File size: 2,744 Bytes
f482083
4dbabbc
 
 
 
 
 
 
f482083
4dbabbc
92d9539
4dbabbc
f482083
 
4dbabbc
 
 
 
 
 
 
f482083
4dbabbc
 
53a63a0
f482083
 
 
4dbabbc
 
 
 
 
 
f482083
4dbabbc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
06b1d52
4dbabbc
f482083
 
4dbabbc
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
---
base_model: NousResearch/Nous-Hermes-2-Yi-34B
inference: true
model_type: llama
quantized_by: mgoin
tags:
- nm-vllm
- sparse
---

## Nous-Hermes-2-Yi-34B-pruned2.4
This repo contains model files for [Nous Hermes 2 - Yi-34B](https://huggingface.co/NousResearch/Nous-Hermes-2-Yi-34B) optimized for [NM-vLLM](https://github.com/neuralmagic/nm-vllm), a high-throughput serving engine for compressed LLMs.

This model was pruned with [SparseGPT](https://arxiv.org/abs/2301.00774), using [SparseML](https://github.com/neuralmagic/sparseml).

## Inference
Install [NM-vLLM](https://github.com/neuralmagic/nm-vllm) for fast inference and low memory-usage: 
```bash
pip install nm-vllm[sparse]
```
Run in a Python pipeline for local inference:
```python
from vllm import LLM, SamplingParams

model = LLM("nm-testing/Nous-Hermes-2-Yi-34B-pruned2.4", sparsity="semi_structured_sparse_w16a16")
prompt = "How to make banana bread?"
formatted_prompt =  f"<|im_start|>User:{prompt}\n<|im_start|>assistant:\n"

sampling_params = SamplingParams(max_tokens=100, temperature=0)
outputs = model.generate(formatted_prompt, sampling_params=sampling_params)
print(outputs[0].outputs[0].text)
"""
To make banana bread, follow these steps:
1. Gather the ingredients:
- 2 ripe bananas
- 2 cups of flour
- 1 teaspoon of baking powder
- 1 teaspoon of salt
- 1 teaspoon of sugar
- 1 teaspoon of vanilla extract
2. Preheat the oven to 350°F.
3. In a mixing bowl, combine the flour, baking powder, salt, sugar, and vanilla extract.
4.
"""
```

## Prompt template

```
<|im_start|>User:{prompt}\n<|im_start|>assistant:\n
```

## Sparsification
For details on how this model was sparsified, see the `recipe.yaml` in this repo and follow the instructions below.

Install [SparseML](https://github.com/neuralmagic/sparseml):
```bash
git clone https://github.com/neuralmagic/sparseml
pip install -e "sparseml[transformers]"
```

Replace the recipe as you like and run this one-shot compression script to apply SparseGPT:
```python
import sparseml.transformers

original_model_name = "NousResearch/Nous-Hermes-2-Yi-34B"
calibration_dataset = "open_platypus"
output_directory = "output/"

recipe = """
test_stage:
  obcq_modifiers:
    SparseGPTModifier:
      sparsity: 0.5
      sequential_update: true
      mask_structure: '2:4'
      targets: ['re:model.layers.\d*$']
"""

# Apply SparseGPT to the model
sparseml.transformers.oneshot(
    model=original_model_name,
    dataset=calibration_dataset,
    recipe=recipe,
    output_dir=output_directory,
)
```

## Slack

For further support, and discussions on these models and AI in general, join [Neural Magic's Slack Community](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ)