711b4377451884dbfd89f87c2dc4df56ce0ed173596387, sha256:45a3ef4bab715a4b0e1963615ed3a83159e01340bdfbda81223d5070e4f59abf]
service: 암호화폐 채굴기(Pearl/Alpha, pool us2.alphapool.tech:5566)
custom: false
doc: crypto-miners-analysis.md
- name: xengpuminer
image: smit1237/xengpuminer:vast
digest: sha256:041ac8d700bcdbe862a99917f4c2c1d509ba24e01002154c3421d04d15e3e6bb
service: 암호화폐 채굴기(XEN GPU, github smit1237/XENGPUMiner)
custom: false
doc: crypto-miners-analysis.md
- name: vllm
image: vastai/vllm_v0.11.0-cuda-12.8-mvc-cuda-12.0/jupyter:latest
digest: sha256:c34ec0b21a10bf16970167f604cd696980e35d04d1ab00fc2e40ec5f6047c8aa
service: Vast.ai 공식 vLLM OpenAI 호환 LLM 추론 서버 + Jupyter
custom: false
doc: vllm-server-analysis.md
- name: vastai-test
images: [vastai/test:bandwidth-test-nvidia, vastai/test:speedtest, vastai/test:common]
digests: [sha256:d38b444d46c6d370e9aeb44877fc0dc77a7039704bcc1ba9ec6dc6b7a6a94394, sha256:d7bcbb6f2382d816e66a01696c64cac9e690ea002e5377fc3d6ec50c895aab3c, sha256:bd76fb63227e275111397c2b8d7eccd0134ad0efe634fdfd39fe5b3df07bf0b8]
service: Vast.ai 공식 하드웨어/네트워크 검증 이미지(CUDA bandwidthTest / Ookla speedtest / 공통 베이스)
custom: false
doc: vastai-benchmark-images-analysis.md
archived:
# ── ComfyUI 임차인 배포 스택 (가장 가치 높은 학습 자료) ──
- name: comfyui-zts
doc: comfyui-zts-analysis.md # ★최고가치 역분석(소스 전체 확보, 2026-06-25)
file: comfyui-zts.tar.gz
origin: 컨테이너 C.40368049 (vastai/comfy_v0.22.0-cuda-12.9-py312/jupyter) : /workspace/zts-comfyui-docker
base_image: vastai/comfy_v0.22.0-cuda-12.9-py312/jupyter:latest
base_digest: sha256:cf3a10c281c26fe4221b9e311b66b7a42ead8662de4fb15295f697fe293f10b7
contents:
- Dockerfile, docker-compose.yml, docker-compose.prod.yml, docker-entrypoint.sh, comfyui.service
- config/provisioning.sh, config/setup.yaml
- custom_nodes_local/zts_depthface_nodes/{__init__.py,nodes.py,requirements.txt} # 자체 커스텀 노드
- scripts/{auto-deploy-local.sh,auto-deploy-vast.sh,cron-install.sh,setup-multi-gpu.sh}
- workflows/*.json (ZIT V2 / wan2.2 i2v / infinite video 등)
- docs/API_MANUAL.md, README.md, .cursorrules, .gitignore
- .env (비밀 포함), .env.example, .git/ (히스토리 포함)
secrets: [".env"]
- name: comfy-zts_depthface_nodes
file: comfy-zts_depthface_nodes.tar.gz
origin: C.40368049 : /workspace/ComfyUI/custom_nodes/zts_depthface_nodes (배포 사본)
note: comfyui-zts 안 custom_nodes_local 과 동일 소스의 배포본.
# ── ECR 추론 서비스 래퍼 (server.py 가 핵심 = 모델을 HTTP 서비스로 감싼 방식) ──
# 3종 공통 역분석: ecr-inference-services-analysis.md (⚠ z-image-turbo/ddcolor 에 HF 토큰 하드코딩)
- name: ecr-z-image-turbo
doc: ecr-inference-services-analysis.md
file: ecr-z-image-turbo.tar.gz
image: 130745477593.dkr.ecr.us-west-2.amazonaws.com/z-image-turbo:latest
digest: sha256:89d511d311ae38dadfdc5a372b2d9e04fd128a43cbf64f1b6486e4ab5ebe5d53
ssh_variant_digest: sha256:d9b07306a0403d7a83644f015201f070aa7587504ab32c4afb8fe0dd986c36cb
workdir: /z-image-turbo
entrypoint: ["python","server.py"]
service: 텍스트→이미지 생성
contents: [server.py, hf_login_download_weights.py, requirements.txt]
- name: ecr-ddcolor
file: ecr-ddcolor.tar.gz
image: 130745477593.dkr.ecr.us-west-2.amazonaws.com/ddcolor:latest
digest: sha256:15a4f62b990b405e9b0b112c98f9e2627fe541bf5d737ac8e9cead82501cf662
ssh_variant_digest: sha256:6769edc57fd16675732e30a2b94764c8ea05ccf9eaa96f8f5cf15cdfe492d856
workdir: /ddcolor
entrypoint: ["python","server.py"]
service: 흑백→컬러 복원
contents: [server.py, colorization_pipeline.py, hf_login_download_weights.py, requirements.txt, DDColor/(공개 업스트림)]
- name: ecr-real-esrgan
file: ecr-real-esrgan.tar.gz
image: 130745477593.dkr.ecr.us-west-2.amazonaws.com/real-esrgan:latest
digest: sha256:e07c858dc29ada90de47fb1f077423b96c0974fe52b572a4dbaee890d5befed8
ssh_variant_digest: sha256:956e839f135e6befc4c1c71c9d9dfcaa38fe65ef71d54c595576c9cfb9b27c1b
workdir: /real-esrgan
entrypoint: ["python","server.py"]
service: 이미지 업스케일
contents: [server.py, inference_realesrgan.py, inference_realesrgan_video.py, realesrgan/, options/, scripts/, setup.py, requirements.txt]
note: ai_integration_pkg 래퍼가 베이스에 pip 설치되어 있었음(detail.txt 히스토리 참조)
# ────────────────────────────────────────────────────────────────────
# 호스트(.54)에만 보존 — genworks로 전송 안 함 (데이터 대용량, 현상 동결)
# 공부 시 .54에서 격리 dockerd 띄워 직접 접근 (아래 README "데이터 접근법" 참조).
# 공부 후 사용자가 삭제 예정.
# ────────────────────────────────────────────────────────────────────
on_host_only:
doc: host-only-containers-volumes-analysis.md # 컨테이너2(C.40368049,C.41508997)+볼륨8 분석. 볼륨 데이터 출처(자체수집 vs 다운로드)는 미확정→호스트 조회 필요
host: 192.168.0.54
docker_data_lv: /var/lib/docker (ubuntu-vg/docker_xfs, 800G, 557G 사용)
containers:
- id: C.40368049
base: vastai/comfy_v0.22.0-cuda-12.9-py312/jupyter:latest
data: ComfyUI 모델/.hf_home 캐시(대용량), custom_nodes, /workspace/zts-comfyui-docker(소스는 genworks에 사본)
- id: C.41508997
base: nvcr.io/nvidia/pytorch_26.01-py3/jupyter:latest
data: /workspace(onstart.sh 등), /root/.cloudflared/{cert.pem, e4f42ef8-...json}(터널 자격증명)
images_private_ecr: # 재pull 불가 — 호스트에만 존재
- z-image-turbo:latest / z-image-turbo/ssh:latest
- ddcolor:latest / ddcolor/ssh:latest
- real-esrgan:latest / real-esrgan/ssh:latest
volumes:
- {id: V.36552643, app: Applio RVC 음성클로닝(Honkai 캐릭터 음성모델), size: 4.9G}
- {id: V.36606356, app: Applio(보조), size: 75M}
- {id: V.36548605, app: VoiceBox TTS, size: 4.3G}
- {id: V.36547480, app: VoiceBox(설정), size: 144K}
- {id: V.40440012, app: stable-diffusion-webui(A1111), size: 301M}
- {id: V.36829118, app: 소형/빈, size: 8K}
- {ids: [V.36606894, V.37315681], app: 빈, size: 0}