|
--- |
|
license: mit |
|
tags: |
|
- deepseek |
|
- int8 |
|
- vllm |
|
- llmcompressor |
|
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B |
|
library_name: transformers |
|
--- |
|
|
|
# DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8 |
|
|
|
## Model Overview |
|
- **Model Architecture:** Qwen2ForCausalLM |
|
- **Input:** Text |
|
- **Output:** Text |
|
- **Model Optimizations:** |
|
- **Weight quantization:** INT8 |
|
- **Activation quantization:** INT8 |
|
- **Release Date:** 2/5/2025 |
|
- **Version:** 1.0 |
|
- **Model Developers:** Neural Magic |
|
|
|
Quantized version of [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B). |
|
|
|
|
|
### Model Optimizations |
|
|
|
This model was obtained by quantizing the weights and activations of [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) 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-Qwen-1.5B-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-Qwen-1.5B" |
|
model_name = model_stub.split("/")[-1] |
|
|
|
num_samples = 2048 |
|
max_seq_len = 8192 |
|
|
|
tokenizer = AutoTokenizer.from_pretrained(model_stub) |
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
model_stub, |
|
device_map="auto", |
|
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=0.9), |
|
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-Qwen-1.5B-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-Qwen-1.5B-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-Qwen-1.5B</th> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-FP8-dynamic</th> |
|
<th>Recovery</th> |
|
</tr> |
|
</thead> |
|
<tbody> |
|
<tr> |
|
<td rowspan="4"><b>Reasoning</b></td> |
|
<td>AIME 2024 (pass@1)</td> |
|
<td>30.05</td> |
|
<td>26.67</td> |
|
<td>88.75%</td> |
|
</tr> |
|
<tr> |
|
<td>MATH-500 (pass@1)</td> |
|
<td>84.66</td> |
|
<td>84.39</td> |
|
<td>99.68%</td> |
|
</tr> |
|
<tr> |
|
<td>GPQA Diamond (pass@1)</td> |
|
<td>35.37</td> |
|
<td>34.43</td> |
|
<td>97.34%</td> |
|
</tr> |
|
<tr> |
|
<td><b>Average Score</b></td> |
|
<td><b>50.03</b></td> |
|
<td><b>48.5</b></td> |
|
<td><b>96.94%</b></td> |
|
</tr> |
|
<tr> |
|
<td rowspan="7"><b>OpenLLM V1</b></td> |
|
<td>ARC-Challenge (Acc-Norm, 25-shot)</td> |
|
<td>37.20</td> |
|
<td>37.03</td> |
|
<td>99.5%</td> |
|
</tr> |
|
<tr> |
|
<td>GSM8K (Strict-Match, 5-shot)</td> |
|
<td>69.98</td> |
|
<td>68.46</td> |
|
<td>97.8%</td> |
|
</tr> |
|
<tr> |
|
<td>HellaSwag (Acc-Norm, 10-shot)</td> |
|
<td>43.86</td> |
|
<td>43.91</td> |
|
<td>100.1%</td> |
|
</tr> |
|
<tr> |
|
<td>MMLU (Acc, 5-shot)</td> |
|
<td>37.38</td> |
|
<td>37.20</td> |
|
<td>99.5%</td> |
|
</tr> |
|
<tr> |
|
<td>TruthfulQA (MC2, 0-shot)</td> |
|
<td>45.21</td> |
|
<td>45.59</td> |
|
<td>100.8%</td> |
|
</tr> |
|
<tr> |
|
<td>Winogrande (Acc, 5-shot)</td> |
|
<td>54.30</td> |
|
<td>54.30</td> |
|
<td>100.0%</td> |
|
</tr> |
|
<tr> |
|
<td><b>Average Score</b></td> |
|
<td><b>47.99</b></td> |
|
<td><b>47.75</b></td> |
|
<td><b>99.5%</b></td> |
|
</tr> |
|
<tr> |
|
<td rowspan="7"><b>OpenLLM V2</b></td> |
|
<td>IFEval (Inst Level Strict Acc, 0-shot)</td> |
|
<td>34.63</td> |
|
<td>35.70</td> |
|
<td>103.1%</td> |
|
</tr> |
|
<tr> |
|
<td>BBH (Acc-Norm, 3-shot)</td> |
|
<td>3.06</td> |
|
<td>2.31</td> |
|
<td>---</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>1.01</td> |
|
<td>1.19</td> |
|
<td>---</td> |
|
</tr> |
|
<tr> |
|
<td>MUSR (Acc-Norm, 0-shot)</td> |
|
<td>0.78</td> |
|
<td>0.29</td> |
|
<td>---</td> |
|
</tr> |
|
<tr> |
|
<td>MMLU-Pro (Acc, 5-shot)</td> |
|
<td>1.32</td> |
|
<td>1.23</td> |
|
<td>---</td> |
|
</tr> |
|
<tr> |
|
<td><b>Average Score</b></td> |
|
<td><b>6.80</b></td> |
|
<td><b>6.79</b></td> |
|
<td><b>---</b></td> |
|
</tr> |
|
<tr> |
|
<td rowspan="4"><b>Coding</b></td> |
|
<td>HumanEval (pass@1)</td> |
|
<td>37.90</td> |
|
<td>36.20</td> |
|
<td><b>95.5%</b></td> |
|
</tr> |
|
<tr> |
|
<td>HumanEval (pass@10)</td> |
|
<td>61.30</td> |
|
<td>63.10</td> |
|
<td>102.9%</td> |
|
</tr> |
|
<tr> |
|
<td>HumanEval+ (pass@10)</td> |
|
<td>33.00</td> |
|
<td>32.70</td> |
|
<td>99.1%</td> |
|
</tr> |
|
<tr> |
|
<td>HumanEval+ (pass@10)</td> |
|
<td>55.90</td> |
|
<td>58.40</td> |
|
<td>104.5%</td> |
|
</tr> |
|
</tbody> |
|
</table> |
|
|
|
|
|
## Inference Performance |
|
|
|
|
|
This model achieves up to 1.2x speedup in single-stream 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-Qwen-1.5B-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-Qwen-1.5B</th> |
|
<td>---</td> |
|
<td>0.8</td> |
|
<td>5667</td> |
|
<td>1.6</td> |
|
<td>2776</td> |
|
<td>0.8</td> |
|
<td>5515</td> |
|
<td>0.8</td> |
|
<td>5466</td> |
|
<td>6.4</td> |
|
<td>705</td> |
|
<td>6.5</td> |
|
<td>697</td> |
|
<td>3.5</td> |
|
<td>1295</td> |
|
<td>18.3</td> |
|
<td>246</td> |
|
</tr> |
|
<tr> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8</th> |
|
<td>1.14</td> |
|
<td>0.7</td> |
|
<td>6635</td> |
|
<td>1.3</td> |
|
<td>3340</td> |
|
<td>0.7</td> |
|
<td>6396</td> |
|
<td>0.7</td> |
|
<td>6343</td> |
|
<td>5.3</td> |
|
<td>845</td> |
|
<td>5.4</td> |
|
<td>832</td> |
|
<td>2.9</td> |
|
<td>1547</td> |
|
<td>21.3</td> |
|
<td>211</td> |
|
</tr> |
|
<tr> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16</th> |
|
<td>1.38</td> |
|
<td>0.5</td> |
|
<td>8293</td> |
|
<td>1.1</td> |
|
<td>4184</td> |
|
<td>0.6</td> |
|
<td>7976</td> |
|
<td>0.6</td> |
|
<td>7504</td> |
|
<td>4.3</td> |
|
<td>1051</td> |
|
<td>4.4</td> |
|
<td>1033</td> |
|
<td>2.5</td> |
|
<td>1819</td> |
|
<td>21.1</td> |
|
<td>213</td> |
|
</tr> |
|
<tr> |
|
<th rowspan="3" valign="top">A100x1</th> |
|
<th>deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B</th> |
|
<td>---</td> |
|
<td>0.6</td> |
|
<td>3359</td> |
|
<td>1.2</td> |
|
<td>1654</td> |
|
<td>0.6</td> |
|
<td>3286</td> |
|
<td>0.6</td> |
|
<td>3241</td> |
|
<td>4.7</td> |
|
<td>424</td> |
|
<td>4.9</td> |
|
<td>411</td> |
|
<td>2.6</td> |
|
<td>778</td> |
|
<td>21.1</td> |
|
<td>95</td> |
|
</tr> |
|
<tr> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8</th> |
|
<td>1.05</td> |
|
<td>0.6</td> |
|
<td>3531</td> |
|
<td>1.1</td> |
|
<td>1807</td> |
|
<td>0.6</td> |
|
<td>3427</td> |
|
<td>0.6</td> |
|
<td>3480</td> |
|
<td>4.5</td> |
|
<td>448</td> |
|
<td>4.5</td> |
|
<td>447</td> |
|
<td>2.4</td> |
|
<td>842</td> |
|
<td>23.5</td> |
|
<td>86</td> |
|
</tr> |
|
<tr> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16</th> |
|
<td>1.03</td> |
|
<td>0.6</td> |
|
<td>3469</td> |
|
<td>1.1</td> |
|
<td>1751</td> |
|
<td>0.6</td> |
|
<td>3403</td> |
|
<td>0.6</td> |
|
<td>3407</td> |
|
<td>4.5</td> |
|
<td>447</td> |
|
<td>4.6</td> |
|
<td>435</td> |
|
<td>2.5</td> |
|
<td>815</td> |
|
<td>23.3</td> |
|
<td>86</td> |
|
</tr> |
|
<tr> |
|
<th rowspan="3" valign="top">H100x1</th> |
|
<th>deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B</th> |
|
<td>---</td> |
|
<td>0.4</td> |
|
<td>2604</td> |
|
<td>0.8</td> |
|
<td>1299</td> |
|
<td>0.4</td> |
|
<td>2543</td> |
|
<td>0.4</td> |
|
<td>2551</td> |
|
<td>3.3</td> |
|
<td>330</td> |
|
<td>3.4</td> |
|
<td>326</td> |
|
<td>1.8</td> |
|
<td>612</td> |
|
<td>14.0</td> |
|
<td>78</td> |
|
</tr> |
|
<tr> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-FP8-dynamic</th> |
|
<td>1.04</td> |
|
<td>0.4</td> |
|
<td>2694</td> |
|
<td>0.8</td> |
|
<td>1364</td> |
|
<td>0.4</td> |
|
<td>2670</td> |
|
<td>0.4</td> |
|
<td>2639</td> |
|
<td>3.2</td> |
|
<td>347</td> |
|
<td>3.2</td> |
|
<td>341</td> |
|
<td>1.6</td> |
|
<td>673</td> |
|
<td>14.1</td> |
|
<td>78</td> |
|
</tr> |
|
<tr> |
|
<th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16</th> |
|
<td>0.84</td> |
|
<td>0.5</td> |
|
<td>2111</td> |
|
<td>1.0</td> |
|
<td>1065</td> |
|
<td>0.5</td> |
|
<td>2068</td> |
|
<td>0.5</td> |
|
<td>2119</td> |
|
<td>4.1</td> |
|
<td>270</td> |
|
<td>4.1</td> |
|
<td>265</td> |
|
<td>2.1</td> |
|
<td>530</td> |
|
<td>15.1</td> |
|
<td>73</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). |
