Deploy

There are three ways to get a mineru-runpod endpoint running. Pick the one that matches what you need to control.

Option A — Deploy from the RunPod Hub (easiest)

This repo is published as a public Hub template. In the RunPod dashboard go to The Hub → Serverless repos, find mineru-runpod, click Deploy. RunPod builds the image on your account, you pick a GPU pool, and you get an endpoint id. No fork, no clone, no local setup.

This is the recommended path if your goal is parsing PDFs, not customising the worker.

Option B — Fork and auto-build (for customization)

Fork this repo into your own GitHub account if you want to:

• Pin different versions of MinerU, vLLM, or other dependencies
• Modify handler.py (custom input validation, extra output formats, etc.)
• Run on a private GitHub repo

Then in the RunPod dashboard:

1. The Hub → Serverless repos → Import Git Repository, point at your fork. Branch main, Dockerfile path Dockerfile.
2. RunPod builds the image (~5–10 min, watchable in the dashboard) and gives you a template_id.
3. Create the endpoint. Pick one:
   • (B1) Dashboard, no local Python needed: Resources → Serverless → New Endpoint, select your template, set idle_timeout=10, workers_min=0, workers_max=3, FlashBoot on, GPU pool ADA_24 (RTX 4090). Save and grab the endpoint id.
   • (B2) As code, reproducible across deployments:

     cp .env.example .env       # fill RUNPOD_API_KEY and MINERU_TEMPLATE_ID
     pip install -e .[deploy]
     python deploy.py --template-id $env:MINERU_TEMPLATE_ID

     Every knob in deploy.py --help matches a setting in the dashboard.

Subsequent pushes to main on your fork rebuild the image automatically; the endpoint picks up the new image on next cold start (or force a redeploy from the dashboard).

Option C — Bring your own image

For full control over the Docker layer, build and push to Docker Hub or GHCR yourself, then:

python deploy.py --image yourhandle/mineru-runpod:0.1

This skips RunPod’s auto-build entirely. Use it when you need custom base images, multi-stage builds, or air-gapped registries.

Endpoint defaults

Setting	Value	Why
gpu_ids	ADA_24	24 GB Ada / RTX 4090. Fits the MinerU2.5-Pro-2605-1.2B VLM comfortably with KV cache; faster per page than the cheaper A5000 (AMPERE_24).
idle_timeout	10 s	Scale workers to zero after 10 s of inactivity
workers_min	0	Pay only when processing
workers_max	3	Concurrency cap; bump for production
execution_timeout	900 s	Per-job cap; covers a several-hundred-page parse
flashboot	true	RunPod’s fast cold-start tech

Override any of these via flags to deploy.py (e.g. --gpu-ids AMPERE_24 --workers-max 5).

See Choosing a GPU for when to deviate from the default 24 GB pool.

What you get after deploy

An endpoint id like abcdef123456 that accepts the documented job input contract via RunPod’s standard /run and /runsync endpoints. From there, see Clients for how to actually call it from your code.