Update VRAM estimates
Browse files
README.md
CHANGED
@@ -12,26 +12,21 @@ pipeline_tag: text-generation
|
|
12 |
|
13 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.14">turboderp's ExLlamaV2 v0.0.14</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 |
|
19 |
-
Conversion was done using the default calibration dataset.
|
20 |
-
|
21 |
-
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.
|
22 |
-
|
23 |
Original model: https://huggingface.co/TIGER-Lab/StructLM-7B
|
24 |
|
|
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
<a href="https://huggingface.co/bartowski/StructLM-7B-exl2/tree/3_5">3.5 bits per weight</a>
|
35 |
|
36 |
|
37 |
## Download instructions
|
|
|
12 |
|
13 |
Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.14">turboderp's ExLlamaV2 v0.0.14</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 |
|
|
|
|
|
|
|
|
|
19 |
Original model: https://huggingface.co/TIGER-Lab/StructLM-7B
|
20 |
|
21 |
+
No GQA - VRAM requirements will be higher
|
22 |
|
23 |
+
| Branch | Bits | lm_head bits | Size (4k) | Size (16k) | Description |
|
24 |
+
| -------------------------------------------------------------- | ---- | ------------ | --------- | ---------- | ----------- |
|
25 |
+
| [8_0](https://huggingface.co/bartowski/StructLM-7B-exl2/tree/8_0) | 8.0 | 8.0 | 9.0 GB | 15.2 GB | Maximum quality that ExLlamaV2 can produce, near unquantized performance. |
|
26 |
+
| [6_5](https://huggingface.co/bartowski/StructLM-7B-exl2/tree/6_5) | 6.5 | 8.0 | 8.2 GB | 14.4 GB | Near unquantized performance at vastly reduced size, **recommended**. |
|
27 |
+
| [5_0](https://huggingface.co/bartowski/StructLM-7B-exl2/tree/5_0) | 5.0 | 6.0 | 6.8 GB | 13.0 GB | Slightly lower quality vs 6.5, but usable on 8GB cards with 4k context. |
|
28 |
+
| [4_25](https://huggingface.co/bartowski/StructLM-7B-exl2/tree/4_25) | 4.25 | 6.0 | 6.1 GB | 12.3 GB | GPTQ equivalent bits per weight. |
|
29 |
+
| [3_5](https://huggingface.co/bartowski/StructLM-7B-exl2/tree/3_5) | 3.5 | 6.0 | 5.5 GB | 11.7 GB | Lower quality, not recommended. |
|
|
|
|
|
30 |
|
31 |
|
32 |
## Download instructions
|