# Mining on rented GPUs (vast.ai / RunPod / Clore) — no Docker-in-Docker

Rental providers give you a **container**, and most don't let you run `docker`
inside it — so the normal one-liner installer (which does `docker run`) won't
work there. You have two options.

---

## Option A — use our image AS the instance image (recommended)

Every provider lets you pick the **docker image the instance boots from**, plus
environment variables. Our miner image is self-contained (it starts itself via
supervisord), so you just point the instance at it — no nesting, no installer.

**Image:** `bobponey/tensorcash-pool-miner:with-model`
(the pool edition — pool endpoint, join token, GPU auto-sizing, **and the 16 GB
model weights baked in**: the instance pulls once from Docker Hub, loads the
model to VRAM in ~1-2 min, and mines. No hidden HuggingFace download.)

If your provider charges by image size or you prefer a smaller pull, the thin
variant `bobponey/tensorcash-pool-miner:latest` (17 GB) works identically but
downloads the model from HuggingFace on first boot (10-40 min of quiet
"nothing happening" in the logs — it is not stuck).

**On-start / command:** leave **empty** — the image already launches mining.

**Shared memory:** vLLM needs a real `/dev/shm`. Set `--shm-size=4g` (or the
provider's "shared memory" field) to at least a few GB.

**One instance = one GPU.** On a multi-GPU instance, either create one instance
per GPU, or set `CUDA_VISIBLE_DEVICES=0` (then `1`, …) per instance.

### Environment variables — just one

```
WALLET=tc1qYOURADDRESS
```

Optional:
```
LABEL=my-rig            # names this machine on the dashboard (default: hostname)
CUDA_VISIBLE_DEVICES=0  # pin to one GPU on a multi-GPU instance
```

Everything else — pool endpoint, join token, broker mode, and the vLLM sizing
for your card (3090/4090 vs 5090 presets, VRAM auto-detected at boot) — is
built into the image. Any of it can still be overridden by setting the
variable explicitly (if you override sizing, set the whole block, not one
value). No comma-containing values are needed, so comma-stripping provider
forms (Clore) are no longer an issue.

> The card must have **≥ 24 GB** VRAM: the Qwen3-8B mining model is ~16.4 GB in
> bf16 and can't be quantized (the proof commits to exact weights). 16 GB cards
> (e.g. RTX 5080) cannot mine.
>
> It must also be **Ampere or newer** (compute capability ≥ 8.0). Turing cards
> have no bf16 and fall back to fp16, which diverges from the checkpoint the
> chain verifies against — every share is rejected. Watch for these on
> marketplaces, since they pass a VRAM filter: **Titan RTX (24 GB)**, **Quadro
> RTX 6000/8000**, and **22 GB modded RTX 2080 Ti** listings. The installer
> refuses them, but you'd be paying for the instance by then.

### Per-provider notes

- **vast.ai** — "Edit Image & Config": set the image, paste the env vars, leave
  the on-start script blank (or `sleep infinity` if it insists on one — mining
  runs regardless), set shared memory ≥ 4 GB.
- **RunPod** — Custom template: Container Image = the image above, Container
  Start Command = empty, add the env vars, set a GPU with ≥ 24 GB. Logs show
  vLLM startup directly (the image logs to stdout).
- **Clore.ai** — create an "on-demand" order with a custom image, paste the env
  vars, set shared memory.

Once it's up, look up your wallet at https://tensorcash.boblabs.eu — shares
appear within a minute or two of the model finishing loading.

---

## Option B — locked container you can't re-image? Run natively (no Docker)

If you're stuck inside a container whose image you can't choose, run the miner
**natively** with our helper. It downloads a flattened copy of the image and
runs it under `proot` (userspace — no root, no Docker), reusing the exact same
Python/vLLM environment.

```
curl -fsSL https://tensorcash.boblabs.eu/install_native.sh | bash -s -- --wallet tc1qYOURADDRESS --gpu 0
```

Requirements inside the rented container: an NVIDIA GPU with the driver present
(`nvidia-smi` works), `curl`, and ~20 GB of free disk. The download is
resumable; installing `aria2` first makes it faster. Add `--service` to keep it
running in the background with auto-restart.