Sun-ZhenXing
111 lines
4.1 KiB
YAML
111 lines
4.1 KiB
YAML
x-defaults: &defaults
|
|
restart: unless-stopped
|
|
logging:
|
|
driver: json-file
|
|
options:
|
|
max-size: ${TURBOOCR_LOG_MAX_SIZE:-100m}
|
|
max-file: '${TURBOOCR_LOG_MAX_FILE:-3}'
|
|
|
|
x-turboocr-common: &turboocr-common
|
|
<<: *defaults
|
|
ports:
|
|
- '${TURBOOCR_HTTP_PORT_OVERRIDE:-8000}:8000'
|
|
- '${TURBOOCR_GRPC_PORT_OVERRIDE:-50051}:50051'
|
|
volumes:
|
|
# Named volume persists TRT engines (GPU) or ONNX model cache (CPU).
|
|
# Must be a named volume — bind-mounting an empty host dir shadows the
|
|
# baked-in language bundles and prevents the server from loading models.
|
|
- turboocr_build_cache:/home/ocr/.cache/turbo-ocr
|
|
environment:
|
|
- TZ=${TZ:-UTC}
|
|
# Language bundle: latin (default), chinese, greek, eslav, arabic, korean, thai
|
|
- OCR_LANG=${TURBOOCR_LANG:-}
|
|
# Set to 1 with OCR_LANG=chinese to use the 84 MB server rec model
|
|
- OCR_SERVER=${TURBOOCR_SERVER:-}
|
|
# Concurrent GPU pipelines (~1.4 GB VRAM each); empty = auto; ignored in CPU mode
|
|
- PIPELINE_POOL_SIZE=${TURBOOCR_PIPELINE_POOL_SIZE:-}
|
|
# Set to 1 to disable PP-DocLayoutV3 layout detection (saves ~300-500 MB VRAM)
|
|
- DISABLE_LAYOUT=${TURBOOCR_DISABLE_LAYOUT:-0}
|
|
# Default PDF mode: ocr (safest) / geometric / auto / auto_verified
|
|
- ENABLE_PDF_MODE=${TURBOOCR_PDF_MODE:-ocr}
|
|
# Skip angle classifier (~0.4 ms savings)
|
|
- DISABLE_ANGLE_CLS=${TURBOOCR_DISABLE_ANGLE_CLS:-0}
|
|
# Max detection input size in pixels
|
|
- DET_MAX_SIDE=${TURBOOCR_DET_MAX_SIDE:-960}
|
|
# PDF render parallelism
|
|
- PDF_DAEMONS=${TURBOOCR_PDF_DAEMONS:-16}
|
|
- PDF_WORKERS=${TURBOOCR_PDF_WORKERS:-4}
|
|
# Maximum pages per PDF request
|
|
- MAX_PDF_PAGES=${TURBOOCR_MAX_PDF_PAGES:-2000}
|
|
# Log level: debug / info / warn / error
|
|
- LOG_LEVEL=${TURBOOCR_LOG_LEVEL:-info}
|
|
# Log format: json (structured) / text (human-readable)
|
|
- LOG_FORMAT=${TURBOOCR_LOG_FORMAT:-json}
|
|
|
|
services:
|
|
turboocr-cuda12:
|
|
<<: *turboocr-common
|
|
profiles: [gpu]
|
|
build:
|
|
context: .
|
|
dockerfile: Dockerfile.cuda12
|
|
args:
|
|
TURBOOCR_VERSION: ${TURBOOCR_VERSION:-v2.1.1}
|
|
NGC_MIRROR: ${TURBOOCR_NGC_MIRROR:-}
|
|
image: ${GLOBAL_REGISTRY:-}alexsuntop/turboocr-cuda12:${TURBOOCR_VERSION:-v2.1.1}
|
|
healthcheck:
|
|
test: [CMD, curl, -fsS, 'http://localhost:8000/health']
|
|
interval: 30s
|
|
timeout: 10s
|
|
retries: 5
|
|
# First start builds 4 TensorRT engines from ONNX. Measured times on an
|
|
# RTX 3070 Laptop: det (~5 min) + rec (~30 min) + cls (~4 min) +
|
|
# layout (~28 min) = ~67-90 min. High-end desktop GPUs finish in ~15 min.
|
|
# Set TURBOOCR_DISABLE_LAYOUT=1 to skip layout and save ~28 min.
|
|
# Subsequent restarts reuse the cached engines and start in seconds.
|
|
start_period: 120m
|
|
deploy:
|
|
resources:
|
|
limits:
|
|
cpus: ${TURBOOCR_CPU_LIMIT:-8.0}
|
|
memory: ${TURBOOCR_MEMORY_LIMIT:-12G}
|
|
reservations:
|
|
cpus: ${TURBOOCR_CPU_RESERVATION:-2.0}
|
|
memory: ${TURBOOCR_MEMORY_RESERVATION:-4G}
|
|
devices:
|
|
- driver: nvidia
|
|
count: ${TURBOOCR_GPU_COUNT:-1}
|
|
capabilities: [gpu]
|
|
shm_size: ${TURBOOCR_SHM_SIZE:-2g}
|
|
|
|
turboocr-cpu:
|
|
<<: *turboocr-common
|
|
profiles: [cpu]
|
|
build:
|
|
context: .
|
|
dockerfile: Dockerfile.cpu
|
|
args:
|
|
TURBOOCR_VERSION: ${TURBOOCR_VERSION:-v2.1.1}
|
|
DOCKER_MIRROR: ${TURBOOCR_DOCKER_MIRROR:-}
|
|
image: ${GLOBAL_REGISTRY:-}alexsuntop/turboocr-cpu:${TURBOOCR_VERSION:-v2.1.1}
|
|
healthcheck:
|
|
test: [CMD, curl, -fsS, 'http://localhost:8000/health']
|
|
interval: 30s
|
|
timeout: 10s
|
|
retries: 5
|
|
# CPU mode uses ONNX Runtime directly — no TRT compilation on first start.
|
|
# Expect startup in under 60 s on most hardware.
|
|
start_period: 2m
|
|
deploy:
|
|
resources:
|
|
limits:
|
|
cpus: ${TURBOOCR_CPU_LIMIT:-8.0}
|
|
memory: ${TURBOOCR_MEMORY_LIMIT:-4G}
|
|
reservations:
|
|
cpus: ${TURBOOCR_CPU_RESERVATION:-2.0}
|
|
memory: ${TURBOOCR_MEMORY_RESERVATION:-1G}
|
|
shm_size: ${TURBOOCR_SHM_SIZE:-512m}
|
|
|
|
volumes:
|
|
turboocr_build_cache:
|