Update README.md
Browse files
README.md
CHANGED
@@ -9,29 +9,19 @@ pipeline_tag: text-generation
|
|
9 |
|
10 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.16">turboderp's ExLlamaV2 v0.0.16</a> for quantization.
|
11 |
|
12 |
-
|
13 |
|
14 |
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
|
15 |
|
16 |
-
Conversion was done using the default calibration dataset.
|
17 |
-
|
18 |
-
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.
|
19 |
-
|
20 |
Original model: https://huggingface.co/AetherResearch/Cerebrum-1.0-8x7b
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
<a href="https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/3_75">3.75 bits per weight</a>
|
30 |
-
|
31 |
-
<a href="https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/3_5">3.5 bits per weight</a>
|
32 |
-
|
33 |
-
<a href="https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/3_0">3.0 bits per weight</a>
|
34 |
-
|
35 |
|
36 |
## Download instructions
|
37 |
|
|
|
9 |
|
10 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.16">turboderp's ExLlamaV2 v0.0.16</a> for quantization.
|
11 |
|
12 |
+
<b>The "main" branch only contains the measurement.json, download one of the other branches for the model (see below)</b>
|
13 |
|
14 |
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
|
15 |
|
|
|
|
|
|
|
|
|
16 |
Original model: https://huggingface.co/AetherResearch/Cerebrum-1.0-8x7b
|
17 |
|
18 |
+
| Branch | Bits | lm_head bits | VRAM (4k) | VRAM (16k) | VRAM (32k) | Description |
|
19 |
+
| ------ | ---- | ------------ | ---- | ---- | ---- | ----------- |
|
20 |
+
| [6_5](https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/6_5) | 6.5 | 8.0 | 38.9 GB | 40.4 GB | 42.4 GB | Near unquantized performance at vastly reduced size, **recommended**. |
|
21 |
+
| [4_25](https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/4_25) | 4.25 | 6.0 | 25.9 GB | 27.4 GB | 29.4 GB | GPTQ equivalent bits per weight, slightly higher quality. |
|
22 |
+
| [3_75](https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/3_75) | 3.5 | 6.0 | 23.0 GB | 24.5 GB | 26.5 GB | Lower quality, but pretty usable. Good for 4k context on 24GB. |
|
23 |
+
| [3_5](https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/3_5) | 3.5 | 6.0 | 21.5 GB | 23.0 GB | 25.0 GB | Lower quality, only use if you need more context on 24GB. |
|
24 |
+
| [3_0](https://huggingface.co/bartowski/Cerebrum-1.0-8x7b-exl2/tree/3_0) | 3.0 | 6.0 | 18.9 GB | 20.4 GB | 22.4 GB | Very low quality, pushes context to max but likely unusable. |
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
## Download instructions
|
27 |
|