Hugging Face's logo

neuralmagic
/
DeepSeek-R1-Distill-Llama-8B-quantized.w8a8

Text Generation
Transformers
Safetensors
llama
deepseek
int8
vllm
llmcompressor
conversational
text-generation-inference
Inference Endpoints
8-bit precision
compressed-tensors
Model card Files Community
DeepSeek-R1-Distill-Llama-8B-quantized.w8a8 / README.md
nm-research's picture
nm-research
Update README.md
f8cfa56 verified
preview code
|
raw
history blame
19.6 kB
---
license: mit
tags:
- deepseek
- int8
- vllm
- llmcompressor
base_model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
library_name: transformers
---
# DeepSeek-R1-Distill-Llama-8B-quantized.w8a8
## Model Overview
- **Model Architecture:** LlamaForCausalLM
- **Input:** Text
- **Output:** Text
- **Model Optimizations:**
- **Weight quantization:** INT8
- **Activation quantization:** INT8
- **Release Date:** 2/1/2025
- **Version:** 1.0
- **Model Developers:** Neural Magic
Quantized version of [DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B).
### Model Optimizations
This model was obtained by quantizing the weights and activations of [DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) to INT8 data type.
This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x).
Weight quantization also reduces disk size requirements by approximately 50%.
Only the weights and activations of the linear operators within transformers blocks are quantized.
Weights are quantized using a symmetric per-channel scheme, whereas quantizations are quantized using a symmetric per-token scheme.
The [GPTQ](https://arxiv.org/abs/2210.17323) algorithm is applied for quantization, as implemented in the [llm-compressor](https://github.com/vllm-project/llm-compressor) library.
## 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
number_gpus = 1
model_name = "neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8"
tokenizer = AutoTokenizer.from_pretrained(model_name)
sampling_params = SamplingParams(temperature=0.6, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])
llm = LLM(model=model_name, tensor_parallel_size=number_gpus, trust_remote_code=True)
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.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
from llmcompressor.transformers import oneshot
# Load model
model_stub = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
model_name = model_stub.split("/")[-1]
num_samples = 1024
max_seq_len = 8192
tokenizer = AutoTokenizer.from_pretrained(model_stub)
model = AutoModelForCausalLM.from_pretrained(
model_stub,
device_map=device_map,
torch_dtype="auto",
)
def preprocess_fn(example):
return {"text": tokenizer.apply_chat_template(example["messages"], add_generation_prompt=False, tokenize=False)}
ds = load_dataset("neuralmagic/LLM_compression_calibration", split="train")
ds = ds.map(preprocess_fn)
# Configure the quantization algorithm and scheme
recipe = [
SmoothQuantModifier(smoothing_strength=0.8),
QuantizationModifier(
targets="Linear",
scheme="W8A8",
ignore=["lm_head"],
dampening_frac=0.1,
),
]
# Apply quantization
oneshot(
model=model,
dataset=ds,
recipe=recipe,
max_seq_length=max_seq_len,
num_calibration_samples=num_samples,
)
# Save to disk in compressed-tensors format
save_path = model_name + "-quantized.w8a8
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")
```
## Evaluation
The model was evaluated on OpenLLM Leaderboard [V1](https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard) and [V2](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/), using the following commands:
OpenLLM Leaderboard V1:
```
lm_eval \
--model vllm \
--model_args pretrained="neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8",dtype=auto,max_model_len=4096,tensor_parallel_size=1,enable_chunked_prefill=True \
--tasks openllm \
--write_out \
--batch_size auto \
--output_path output_dir \
--show_config
```
OpenLLM Leaderboard V2:
```
lm_eval \
--model vllm \
--model_args pretrained="neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8",dtype=auto,max_model_len=4096,tensor_parallel_size=1,enable_chunked_prefill=True \
--apply_chat_template \
--fewshot_as_multiturn \
--tasks leaderboard \
--write_out \
--batch_size auto \
--output_path output_dir \
--show_config
```
### Accuracy
<table>
<thead>
<tr>
<th>Category</th>
<th>Metric</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</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>45.05</td>
<td>45.22</td>
<td>100.4%</td>
</tr>
<tr>
<td>GSM8K (Strict-Match, 5-shot)</td>
<td>62.77</td>
<td>62.09</td>
<td>98.9%</td>
</tr>
<tr>
<td>HellaSwag (Acc-Norm, 10-shot)</td>
<td>76.78</td>
<td>76.80</td>
<td>100.0%</td>
</tr>
<tr>
<td>MMLU (Acc, 5-shot)</td>
<td>55.65</td>
<td>55.53</td>
<td>99.8%</td>
</tr>
<tr>
<td>TruthfulQA (MC2, 0-shot)</td>
<td>50.55</td>
<td>49.89</td>
<td>98.7%</td>
</tr>
<tr>
<td>Winogrande (Acc, 5-shot)</td>
<td>68.51</td>
<td>67.40</td>
<td>98.4%</td>
</tr>
<tr>
<td><b>Average Score</b></td>
<td><b>59.88</b></td>
<td><b>59.49</b></td>
<td><b>99.3%</b></td>
</tr>
<tr>
<td rowspan="7"><b>OpenLLM V2</b></td>
<td>IFEval (Inst Level Strict Acc, 0-shot)</td>
<td>38.34</td>
<td>39.07</td>
<td>101.9%</td>
</tr>
<tr>
<td>BBH (Acc-Norm, 3-shot)</td>
<td>38.19</td>
<td>39.57</td>
<td>103.6%</td>
</tr>
<tr>
<td>Math-Hard (Exact-Match, 4-shot)</td>
<td>0.00</td>
<td>0.00</td>
<td>---</td>
</tr>
<tr>
<td>GPQA (Acc-Norm, 0-shot)</td>
<td>28.87</td>
<td>27.28</td>
<td>94.5%</td>
</tr>
<tr>
<td>MUSR (Acc-Norm, 0-shot)</td>
<td>33.31</td>
<td>34.50</td>
<td>103.6%</td>
</tr>
<tr>
<td>MMLU-Pro (Acc, 5-shot)</td>
<td>20.10</td>
<td>20.60</td>
<td>102.4%</td>
</tr>
<tr>
<td><b>Average Score</b></td>
<td><b>26.47</b></td>
<td><b>26.84</b></td>
<td><b>101.4%</b></td>
</tr>
<tr>
<td rowspan="4"><b>Coding</b></td>
<td>HumanEval (pass@1)</td>
<td>49.90</td>
<td>50.90</td>
<td><b>102.0%</b></td>
</tr>
<tr>
<td>HumanEval (pass@10)</td>
<td>68.90</td>
<td>68.70</td>
<td>99.7%</td>
</tr>
<tr>
<td>HumanEval+ (pass@10)</td>
<td>44.10</td>
<td>46.70</td>
<td>105.9%</td>
</tr>
<tr>
<td>HumanEval+ (pass@10)</td>
<td>62.90</td>
<td>64.30</td>
<td>102.2%</td>
</tr>
</tbody>
</table>
## Inference Performance
This model achieves up to 1.6x speedup in single-stream deployment and up to 1.4x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario.
The following performance benchmarks were conducted with [vLLM](https://docs.vllm.ai/en/latest/) version 0.7.2, and [GuideLLM](https://github.com/neuralmagic/guidellm).
<details>
<summary>Benchmarking Command</summary>
```
guidellm --model neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8 --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 0.7.2)
<table>
<thead>
<tr>
<th></th>
<th></th>
<th></th>
<th style="text-align: center;" colspan="2" >Instruction Following<br>256 / 128</th>
<th style="text-align: center;" colspan="2" >Multi-turn Chat<br>512 / 256</th>
<th style="text-align: center;" colspan="2" >Docstring Generation<br>768 / 128</th>
<th style="text-align: center;" colspan="2" >RAG<br>1024 / 128</th>
<th style="text-align: center;" colspan="2" >Code Completion<br>256 / 1024</th>
<th style="text-align: center;" colspan="2" >Code Fixing<br>1024 / 1024</th>
<th style="text-align: center;" colspan="2" >Large Summarization<br>4096 / 512</th>
<th style="text-align: center;" colspan="2" >Large RAG<br>10240 / 1536</th>
</tr>
<tr>
<th>Hardware</th>
<th>Model</th>
<th>Average cost reduction</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
<th>Latency (s)</th>
<th>QPD</th>
</tr>
</thead>
<tbody style="text-align: center" >
<tr>
<th rowspan="3" valign="top">A6000x1</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<td>---</td>
<td>3.0</td>
<td>1511</td>
<td>6.0</td>
<td>755</td>
<td>3.0</td>
<td>1483</td>
<td>3.1</td>
<td>1462</td>
<td>23.6</td>
<td>191</td>
<td>24.0</td>
<td>188</td>
<td>12.7</td>
<td>353</td>
<td>41.1</td>
<td>110</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
<td>1.53</td>
<td>1.9</td>
<td>2356</td>
<td>3.8</td>
<td>1175</td>
<td>2.0</td>
<td>2291</td>
<td>2.0</td>
<td>2207</td>
<td>15.2</td>
<td>297</td>
<td>15.5</td>
<td>290</td>
<td>8.5</td>
<td>531</td>
<td>28.6</td>
<td>157</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
<td>2.35</td>
<td>1.2</td>
<td>3870</td>
<td>2.3</td>
<td>1918</td>
<td>1.3</td>
<td>3492</td>
<td>1.3</td>
<td>3335</td>
<td>9.1</td>
<td>492</td>
<td>9.5</td>
<td>472</td>
<td>5.8</td>
<td>771</td>
<td>22.7</td>
<td>198</td>
</tr>
<tr>
<th rowspan="3" valign="top">A100x1</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<td>---</td>
<td>1.5</td>
<td>1308</td>
<td>3.1</td>
<td>657</td>
<td>1.6</td>
<td>1274</td>
<td>1.6</td>
<td>1263</td>
<td>12.1</td>
<td>166</td>
<td>12.4</td>
<td>162</td>
<td>6.5</td>
<td>308</td>
<td>25.6</td>
<td>78</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
<td>1.30</td>
<td>1.1</td>
<td>1763</td>
<td>2.3</td>
<td>882</td>
<td>1.2</td>
<td>1716</td>
<td>1.2</td>
<td>1698</td>
<td>9.0</td>
<td>223</td>
<td>9.2</td>
<td>218</td>
<td>4.9</td>
<td>409</td>
<td>25.7</td>
<td>78</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
<td>1.76</td>
<td>0.8</td>
<td>2501</td>
<td>1.6</td>
<td>1236</td>
<td>0.9</td>
<td>2350</td>
<td>0.9</td>
<td>2287</td>
<td>6.4</td>
<td>316</td>
<td>6.6</td>
<td>306</td>
<td>3.7</td>
<td>544</td>
<td>24.7</td>
<td>82</td>
</tr>
<tr>
<th rowspan="3" valign="top">H100x1</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<td>---</td>
<td>1.0</td>
<td>1146</td>
<td>1.9</td>
<td>574</td>
<td>1.0</td>
<td>1128</td>
<td>1.0</td>
<td>1111</td>
<td>7.6</td>
<td>144</td>
<td>7.7</td>
<td>142</td>
<td>4.1</td>
<td>266</td>
<td>16.3</td>
<td>67</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-FP8-dynamic</th>
<td>1.25</td>
<td>0.7</td>
<td>1567</td>
<td>1.4</td>
<td>758</td>
<td>0.7</td>
<td>1484</td>
<td>0.7</td>
<td>1462</td>
<td>5.7</td>
<td>191</td>
<td>5.8</td>
<td>189</td>
<td>3.2</td>
<td>347</td>
<td>22.5</td>
<td>49</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
<td>1.30</td>
<td>0.7</td>
<td>1527</td>
<td>1.4</td>
<td>768</td>
<td>0.7</td>
<td>1495</td>
<td>0.7</td>
<td>1463</td>
<td>5.6</td>
<td>194</td>
<td>5.7</td>
<td>190</td>
<td>3.1</td>
<td>350</td>
<td>14.7</td>
<td>74</td>
</tr>
</tbody>
</table>
**Use case profiles: prompt tokens / generation tokens
**QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).
### Multi-stream asynchronous performance (measured with vLLM version 0.7.2)
<table>
<thead>
<tr>
<th></th>
<th></th>
<th></th>
<th style="text-align: center;" colspan="2" >Instruction Following<br>256 / 128</th>
<th style="text-align: center;" colspan="2" >Multi-turn Chat<br>512 / 256</th>
<th style="text-align: center;" colspan="2" >Docstring Generation<br>768 / 128</th>
<th style="text-align: center;" colspan="2" >RAG<br>1024 / 128</th>
<th style="text-align: center;" colspan="2" >Code Completion<br>256 / 1024</th>
<th style="text-align: center;" colspan="2" >Code Fixing<br>1024 / 1024</th>
<th style="text-align: center;" colspan="2" >Large Summarization<br>4096 / 512</th>
<th style="text-align: center;" colspan="2" >Large RAG<br>10240 / 1536</th>
</tr>
<tr>
<th>Hardware</th>
<th>Model</th>
<th>Average cost reduction</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
<th>Maximum throughput (QPS)</th>
<th>QPD</th>
</tr>
</thead>
<tbody style="text-align: center" >
<tr>
<th rowspan="3" valign="top">A6000x1</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<td>---</td>
<td>12.6</td>
<td>56742</td>
<td>5.7</td>
<td>25687</td>
<td>6.5</td>
<td>29349</td>
<td>5.2</td>
<td>23259</td>
<td>1.6</td>
<td>7250</td>
<td>1.2</td>
<td>5181</td>
<td>0.8</td>
<td>3445</td>
<td>0.1</td>
<td>616</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
<td>1.34</td>
<td>17.4</td>
<td>78101</td>
<td>7.6</td>
<td>34351</td>
<td>8.8</td>
<td>39790</td>
<td>7.0</td>
<td>31532</td>
<td>2.3</td>
<td>10405</td>
<td>1.5</td>
<td>6960</td>
<td>1.0</td>
<td>4355</td>
<td>0.2</td>
<td>785</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
<td>0.91</td>
<td>10.9</td>
<td>48964</td>
<td>5.1</td>
<td>22989</td>
<td>4.8</td>
<td>21791</td>
<td>3.8</td>
<td>17039</td>
<td>2.2</td>
<td>9726</td>
<td>1.2</td>
<td>5434</td>
<td>0.6</td>
<td>2544</td>
<td>0.1</td>
<td>578</td>
</tr>
<tr>
<th rowspan="3" valign="top">A100x1</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<td>---</td>
<td>24.5</td>
<td>49296</td>
<td>11.3</td>
<td>22657</td>
<td>13.0</td>
<td>26047</td>
<td>10.5</td>
<td>21020</td>
<td>3.5</td>
<td>7029</td>
<td>2.5</td>
<td>4995</td>
<td>1.7</td>
<td>3503</td>
<td>0.3</td>
<td>659</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
<td>1.27</td>
<td>30.8</td>
<td>62042</td>
<td>14.1</td>
<td>28419</td>
<td>17.2</td>
<td>34554</td>
<td>13.8</td>
<td>27719</td>
<td>4.6</td>
<td>9299</td>
<td>3.1</td>
<td>6215</td>
<td>2.2</td>
<td>4331</td>
<td>0.4</td>
<td>807</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
<td>0.97</td>
<td>22.7</td>
<td>45708</td>
<td>10.5</td>
<td>21216</td>
<td>11.1</td>
<td>22353</td>
<td>8.9</td>
<td>17939</td>
<td>3.9</td>
<td>7758</td>
<td>2.6</td>
<td>5241</td>
<td>1.6</td>
<td>3196</td>
<td>0.4</td>
<td>718</td>
</tr>
<tr>
<th rowspan="3" valign="top">H100x1</th>
<th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
<td>---</td>
<td>49.0</td>
<td>53593</td>
<td>22.6</td>
<td>24750</td>
<td>28.3</td>
<td>30971</td>
<td>22.9</td>
<td>25035</td>
<td>7.2</td>
<td>7912</td>
<td>5.1</td>
<td>5561</td>
<td>3.6</td>
<td>3939</td>
<td>0.6</td>
<td>703</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-FP8-dynamic</th>
<td>1.14</td>
<td>57.1</td>
<td>62517</td>
<td>26.0</td>
<td>28440</td>
<td>34.5</td>
<td>37781</td>
<td>28.7</td>
<td>31360</td>
<td>7.2</td>
<td>7877</td>
<td>5.4</td>
<td>5923</td>
<td>4.3</td>
<td>4697</td>
<td>0.7</td>
<td>782</td>
</tr>
<tr>
<th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
<td>1.01</td>
<td>49.8</td>
<td>54452</td>
<td>22.9</td>
<td>25035</td>
<td>28.5</td>
<td>31162</td>
<td>23.0</td>
<td>25200</td>
<td>6.8</td>
<td>7493</td>
<td>5.0</td>
<td>5431</td>
<td>3.7</td>
<td>4079</td>
<td>0.7</td>
<td>787</td>
</tr>
</tbody>
</table>
**Use case profiles: prompt tokens / generation tokens
**QPS: Queries per second.
**QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).