File size: 1,721 Bytes
7ffca1d
b148141
 
 
 
7ffca1d
b148141
7ffca1d
 
 
b148141
 
 
 
 
 
7ffca1d
 
b148141
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
---
title: VRAM Calculator
emoji: 🧮
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
short_description: Calculate VRAM requirements for HuggingFace models
tags:
  - vram
  - gpu
  - inference
  - deployment
---

# 🧮 VRAM & Instance Type Calculator

Estimate GPU memory requirements for any HuggingFace model and get cloud instance recommendations.

## Features

- **Automatic model analysis**: Fetches parameter count, dtype, and architecture from HF Hub
- **KV cache estimation**: Calculates memory for different context lengths
- **GPU recommendations**: Shows which GPUs can run the model (RTX 3090 → H100)
- **Cloud instance mapping**: Suggests AWS/GCP instance types with pricing
- **Quantization guidance**: Suggests INT8/INT4 options for large models

## How it works

1. Fetches `safetensors` metadata for parameter count and dtype
2. Downloads `config.json` for architecture details (layers, hidden size, KV heads)
3. Calculates:
   - Model weights: `params × dtype_bytes`
   - KV cache: `2 × layers × batch × seq_len × kv_heads × head_dim × dtype_bytes`
   - Adds ~15% overhead for activations

## Limitations

- Estimates are for inference, not training
- Actual VRAM varies by serving framework (vLLM vs TGI vs vanilla)
- GGUF/quantized models have different memory profiles
- Does not account for tensor parallelism across multiple GPUs

## Usage

```python
# Or run locally:
pip install gradio huggingface_hub
python app.py
```

## Contributing

PRs welcome! Ideas for improvement:
- Add support for GGUF models
- Include throughput estimates
- Add more cloud providers (Azure, Lambda Labs)
- Support tensor parallelism calculations