Update model card: remove MINT/SWAN branding, optimised by baa.ai
Browse files
README.md
CHANGED
|
@@ -4,20 +4,15 @@ tags:
|
|
| 4 |
- mlx
|
| 5 |
- quantized
|
| 6 |
- mixed-precision
|
| 7 |
-
- swan
|
| 8 |
license: other
|
| 9 |
license_name: polyform-noncommercial
|
| 10 |
base_model: THU-KEG/GLM-5-0817
|
| 11 |
base_model_relation: quantized
|
| 12 |
---
|
| 13 |
|
| 14 |
-
<p align="center">
|
| 15 |
-
<img src="https://huggingface.co/spaces/baa-ai/MINT/resolve/main/baa-logo.svg" width="300" alt="baa.ai">
|
| 16 |
-
</p>
|
| 17 |
-
|
| 18 |
# GLM-5-SWAN-5bit-MLX
|
| 19 |
|
| 20 |
-
Mixed-precision quantized version of [THUDM/GLM-5](https://huggingface.co/THUDM/GLM-5)
|
| 21 |
|
| 22 |
> GLM-5 (355B parameters). Experimental.
|
| 23 |
|
|
@@ -31,25 +26,6 @@ Mixed-precision quantized version of [THUDM/GLM-5](https://huggingface.co/THUDM/
|
|
| 31 |
| WikiText-2 PPL | — |
|
| 32 |
|
| 33 |
|
| 34 |
-
|
| 35 |
-
## 🚀 Create Your Own Custom Quantization
|
| 36 |
-
|
| 37 |
-
**Don't see the size you need?** Use [**MINT-UI**](https://github.com/baa-ai/MINT-UI) to create a custom-sized quantization targeting your exact memory budget:
|
| 38 |
-
|
| 39 |
-
```bash
|
| 40 |
-
pip install mint-ui
|
| 41 |
-
mint-ui
|
| 42 |
-
```
|
| 43 |
-
|
| 44 |
-
MINT-UI analyzes any model in **under 60 seconds** using a cutting-edge allocation technique — no calibration data needed. Specify your exact memory target (e.g., "fit in 24 GB for RTX 4090") and MINT returns a near-optimal per-tensor bit-width allocation.
|
| 45 |
-
|
| 46 |
-
- ⚡ **60 seconds** analysis (vs hours for GPTQ/AWQ calibration)
|
| 47 |
-
- 🎯 **Any target size** — not limited to uniform 4-bit or 8-bit
|
| 48 |
-
- 🧠 **Data-free** — no calibration dataset required
|
| 49 |
-
- 💻 **Runs on any Mac** — even 32 GB machines can analyze 400B models
|
| 50 |
-
|
| 51 |
-
👉 **[Get MINT-UI](https://github.com/baa-ai/MINT-UI)** | 📄 **[MINT Paper](https://github.com/baa-ai/MINT) | [MINT-UI](https://github.com/baa-ai/MINT-UI)** | 🤗 **[All Models](https://huggingface.co/baa-ai)**
|
| 52 |
-
|
| 53 |
## Usage
|
| 54 |
|
| 55 |
```python
|
|
@@ -60,11 +36,6 @@ response = generate(model, tokenizer, prompt="Hello!", max_tokens=256)
|
|
| 60 |
print(response)
|
| 61 |
```
|
| 62 |
|
| 63 |
-
## About SWAN
|
| 64 |
-
|
| 65 |
-
SWAN uses data-free per-tensor sensitivity analysis with composite scoring to allocate bit-widths across model layers.
|
| 66 |
-
|
| 67 |
-
- [Paper](https://huggingface.co/spaces/baa-ai/MINT) | [Code](https://github.com/baa-ai/MINT) | [MINT-UI](https://github.com/baa-ai/MINT-UI) | [Models](https://huggingface.co/baa-ai)
|
| 68 |
|
| 69 |
---
|
| 70 |
*Quantized by [baa.ai](https://baa.ai)*
|
|
|
|
| 4 |
- mlx
|
| 5 |
- quantized
|
| 6 |
- mixed-precision
|
|
|
|
| 7 |
license: other
|
| 8 |
license_name: polyform-noncommercial
|
| 9 |
base_model: THU-KEG/GLM-5-0817
|
| 10 |
base_model_relation: quantized
|
| 11 |
---
|
| 12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
# GLM-5-SWAN-5bit-MLX
|
| 14 |
|
| 15 |
+
Mixed-precision quantized version of [THUDM/GLM-5](https://huggingface.co/THUDM/GLM-5) optimised by [baa.ai](https://baa.ai).
|
| 16 |
|
| 17 |
> GLM-5 (355B parameters). Experimental.
|
| 18 |
|
|
|
|
| 26 |
| WikiText-2 PPL | — |
|
| 27 |
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
## Usage
|
| 30 |
|
| 31 |
```python
|
|
|
|
| 36 |
print(response)
|
| 37 |
```
|
| 38 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
---
|
| 41 |
*Quantized by [baa.ai](https://baa.ai)*
|