# phi-2 GPU requirements

phi-2 has 2.8B parameters. To serve it, plan for about 6.7 GB of GPU memory in 16-bit, 3.3 GB in 8-bit or 1.7 GB in 4-bit, including about 20% for the KV cache and activations.

[Deploy it on Arrakis](/console/#/models?model=microsoft/phi-2).

## What it takes to serve it

| Precision | Memory and the smallest setup |
|---|---|
| 16-bit | 6.7 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |
| 8-bit | 3.3 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |
| 4-bit | 1.7 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |

## What it takes to fine-tune it

| Method | Memory and the smallest setup |
|---|---|
| QLoRA | 1.9 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |
| LoRA | 6.7 GB: one [RTX 4090](/gpus/rtx-4090.html.md) or [RTX 3090](/gpus/rtx-3090.html.md), RTX 3090 from $0.19 per GPU-hour |
| Full fine-tuning | 44 GB: one [L40S](/gpus/l40s.html.md), [RTX 6000 Ada](/gpus/rtx-6000-ada.html.md) or [RTX A6000](/gpus/rtx-a6000.html.md), L40S from $0.44 per GPU-hour |

## About the model

| | |
|---|---|
| Parameters | `2.8B` |
| Context window (tokens) | `2,048` |
| Published in | `BF16` |
| License | `mit` |
| Released | `2023-12-13` |
| Weights | [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) |

## Serve it yourself

```
vllm serve microsoft/phi-2 --tensor-parallel-size 1
```

## Questions

**How much GPU memory does phi-2 need?** To serve it, about 7 GB in 16-bit, 3 GB in 8-bit, 2 GB in 4-bit, including about 20% for the KV cache and activations. To fine-tune it, about 2 GB with QLoRA, 7 GB with LoRA and 44 GB for full fine-tuning with Adam, before activations.

**Can I run phi-2 on one GPU?** Yes. In 16-bit it needs about 6.7 GB, which fits on one RTX 4090 or RTX 3090.

**What does it cost to run phi-2?** In 16-bit it needs one RTX 4090 or RTX 3090; in 4-bit, one RTX 4090 or RTX 3090. Each setup on this page shows the lowest price per GPU-hour on Arrakis right now, and you pay by the hour from a prepaid balance.

**How do I deploy phi-2 on Arrakis?** Open Models in the Arrakis console, paste microsoft/phi-2, choose vLLM, SGLang or llama.cpp, and rent a GPU it fits on. You get an OpenAI-compatible endpoint with its own API key, and root SSH to the machine.


More models: [/models](/models.html.md).