nmanga-list

To install dependencies:

bun install

To run:

bun run index.ts

This project was created using bun init in bun v1.3.12. Bun is a fast all-in-one JavaScript runtime.

Resource limits for large manga

Large manga can use a lot of RAM/CPU because image conversion + embedding run in worker threads.

Docker limits

docker/docker-compose.yml now includes an app service with hard caps:

• cpus: 2
• mem_limit: 2g

Tune those values based on your machine. Example:

app:
	cpus: 1.5
	mem_limit: 1536m

App-level throttles

Use these environment variables to reduce workload:

• IMAGE_QUEUE_WORKERS (default: 1)
  Number of manga processed at same time.
• IMAGE_PAGE_CONCURRENCY (default: 2)
  Number of pages fetched/converted at same time per worker.
• MANGA_REFRESH_INTERVAL_HOURS (default: 2)
  Minimum hours before /api/scrape re-fetches manga metadata/tags again. Within this interval, scrape request returns fresh and skips re-fetch.

Safe low-memory profile:

IMAGE_QUEUE_WORKERS=1
IMAGE_PAGE_CONCURRENCY=1

nhentai 429 handling with temporary proxies

When nhentai responds with HTTP 429, the app now:

• Reads Retry-After when available (falls back to 60s, capped at 10m)
• Shares one cooldown across all concurrent workers in the same app process
• Uses rotating proxies only while that cooldown is active

Proxy list is loaded once on app startup from ProxyScrape by default.

Optional env vars:

• NMANGA_NHENTAI_PROXY_ENABLED (default: true)
  Set to false to disable proxy usage entirely.
• NMANGA_NHENTAI_PROXY_LIST_URL (default: ProxyScrape elite list URL)
  Override with your own provider URL that returns one proxy per line in protocol://ip:port format.
• NMANGA_NHENTAI_FETCH_TIMEOUT_MS (default: 15000)
  Per-request timeout for nhentai direct/proxy fetches. Increase if your network/proxy path is slow; decrease to fail over faster.

Remote worker via CLI

You can run additional queue workers on another machine using the same database and S3/MinIO bucket:

bun run cli remote-worker

By default this command starts a worker subprocess with:

• env file: .env.remote
• concurrency: 1
• poll interval: 1000ms

Examples:

# 2 worker loops
bun run cli remote-worker --concurrency 2

# custom env file
bun run cli remote-worker --env-file /etc/nmanga/worker.env --concurrency 2

.env.remote must contain the server connection settings — not just the API URLs — because the Python embedding worker on the same machine connects to the database and bucket directly:

API_URL=http://10.0.0.11:3006
API_WS_URL=ws://10.0.0.11:3008/trpc
DATABASE_URL=postgresql://nmanga:CHANGE_ME@10.0.0.11:5432/nmanga?schema=public
MINIO_ENDPOINT=10.0.0.11
MINIO_PORT=9000
MINIO_USE_SSL=false
MINIO_ACCESS_KEY=minioadmin
MINIO_SECRET_KEY=minioadmin
MINIO_BUCKET=nmanga

Real environment variables take precedence; with bun --env-file the root .env is not loaded, and the Python workers read .env.remote themselves (falling back to .env for keys missing there).

All normal scraper/proxy env vars are supported, including:

• NMANGA_NHENTAI_PROXY_ENABLED
• NMANGA_NHENTAI_PROXY_LIST_URL
• NMANGA_NHENTAI_FETCH_TIMEOUT_MS

AMD GPU (ROCm) embedding worker

On Windows, uv sync installs ROCm-enabled PyTorch from AMD's index (torch 2.12.0+rocm10.1.0 for the Radeon RX 9070/XT, target device-gfx1201) instead of the CPU wheel. Other platforms (Docker/Linux) keep resolving regular PyPI wheels. The first sync downloads roughly 1.4 GB (ROCm runtime).

Requirements:

• Latest AMD Adrenalin driver
• Python 3.13 / 3.14 (managed by uv)

Verify the GPU is visible:

uv run python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0), torch.version.hip)"

Embedding workers use the GPU by default (--device cuda); rocm and hip are accepted as aliases:

uv run scripts/embed_worker.py --once   # continuous DB embedding worker
uv run scripts/reembed_images_gpu.py    # re-embed stored pages

For a different AMD GPU, change the device-gfx1201 extras on the torch/torchvision dependencies in pyproject.toml to the matching target (e.g. device-gfx1100 for the RX 7900 XTX), then run uv lock && uv sync.

If inference crashes or resets the GPU on RX 9070 series, run with TORCH_BLAS_PREFER_HIPBLASLT=0 to disable hipBLASLt.

Nginx Proxy Manager setup

This app serves HTTP API on port 3000 and tRPC WebSocket on port 3001.

In Nginx Proxy Manager, create a Proxy Host:

• Domain Names: nmanga.domain.dev
• Scheme: http
• Forward Hostname / IP: your app container or host (example: app)
• Forward Port: 3000
• Websockets Support: enabled

Then add this in Advanced:

location /trpc {
	set $nmanga_upstream http://app:3000;
	if ($http_upgrade = "websocket") {
		set $nmanga_upstream http://app:3001;
	}

	proxy_pass $nmanga_upstream;
	proxy_http_version 1.1;
	proxy_set_header Host $host;
	proxy_set_header X-Real-IP $remote_addr;
	proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
	proxy_set_header X-Forwarded-Proto $scheme;
	proxy_set_header Upgrade $http_upgrade;
	proxy_set_header Connection "upgrade";
	proxy_read_timeout 600s;
}

Replace app with your actual upstream hostname/IP used by NPM.