Update VRAM estimates
Browse files
README.md
CHANGED
@@ -12,7 +12,7 @@ pipeline_tag: text-generation
|
|
12 |
|
13 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.13">turboderp's ExLlamaV2 v0.0.13</a> for quantization.
|
14 |
|
15 |
-
|
16 |
|
17 |
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
|
18 |
|
@@ -20,11 +20,11 @@ Original model: https://huggingface.co/serpdotai/sparsetral-16x7B-v2
|
|
20 |
|
21 |
| Branch | Bits | lm_head bits | VRAM (4k) | VRAM (16k) | VRAM (32k) | Description |
|
22 |
| ----- | ---- | ------- | ------ | ------ | ------ | ------------ |
|
23 |
-
| [8_0](https://huggingface.co/
|
24 |
-
| [6_5](https://huggingface.co/
|
25 |
-
| [5_0](https://huggingface.co/
|
26 |
-
| [4_25](https://huggingface.co/
|
27 |
-
| [3_5](https://huggingface.co/
|
28 |
|
29 |
## Download instructions
|
30 |
|
|
|
12 |
|
13 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.13">turboderp's ExLlamaV2 v0.0.13</a> for quantization.
|
14 |
|
15 |
+
<b>The "main" branch only contains the measurement.json, download one of the other branches for the model (see below)</b>
|
16 |
|
17 |
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
|
18 |
|
|
|
20 |
|
21 |
| Branch | Bits | lm_head bits | VRAM (4k) | VRAM (16k) | VRAM (32k) | Description |
|
22 |
| ----- | ---- | ------- | ------ | ------ | ------ | ------------ |
|
23 |
+
| [8_0](https://huggingface.co/bartowski/sparsetral-16x7B-v2-exl2/tree/8_0) | 8.0 | 8.0 | 8.3 GB | 9.7 GB | 11.8 GB | Maximum quality that ExLlamaV2 can produce, near unquantized performance. |
|
24 |
+
| [6_5](https://huggingface.co/bartowski/sparsetral-16x7B-v2-exl2/tree/6_5) | 6.5 | 8.0 | 7.1 GB | 8.5 GB | 10.6 GB | Very similar to 8.0, good tradeoff of size vs performance, **recommended**. |
|
25 |
+
| [5_0](https://huggingface.co/bartowski/sparsetral-16x7B-v2-exl2/tree/5_0) | 5.0 | 6.0 | 5.7 GB | 7.1 GB | 9.2 GB | Slightly lower quality vs 6.5, but usable on 8GB cards. |
|
26 |
+
| [4_25](https://huggingface.co/bartowski/sparsetral-16x7B-v2-exl2/tree/4_25) | 4.25 | 6.0 | 5.1 GB | 6.5 GB | 8.6 GB | GPTQ equivalent bits per weight, slightly higher quality. |
|
27 |
+
| [3_5](https://huggingface.co/bartowski/sparsetral-16x7B-v2-exl2/tree/3_5) | 3.5 | 6.0 | 4.4 GB | 5.8 GB | 7.9 GB | Lower quality, only use if you have to. |
|
28 |
|
29 |
## Download instructions
|
30 |
|