Text-to-Video
Diffusers
Safetensors
MiniMax H3
video-generation
quantized
torchao
nvfp4
int8
nf4
bitsandbytes
Instructions to use corechan/MiniMax_H3_Torchao018 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use corechan/MiniMax_H3_Torchao018 with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("corechan/MiniMax_H3_Torchao018", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: minimax-h3-community-license
|
| 4 |
+
license_link: https://huggingface.co/corechan/MiniMax-H3_NVFP4_INT8/resolve/main/LICENSE
|
| 5 |
+
tags:
|
| 6 |
+
- text-generation
|
| 7 |
+
- text-to-video
|
| 8 |
+
- video-generation
|
| 9 |
+
- quantized
|
| 10 |
+
- nvfp4
|
| 11 |
+
- int8
|
| 12 |
+
- torchao
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# MiniMax-H3 (torchao NVFP4 + INT8)
|
| 16 |
+
|
| 17 |
+
This repository provides an optimized quantized version of the **MiniMax-H3** model targeting **`torchao` (>= 0.18.0)** to significantly reduce VRAM consumption during inference while maintaining generation fidelity.
|
| 18 |
+
|
| 19 |
+
- **Transformer (DiT):** Quantized to **NVFP4** (NVIDIA FP4)
|
| 20 |
+
- **Text Encoder:** Quantized to **INT8**
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## Original Model Details
|
| 25 |
+
* **Developed by:** MiniMax
|
| 26 |
+
* **Original Repository:** [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3)
|
| 27 |
+
|
| 28 |
+
## Quantization Details
|
| 29 |
+
* **Framework:** [`torchao`](https://github.com/pytorch/ao) (version `0.18.0` or higher required)
|
| 30 |
+
* **Transformer Architecture:** NVFP4 (NVIDIA FP4 Linear Quantization)
|
| 31 |
+
* **Text Encoder:** INT8 Linear Quantization
|
| 32 |
+
* **Target:** Linear layers
|
| 33 |
+
|
| 34 |
+
### Requirements
|
| 35 |
+
```bash
|
| 36 |
+
pip install torchao>=0.18.0
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
---
|
| 40 |
+
|
| 41 |
+
## License and Terms of Use ⚠️ IMPORTANT
|
| 42 |
+
|
| 43 |
+
This model is a derivative work of MiniMax-H3 and is distributed under the **MiniMax H3 Community License Agreement**. By downloading or using this model, you and your downstream users agree to be bound by all terms, conditions, and the Acceptable Use Policy of the original license.
|
| 44 |
+
|
| 45 |
+
### Key Terms to Remember:
|
| 46 |
+
1. **Commercial Use Threshold:** Free commercial use, modification, and redistribution are permitted **ONLY** for entities with an annual revenue of **less than $20 million USD**. If your annual revenue exceeds this threshold, you must obtain a separate written license from MiniMax.
|
| 47 |
+
2. **Excluded Territories:** This community license **DOES NOT** grant any rights to use, display, or distribute this model or its outputs within the following territories: **The United States, The European Union, The United Kingdom, and South Korea**. Users in these regions must contact MiniMax directly for separate authorization.
|
| 48 |
+
3. **No Cross-Model Improvement:** You may not use this model or its outputs to train, align, or improve any AI models other than MiniMax-H3 or its direct derivatives.
|
| 49 |
+
4. **Attribution:** Any commercial products or interfaces utilizing this model must display appropriate credit to MiniMax.
|
| 50 |
+
|
| 51 |
+
For the full license terms, please read the `LICENSE` file in this repository or refer to the official [MiniMax H3 Community License Agreement](https://huggingface.co).
|
| 52 |
+
|
| 53 |
+
---
|
| 54 |
+
|
| 55 |
+
## 日本語概要 (Japanese Summary)
|
| 56 |
+
|
| 57 |
+
本リポジトリは **MiniMax-H3** を **`torchao` (0.18.0+)** 向けに量子化したモデルです。
|
| 58 |
+
|
| 59 |
+
* **Transformer (DiT部):** **NVFP4 (NVIDIA FP4)** 量子化
|
| 60 |
+
* **Text Encoder (テキストエンコーダー):** **INT8** 量子化
|
| 61 |
+
|
| 62 |
+
### 必要な環境
|
| 63 |
+
利用には `torchao >= 0.18.0` がインストールされている必要があります。
|
| 64 |
+
|
| 65 |
+
### ライセンスに関する重要な注意点
|
| 66 |
+
本モデルは独自の **MiniMax H3 Community License Agreement** に基づいて配布されています。ダウンロードおよび利用にあたっては、以下の制限が自動的に適用されます。
|
| 67 |
+
|
| 68 |
+
* **商用利用の制限**: 年間売上が **2,000万米ドル(約31億円)未満** の組織・個人に限り、無料での商用利用・改変・再配布が認められています。これを超える場合はMiniMax社との個別契約が必要です。
|
| 69 |
+
* **地域制限(適用外地域)**: **アメリカ、EU、イギリス、韓国** の4地域は本ライセンスの適用外(除外地域)です。これらの地域での利用や、出力結果の表示・頒布には別途許諾が必要です。
|
| 70 |
+
* **禁止事項**: 本モデルの出力成果物を、MiniMax-H3シリーズ以外の別のAIモデルの学習や改善(蒸留など)に利用することは禁止されています。
|
| 71 |
+
* **帰属表示**: 商用プロダクトで利用する際は、MiniMax社のベースモデルである旨のクレジット表記が必要です。
|
| 72 |
+
|
| 73 |
+
---
|
| 74 |
+
|
| 75 |
+
### Disclaimer
|
| 76 |
+
The author of this repository is not affiliated with MiniMax. This quantized version is provided "as-is" for the community, and users are solely responsible for ensuring compliance with the original license terms.
|
| 77 |
+
```
|