Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
library_name: exllamav3
|
| 2 |
+
tags:
|
| 3 |
+
- exl3
|
| 4 |
+
- exllamav3
|
| 5 |
+
- quantized
|
| 6 |
+
- dflash
|
| 7 |
+
- speculative-decoding
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# Gemma 4 12B IT DFlash (EXL3)
|
| 11 |
+
|
| 12 |
+
This is an **ExLlamaV3 (EXL3)** quantized version of the `gemma4-12B-it-DFlash` model.
|
| 13 |
+
|
| 14 |
+
DFlash models are specifically designed to act as **Draft Models** for Speculative Decoding, allowing you to generate text significantly faster when paired with a larger main model. These were quantized using the Ampere-optimized `mul1` codebook for maximum decoding speed on modern GPUs.
|
| 15 |
+
|
| 16 |
+
## 📥 Available Branches / Downloads
|
| 17 |
+
|
| 18 |
+
The model weights are stored in separate branches. **Please click a branch link below to view and download the specific size you need.**
|
| 19 |
+
|
| 20 |
+
| Target BPW | Head BPW | Branch (Download Link) |
|
| 21 |
+
|:---|:---:|:---|
|
| 22 |
+
| **6.00** | 6 | [6.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/6.0bpw_h6) |
|
| 23 |
+
| **5.00** | 6 | [5.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/5.0bpw_h6) |
|
| 24 |
+
| **4.00** | 6 | [4.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/4.0bpw_h6) |
|
| 25 |
+
| **3.50** | 6 | [3.5bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/3.5bpw_h6) |
|
| 26 |
+
| **3.00** | 6 | [3.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/3.0bpw_h6) |
|
| 27 |
+
| **2.50** | 6 | [2.5bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/2.5bpw_h6) |
|
| 28 |
+
|
| 29 |
+
## 💻 How to Download
|
| 30 |
+
|
| 31 |
+
It is highly recommended to use the `huggingface-cli` to download a specific branch to avoid downloading the entire repository (which contains all sizes).
|
| 32 |
+
|
| 33 |
+
```bash
|
| 34 |
+
pip install -U "huggingface_hub[cli]"
|
| 35 |
+
```
|
| 36 |
+
|
| 37 |
+
**Example: Downloading the 4.0 BPW branch**
|
| 38 |
+
```bash
|
| 39 |
+
huggingface-cli download DBMe/gemma4-12B-it-DFlash-exl3 \
|
| 40 |
+
--revision 4.0bpw_h6 \
|
| 41 |
+
--local-dir gemma4-12B-it-DFlash-exl3-4.0bpw
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
## ⚙️ Quantization Details
|
| 45 |
+
* **Engine:** ExLlamaV3
|
| 46 |
+
* **Codebook:** `mul1` (Ampere-optimized)
|
| 47 |
+
* **Head BPW:** 6.0 (Preserves output layer accuracy for drafting)
|
| 48 |
+
* **Calibration:** Uncalibrated (Standard practice for DFlash architectures)
|