Update README.md
Browse files
README.md
CHANGED
@@ -19,26 +19,19 @@ pipeline_tag: text-generation
|
|
19 |
|
20 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.16">turboderp's ExLlamaV2 v0.0.16</a> for quantization.
|
21 |
|
22 |
-
|
23 |
|
24 |
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
|
25 |
|
26 |
-
Conversion was done using the default calibration dataset.
|
27 |
-
|
28 |
-
Default arguments used except when the bits per weight is above 6.0, at that point the lm_head layer is quantized at 8 bits per weight instead of the default 6.
|
29 |
-
|
30 |
Original model: https://huggingface.co/mlabonne/Beyonder-4x7B-v3
|
31 |
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
<a href="https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/4_25">4.25 bits per weight</a>
|
40 |
-
|
41 |
-
<a href="https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/3_5">3.5 bits per weight</a>
|
42 |
|
43 |
|
44 |
## Download instructions
|
|
|
19 |
|
20 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.16">turboderp's ExLlamaV2 v0.0.16</a> for quantization.
|
21 |
|
22 |
+
<b>The "main" branch only contains the measurement.json, download one of the other branches for the model (see below)</b>
|
23 |
|
24 |
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
|
25 |
|
|
|
|
|
|
|
|
|
26 |
Original model: https://huggingface.co/mlabonne/Beyonder-4x7B-v3
|
27 |
|
28 |
+
| Branch | Bits | lm_head bits | VRAM (4k) | VRAM (16k) | VRAM (32k) | Description |
|
29 |
+
| ------ | ---- | ------------ | ---- | ---- | ---- | ----------- |
|
30 |
+
| [8_0](https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/8_0) | 8.0 | 8.0 | 24.8 GB | 26.3 GB | 28.3 GB | Maximum quality that ExLlamaV2 can produce, near unquantized performance. |
|
31 |
+
| [6_5](https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/6_5) | 6.5 | 8.0 | 20.3 GB | 21.8 GB | 23.8 GB | Near unquantized performance at vastly reduced size, **recommended**. |
|
32 |
+
| [5_0](https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/5_0) | 5.0 | 6.0 | 15.8 GB | 17.3 GB | 19.3 GB | Slightly lower quality vs 6.5. |
|
33 |
+
| [4_25](https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/4_25) | 4.25 | 6.0 | 14.0 GB | 15.5 GB | 17.5 GB | GPTQ equivalent bits per weight, slightly higher quality, great for 16gb cards with 16k context. |
|
34 |
+
| [3_5](https://huggingface.co/bartowski/Beyonder-4x7B-v3-exl2/tree/3_5) | 3.5 | 6.0 | 11.3 GB | 12.8 GB | 14.8 GB | Lower quality, not recommended, only suitable for 12GB cards. |
|
|
|
|
|
|
|
35 |
|
36 |
|
37 |
## Download instructions
|