Add default model card
Browse files
README.md
CHANGED
@@ -1,13 +1,69 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
inference: false
|
|
|
3 |
---
|
4 |
# Nitral-AI/Hathor-L3-8B-v.02 AWQ
|
5 |
|
6 |
-
** PROCESSING .... ETA 30mins **
|
7 |
-
|
8 |
- Model creator: [Nitral-AI](https://huggingface.co/Nitral-AI)
|
9 |
- Original model: [Hathor-L3-8B-v.02](https://huggingface.co/Nitral-AI/Hathor-L3-8B-v.02)
|
10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
### About AWQ
|
12 |
|
13 |
AWQ is an efficient, accurate and blazing-fast low-bit weight quantization method, currently supporting 4-bit quantization. Compared to GPTQ, it offers faster Transformers-based inference with equivalent or better quality compared to the most commonly used GPTQ settings.
|
|
|
1 |
---
|
2 |
+
library_name: transformers
|
3 |
+
tags:
|
4 |
+
- 4-bit
|
5 |
+
- AWQ
|
6 |
+
- text-generation
|
7 |
+
- autotrain_compatible
|
8 |
+
- endpoints_compatible
|
9 |
+
pipeline_tag: text-generation
|
10 |
inference: false
|
11 |
+
quantized_by: Suparious
|
12 |
---
|
13 |
# Nitral-AI/Hathor-L3-8B-v.02 AWQ
|
14 |
|
|
|
|
|
15 |
- Model creator: [Nitral-AI](https://huggingface.co/Nitral-AI)
|
16 |
- Original model: [Hathor-L3-8B-v.02](https://huggingface.co/Nitral-AI/Hathor-L3-8B-v.02)
|
17 |
|
18 |
+
|
19 |
+
|
20 |
+
## How to use
|
21 |
+
|
22 |
+
### Install the necessary packages
|
23 |
+
|
24 |
+
```bash
|
25 |
+
pip install --upgrade autoawq autoawq-kernels
|
26 |
+
```
|
27 |
+
|
28 |
+
### Example Python code
|
29 |
+
|
30 |
+
```python
|
31 |
+
from awq import AutoAWQForCausalLM
|
32 |
+
from transformers import AutoTokenizer, TextStreamer
|
33 |
+
|
34 |
+
model_path = "solidrust/Hathor-L3-8B-v.02-AWQ"
|
35 |
+
system_message = "You are Hathor-L3-8B-v.02, incarnated as a powerful AI. You were created by Nitral-AI."
|
36 |
+
|
37 |
+
# Load model
|
38 |
+
model = AutoAWQForCausalLM.from_quantized(model_path,
|
39 |
+
fuse_layers=True)
|
40 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path,
|
41 |
+
trust_remote_code=True)
|
42 |
+
streamer = TextStreamer(tokenizer,
|
43 |
+
skip_prompt=True,
|
44 |
+
skip_special_tokens=True)
|
45 |
+
|
46 |
+
# Convert prompt to tokens
|
47 |
+
prompt_template = """\
|
48 |
+
<|im_start|>system
|
49 |
+
{system_message}<|im_end|>
|
50 |
+
<|im_start|>user
|
51 |
+
{prompt}<|im_end|>
|
52 |
+
<|im_start|>assistant"""
|
53 |
+
|
54 |
+
prompt = "You're standing on the surface of the Earth. "\
|
55 |
+
"You walk one mile south, one mile west and one mile north. "\
|
56 |
+
"You end up exactly where you started. Where are you?"
|
57 |
+
|
58 |
+
tokens = tokenizer(prompt_template.format(system_message=system_message,prompt=prompt),
|
59 |
+
return_tensors='pt').input_ids.cuda()
|
60 |
+
|
61 |
+
# Generate output
|
62 |
+
generation_output = model.generate(tokens,
|
63 |
+
streamer=streamer,
|
64 |
+
max_new_tokens=512)
|
65 |
+
```
|
66 |
+
|
67 |
### About AWQ
|
68 |
|
69 |
AWQ is an efficient, accurate and blazing-fast low-bit weight quantization method, currently supporting 4-bit quantization. Compared to GPTQ, it offers faster Transformers-based inference with equivalent or better quality compared to the most commonly used GPTQ settings.
|