|
--- |
|
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-3.1-8b-base |
|
library_name: transformers |
|
--- |
|
|
|
# granite-3.1-8b-base-FP8-dynamic |
|
|
|
## Model Overview |
|
- **Model Architecture:** granite-3.1-8b-base |
|
- **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-3.1-8b-base](https://huggingface.co/ibm-granite/granite-3.1-8b-base). |
|
It achieves an average score of 66.94 on the [OpenLLM](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) benchmark (version 1), whereas the unquantized model achieves 67.44. |
|
|
|
### Model Optimizations |
|
|
|
This model was obtained by quantizing the weights and activations of [ibm-granite/granite-3.1-8b-base](https://huggingface.co/ibm-granite/granite-3.1-8b-base) 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/granite-3.1-8b-base-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. |
|
|
|
<details> |
|
<summary>Model Creation Code</summary> |
|
|
|
```bash |
|
python quantize.py --model_id ibm-granite/granite-3.1-8b-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-3-8B-base")') |
|
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() |
|
``` |
|
</details> |
|
|
|
## Evaluation |
|
|
|
The model was evaluated on OpenLLM Leaderboard [V1](https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard), OpenLLM Leaderboard [V2](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/) and on [HumanEval](https://github.com/neuralmagic/evalplus), using the following commands: |
|
|
|
<details> |
|
<summary>Evaluation Commands</summary> |
|
|
|
OpenLLM Leaderboard V1: |
|
``` |
|
lm_eval \ |
|
--model vllm \ |
|
--model_args pretrained="neuralmagic/granite-3.1-8b-base-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/granite-3.1-8b-base-FP8-dynamic \ |
|
--bs 16 \ |
|
--temperature 0.2 \ |
|
--n_samples 50 \ |
|
--root "." \ |
|
--dataset humaneval |
|
``` |
|
##### Sanitization |
|
``` |
|
python3 evalplus/sanitize.py \ |
|
humaneval/neuralmagic--granite-3.1-8b-base-FP8-dynamic_vllm_temp_0.2 |
|
``` |
|
##### Evaluation |
|
``` |
|
evalplus.evaluate \ |
|
--dataset humaneval \ |
|
--samples humaneval/neuralmagic--granite-3.1-8b-base-FP8-dynamic_vllm_temp_0.2-sanitized |
|
``` |
|
</details> |
|
|
|
### Accuracy |
|
|
|
<table> |
|
<thead> |
|
<tr> |
|
<th>Category</th> |
|
<th>Metric</th> |
|
<th>ibm-granite/granite-3.1-8b-base</th> |
|
<th>neuralmagic/granite-3.1-8b-base-FP8-dynamic</th> |
|
<th>Recovery (%)</th> |
|
</tr> |
|
</thead> |
|
<tbody> |
|
<tr> |
|
<td rowspan="7"><b>OpenLLM V1</b></td> |
|
<td>ARC-Challenge (Acc-Norm, 25-shot)</td> |
|
<td>64.68</td> |
|
<td>64.16</td> |
|
<td>99.20</td> |
|
</tr> |
|
<tr> |
|
<td>GSM8K (Strict-Match, 5-shot)</td> |
|
<td>60.88</td> |
|
<td>58.45</td> |
|
<td>95.99</td> |
|
</tr> |
|
<tr> |
|
<td>HellaSwag (Acc-Norm, 10-shot)</td> |
|
<td>83.52</td> |
|
<td>83.46</td> |
|
<td>99.93</td> |
|
</tr> |
|
<tr> |
|
<td>MMLU (Acc, 5-shot)</td> |
|
<td>63.33</td> |
|
<td>63.35</td> |
|
<td>100.03</td> |
|
</tr> |
|
<tr> |
|
<td>TruthfulQA (MC2, 0-shot)</td> |
|
<td>51.33</td> |
|
<td>51.56</td> |
|
<td>100.45</td> |
|
</tr> |
|
<tr> |
|
<td>Winogrande (Acc, 5-shot)</td> |
|
<td>80.90</td> |
|
<td>80.66</td> |
|
<td>99.70</td> |
|
</tr> |
|
<tr> |
|
<td><b>Average Score</b></td> |
|
<td><b>67.44</b></td> |
|
<td><b>66.94</b></td> |
|
<td><b>99.26</b></td> |
|
</tr> |
|
<tr> |
|
<td rowspan="2"><b>Coding</b></td> |
|
<td>HumanEval Pass@1</td> |
|
<td>44.10</td> |
|
<td>44.80</td> |
|
<td><b>101.59</b></td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
|
|
|
|
|
|
## Inference Performance |
|
|
|
|
|
This model achieves up to 1.5x speedup in single-stream deployment and up to 1.1x speedup in multi-stream asynchronous 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). |
|
|
|
<details> |
|
<summary>Benchmarking Command</summary> |
|
|
|
``` |
|
guidellm --model neuralmagic/granite-3.1-8b-base-FP8-dynamic --target "http://localhost:8000/v1" --data-type emulated --data "prompt_tokens=<prompt_tokens>,generated_tokens=<generated_tokens>" --max seconds 360 --backend aiohttp_server |
|
``` |
|
|
|
</details> |
|
|
|
|
|
### 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-3.1-8b-base</td> |
|
<td></td> |
|
<td>25.1</td> |
|
<td>3.2</td> |
|
<td>25.3</td> |
|
<td>3.2</td> |
|
<td>3.2</td> |
|
<td>6.3</td> |
|
<td>13.4</td> |
|
</tr> |
|
<tr> |
|
<td>granite-3.1-8b-base-FP8-dynamic<br>(this model)</td> |
|
<td>1.47</td> |
|
<td>16.8</td> |
|
<td>2.2</td> |
|
<td>17.1</td> |
|
<td>2.2</td> |
|
<td>2.1</td> |
|
<td>4.2</td> |
|
<td>9.3</td> |
|
</tr> |
|
<tr> |
|
<td>granite-3.1-8b-base-quantized.w4a16</td> |
|
<td>2.72</td> |
|
<td>8.9</td> |
|
<td>1.2</td> |
|
<td>9.2</td> |
|
<td>1.2</td> |
|
<td>1.1</td> |
|
<td>2.3</td> |
|
<td>5.3</td> |
|
</tr> |
|
</table> |
|
|
|
|
|
### Multi-stream asynchronous performance (measured with vLLM version 0.6.6.post1) |
|
<table> |
|
<tr> |
|
<td></td> |
|
<td></td> |
|
<td></td> |
|
<th style="text-align: center;" colspan="7" >Maximum Throughput (Queries per Second)</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-3.1-8b-base</td> |
|
<td></td> |
|
<td>1.4</td> |
|
<td>7.8</td> |
|
<td>1.1</td> |
|
<td>6.2</td> |
|
<td>15.5</td> |
|
<td>6.0</td> |
|
<td>0.7</td> |
|
</tr> |
|
<tr> |
|
<td>granite-3.1-8b-base-FP8-dynamic<br>(this model)</td> |
|
<td>1.12</td> |
|
<td>2.1</td> |
|
<td>7.4</td> |
|
<td>1.3</td> |
|
<td>5.9</td> |
|
<td>15.3</td> |
|
<td>6.9</td> |
|
<td>0.8</td> |
|
</tr> |
|
<tr> |
|
<td>granite-3.1-2b-base-quantized.w4a16</td> |
|
<td>1.29</td> |
|
<td>2.4</td> |
|
<td>8.9</td> |
|
<td>1.4</td> |
|
<td>7.1</td> |
|
<td>17.8</td> |
|
<td>7.8</td> |
|
<td>1.0</td> |
|
</tr> |
|
</table> |
|
|
|
|
|
|
