File size: 10,242 Bytes
d53b1b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6043ef6
d53b1b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c8b4ee2
 
 
 
d53b1b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b758978
d53b1b3
 
 
 
 
 
 
 
 
 
b758978
 
 
 
 
 
 
 
 
492f5db
b758978
d53b1b3
b758978
d53b1b3
 
 
 
6043ef6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
---
tags:
- fp8
- vllm
license: apache-2.0
license_link: https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md
language:
  - en
base_model: ibm-granite/granite-3.1-2b-instruct
library_name: transformers
---

# granite-3.1-2b-instruct-FP8-dynamic

## Model Overview
- **Model Architecture:** granite-3.1-2b-instruct
  - **Input:** Text
  - **Output:** Text
- **Model Optimizations:**
  - **Weight quantization:** FP8
  - **Activation quantization:** FP8
- **Release Date:** 1/8/2025
- **Version:** 1.0
- **Model Developers:** Neural Magic

Quantized version of [ibm-granite/granite-3.1-2b-instruct](https://huggingface.co/ibm-granite/granite-3.1-2b-instruct).
It achieves an average score of 61.84 on the [OpenLLM](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) benchmark (version 1), whereas the unquantized model achieves 61.98.

### Model Optimizations

This model was obtained by quantizing the weights and activations of [ibm-granite/granite-3.1-2b-instruct](https://huggingface.co/ibm-granite/granite-3.1-2b-instruct) to FP8 data type, ready for inference with vLLM >= 0.5.2.
This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. Only the weights and activations of the linear operators within transformers blocks are quantized. 

## Deployment

### Use with vLLM

This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below.

```python
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

max_model_len, tp_size = 4096, 1
model_name = "neuralmagic-ent/granite-3.1-2b-instruct-FP8-dynamic"
tokenizer = AutoTokenizer.from_pretrained(model_name)
llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True)
sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])

messages_list = [
    [{"role": "user", "content": "Who are you? Please respond in pirate speak!"}],
]

prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list]

outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)

generated_text = [output.outputs[0].text for output in outputs]
print(generated_text)
```

vLLM also supports OpenAI-compatible serving. See the [documentation](https://docs.vllm.ai/en/latest/) for more details.

## Creation

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below. 


```bash
python quantize.py --model_id ibm-granite/granite-3.1-2b-base --save_path "output_dir/"
```

```python
import argparse
from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot
import os

def main():
    parser = argparse.ArgumentParser(description='Quantize a transformer model to FP8')
    parser.add_argument('--model_id', type=str, required=True,
                        help='The model ID from HuggingFace (e.g., "meta-llama/Meta-Llama-3-2b-Instruct")')
    parser.add_argument('--save_path', type=str, default='.',
                        help='Custom path to save the quantized model. If not provided, will use model_name-FP8-dynamic')
    args = parser.parse_args()

    # Load model
    model = AutoModelForCausalLM.from_pretrained(
        args.model_id, device_map="auto", torch_dtype="auto", trust_remote_code=True,
    )
    tokenizer = AutoTokenizer.from_pretrained(args.model_id)

    # Configure the quantization algorithm and scheme
    recipe = QuantizationModifier(
        targets="Linear", scheme="FP8_DYNAMIC", ignore=["lm_head"]
    )

    # Apply quantization
    oneshot(model=model, recipe=recipe)

    save_path = os.path.join(args.save_path, args.model_id.split("/")[1] + "-FP8-dynamic")
    os.makedirs(save_path, exist_ok=True)

    # Save to disk in compressed-tensors format
    model.save_pretrained(save_path)
    tokenizer.save_pretrained(save_path)
    print(f"Model and tokenizer saved to: {save_path}")

if __name__ == "__main__":
    main()
```

## Evaluation

The model was evaluated on OpenLLM Leaderboard [V1](https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard) and on [HumanEval](https://github.com/neuralmagic/evalplus), using the following commands:

OpenLLM Leaderboard V1:
```
lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/granite-3.1-2b-instruct-FP8-dynamic",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1,gpu_memory_utilization=0.8,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks openllm \
  --write_out \
  --batch_size auto \
  --output_path output_dir \
  --show_config
```

#### HumanEval
##### Generation
```
python3 codegen/generate.py \
  --model neuralmagic-ent/granite-3.1-2b-instruct-FP8-dynamic \
  --bs 16 \
  --temperature 0.2 \
  --n_samples 50 \
  --root "." \
  --dataset humaneval
```
##### Sanitization
```
python3 evalplus/sanitize.py \
  humaneval/neuralmagic-ent--granite-3.1-2b-instruct-FP8-dynamic_vllm_temp_0.2
```
##### Evaluation
```
evalplus.evaluate \
  --dataset humaneval \
  --samples humaneval/neuralmagic-ent--granite-3.1-2b-instruct-FP8-dynamic_vllm_temp_0.2-sanitized
```

### Accuracy

#### OpenLLM Leaderboard V1 evaluation scores

| Metric                                  | ibm-granite/granite-3.1-2b-instruct             | neuralmagic-ent/granite-3.1-2b-instruct-FP8-dynamic |
|-----------------------------------------|:---------------------------------:|:-------------------------------------------:|
| ARC-Challenge (Acc-Norm, 25-shot)       | 55.63                             | 55.03                                      |
| GSM8K (Strict-Match, 5-shot)            | 60.96                             | 61.49                                       |
| HellaSwag (Acc-Norm, 10-shot)           | 75.21                             | 75.26                                       |
| MMLU (Acc, 5-shot)                      | 54.38                             | 54.24                                        |
| TruthfulQA (MC2, 0-shot)                | 55.93                             | 55.42                                        |
| Winogrande (Acc, 5-shot)                | 69.67                             | 69.61                                        |
| **Average Score**                       | **61.98**                         | **61.84**                                   |
| **Recovery**                            | **100.00**                        | **99.78**                                   |

| Metric                                  | ibm-granite/granite-3.1-2b-instruct             | neuralmagic-ent/granite-3.1-2b-instruct-FP8-dynamic |
|-----------------------------------------|:---------------------------------:|:-------------------------------------------:|
| IFEval (Inst Level Strict Acc, 0-shot)| 67.99                           | 66.79                                         |
| BBH (Acc-Norm, 3-shot)            | 44.11                             | 44.24                                        |
| Math-Hard (Exact-Match, 4-shot)   | 8.66                            | 7.89                                       |
| GPQA (Acc-Norm, 0-shot)           | 28.30                             | 26.90                                        |
| MUSR (Acc-Norm, 0-shot)           | 35.12                             | 35.12                                         |
| MMLU-Pro (Acc, 5-shot)            | 26.87                             | 28.33                                        |
| **Average Score**                 | **35.17**                         | **34.88**                                    |
| **Recovery**                      | **100.00**                         | **99.16**                                    |

#### HumanEval pass@1 scores
| Metric                                  | ibm-granite/granite-3.1-2b-instruct             | neuralmagic-ent/granite-3.1-2b-instruct-FP8-dynamic |
|-----------------------------------------|:---------------------------------:|:-------------------------------------------:|
| HumanEval Pass@1                        | 53.40                             |  54.90                                     |


## Inference Performance


This model achieves up to 1.2x speedup in single-stream deployment on L40 GPUs.
The following performance benchmarks were conducted with [vLLM](https://docs.vllm.ai/en/latest/) version 0.6.6.post1, and [GuideLLM](https://github.com/neuralmagic/guidellm).

### Single-stream performance (measured with vLLM version 0.6.6.post1)
<table>
  <tr>
    <td></td>
    <td></td>
    <td></td>
    <th style="text-align: center;" colspan="7" >Latency (s)</th>
  </tr>
  <tr>
    <th>GPU class</th>
    <th>Model</th>
    <th>Speedup</th>
    <th>Code Completion<br>prefill: 256 tokens<br>decode: 1024 tokens</th>
    <th>Docstring Generation<br>prefill: 768 tokens<br>decode: 128 tokens</th>
    <th>Code Fixing<br>prefill: 1024 tokens<br>decode: 1024 tokens</th>
    <th>RAG<br>prefill: 1024 tokens<br>decode: 128 tokens</th>
    <th>Instruction Following<br>prefill: 256 tokens<br>decode: 128 tokens</th>
    <th>Multi-turn Chat<br>prefill: 512 tokens<br>decode: 256 tokens</th>
    <th>Large Summarization<br>prefill: 4096 tokens<br>decode: 512 tokens</th>
  </tr>
  <tr>
    <td style="vertical-align: middle;" rowspan="3" >L40</td>
    <td>granite-3.1-2b-instruct</td>
    <td></td>
    <td>9.3</td>
    <td>1.2</td>
    <td>9.4</td>
    <td>1.2</td>
    <td>1.2</td>
    <td>2.3</td>
    <td>5.0</td>
  </tr>
  <tr>
    <td>granite-3.1-2b-instruct-FP8-dynamic<br>(this model)</td>
    <td>1.26</td>
    <td>7.3</td>
    <td>0.9</td>
    <td>7.4</td>
    <td>1.0</td>
    <td>0.9</td>
    <td>1.8</td>
    <td>4.1</td>
  </tr>
  <tr>
    <td>granite-3.1-2b-instruct-quantized.w4a16</td>
    <td>1.88</td>
    <td>4.8</td>
    <td>0.6</td>
    <td>4.9</td>
    <td>0.6</td>
    <td>0.6</td>
    <td>1.2</td>
    <td>2.8</td>
  </tr>
</table>