DeepSeek-R1-Distill-Llama-70B-quantized.w4a16

DeepSeek-R1-Distill-Llama-70B-quantized.w4a16

Model Overview

• Model Architecture: LlamaForCausalLM
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT4
• Release Date: 2/7/2025
• Version: 1.0
• Model Developers: Neural Magic

Quantized version of DeepSeek-R1-Distill-Llama-70B.

Model Optimizations

This model was obtained by quantizing the weights of DeepSeek-R1-Distill-Llama-70B to INT4 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%.

Only the weights of the linear operators within transformers blocks are quantized. Weights are quantized using a symmetric per-group scheme, with group size 128. The GPTQ algorithm is applied for quantization, as implemented in the llm-compressor library.

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

number_gpus = 1
model_name = "neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16"

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 for more details.

Creation

This model was created with llm-compressor by running the code snippet below.

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
from llmcompressor.transformers import oneshot
from llmcompressor.transformers.compression.helpers import calculate_offload_device_map

# Load model
model_stub = "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
model_name = model_stub.split("/")[-1]

num_samples = 3072
max_seq_len = 8192

tokenizer = AutoTokenizer.from_pretrained(model_stub)

device_map = calculate_offload_device_map(
    model_stub,
    reserve_for_hessians=True,
    num_gpus=2,
    torch_dtype="auto",
)

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 = QuantizationModifier(
    targets="Linear",
    scheme="W4A16",
    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.w4a16
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 and V2, using the following commands:

OpenLLM Leaderboard V1:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16",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-70B-quantized.w4a16",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-70B</th> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <th>Recovery</th> </tr> </thead> <tbody> <tr> <td rowspan="4"><b>Reasoning</b></td> <td>AIME 2024 (pass@1)</td> <td>67.83</td> <td>65.61</td> <td>96.73%</td> </tr> <tr> <td>MATH-500 (pass@1)</td> <td>95.29</td> <td>95.19</td> <td>99.9%</td> </tr> <tr> <td>GPQA Diamond (pass@1)</td> <td>65.57</td> <td>64.04</td> <td>97.67%</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>76.23</b></td> <td><b>74.95</b></td> <td><b>98.23%</b></td> </tr> <tr> <td rowspan="7"><b>OpenLLM V1</b></td> <td>ARC-Challenge (Acc-Norm, 25-shot)</td> <td>63.65</td> <td>63.31</td> <td>99.5%</td> </tr> <tr> <td>GSM8K (Strict-Match, 5-shot)</td> <td>93.03</td> <td>93.03</td> <td>100.0%</td> </tr> <tr> <td>HellaSwag (Acc-Norm, 10-shot)</td> <td>84.85</td> <td>84.43</td> <td>99.5%</td> </tr> <tr> <td>MMLU (Acc, 5-shot)</td> <td>78.04</td> <td>77.15</td> <td>98.9%</td> </tr> <tr> <td>TruthfulQA (MC2, 0-shot)</td> <td>56.67</td> <td>57.79</td> <td>102.0%</td> </tr> <tr> <td>Winogrande (Acc, 5-shot)</td> <td>78.22</td> <td>79.48</td> <td>101.6%</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>75.74</b></td> <td><b>75.86</b></td> <td><b>100.2%</b></td> </tr> <tr> <td rowspan="7"><b>OpenLLM V2</b></td> <td>IFEval (Inst Level Strict Acc, 0-shot)</td> <td>42.45</td> <td>42.41</td> <td>99.9%</td> </tr> <tr> <td>BBH (Acc-Norm, 3-shot)</td> <td>21.26</td> <td>16.96</td> <td>79.8%</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>9.51</td> <td>6.68</td> <td>---</td> </tr> <tr> <td>MUSR (Acc-Norm, 0-shot)</td> <td>14.87</td> <td>12.91</td> <td>---</td> </tr> <tr> <td>MMLU-Pro (Acc, 5-shot)</td> <td>4.27</td> <td>2.38</td> <td>---</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>15.39</b></td> <td><b>13.56</b></td> <td><b>88.1%</b></td> </tr> <tr> <td rowspan="4"><b>Coding</b></td> <td>HumanEval (pass@1)</td> <td>81.10</td> <td>80.20</td> <td><b>98.9%</b></td> </tr> <tr> <td>HumanEval (pass@10)</td> <td>87.60</td> <td>89.30</td> <td>101.9%</td> </tr> <tr> <td>HumanEval+ (pass@10)</td> <td>75.20</td> <td>73.00</td> <td>97.1%</td> </tr> <tr> <td>HumanEval+ (pass@10)</td> <td>83.10</td> <td>83.70</td> <td>100.7%</td> </tr> </tbody> </table>

Inference Performance

This model achieves up to 3.0x speedup in single-stream deployment and up to 2.6x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario. The following performance benchmarks were conducted with vLLM version 0.7.2, and GuideLLM.

<details> <summary>Benchmarking Command</summary>

guidellm --model neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 --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></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>GPU class</th> <th>Number of GPUs</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">A6000</th> <td>4</td> <th>deepseek-ai/DeepSeek-R1-Distill-Llama-70B</th> <td>---</td> <td>7.4</td> <td>152</td> <td>14.9</td> <td>76</td> <td>7.5</td> <td>149</td> <td>7.7</td> <td>146</td> <td>57.2</td> <td>20</td> <td>58.9</td> <td>19</td> <td>31.9</td> <td>35</td> <td>98.4</td> <td>11</td> </tr> <tr> <td>2</td> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8</th> <td>1.93</td> <td>7.7</td> <td>292</td> <td>15.2</td> <td>148</td> <td>7.8</td> <td>287</td> <td>8.0</td> <td>282</td> <td>60.7</td> <td>37</td> <td>60.2</td> <td>37</td> <td>32.3</td> <td>70</td> <td>104.0</td> <td>22</td> </tr> <tr> <td>2</td> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <td>2.83</td> <td>4.9</td> <td>457</td> <td>10.0</td> <td>225</td> <td>5.5</td> <td>411</td> <td>5.8</td> <td>389</td> <td>38.9</td> <td>58</td> <td>39.2</td> <td>57</td> <td>23.7</td> <td>95</td> <td>76.6</td> <td>29</td> </tr> <tr> <th rowspan="3" valign="top">A100</th> <td>2</td> <th>deepseek-ai/DeepSeek-R1-Distill-Llama-70B</th> <td>---</td> <td>6.4</td> <td>157</td> <td>12.8</td> <td>79</td> <td>6.6</td> <td>153</td> <td>6.7</td> <td>151</td> <td>50.4</td> <td>20</td> <td>50.8</td> <td>20</td> <td>27.0</td> <td>37</td> <td>85.4</td> <td>12</td> </tr> <tr> <td>2</td> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8</th> <td>1.48</td> <td>4.1</td> <td>245</td> <td>8.2</td> <td>123</td> <td>4.2</td> <td>238</td> <td>4.3</td> <td>235</td> <td>32.4</td> <td>31</td> <td>32.8</td> <td>31</td> <td>17.6</td> <td>57</td> <td>90.8</td> <td>11</td> </tr> <tr> <td>1</td> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <td>2.69</td> <td>4.6</td> <td>440</td> <td>9.2</td> <td>220</td> <td>4.9</td> <td>407</td> <td>5.2</td> <td>389</td> <td>35.3</td> <td>57</td> <td>36.3</td> <td>55</td> <td>21.2</td> <td>95</td> <td>68.1</td> <td>30</td> </tr> <tr> <th rowspan="3" valign="top">H100</th> <td>2</td> <th>deepseek-ai/DeepSeek-R1-Distill-Llama-70B</th> <td>---</td> <td>3.8</td> <td>149</td> <td>7.6</td> <td>74</td> <td>3.9</td> <td>146</td> <td>3.9</td> <td>144</td> <td>30.0</td> <td>19</td> <td>30.4</td> <td>19</td> <td>16.1</td> <td>35</td> <td>56.5</td> <td>10</td> </tr> <tr> <td>2</td> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic</th> <td>1.39</td> <td>2.7</td> <td>210</td> <td>5.3</td> <td>106</td> <td>2.7</td> <td>207</td> <td>2.8</td> <td>203</td> <td>21.1</td> <td>27</td> <td>21.4</td> <td>26</td> <td>11.5</td> <td>49</td> <td>47.2</td> <td>12</td> </tr> <tr> <td>1</td> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <td>1.83</td> <td>4.0</td> <td>277</td> <td>7.9</td> <td>138</td> <td>4.1</td> <td>266</td> <td>4.2</td> <td>262</td> <td>31.2</td> <td>35</td> <td>31.8</td> <td>34</td> <td>17.8</td> <td>61</td> <td>61.4</td> <td>18</td> </tr> </tbody> </table>

**Use case profiles: prompt tokens / generation tokens

**QPD: Queries per dollar, based on on-demand cost at Lambda Labs (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">A6000x4</th> <th>deepseek-ai/DeepSeek-R1-Distill-Llama-70B</th> <td>---</td> <td>3.65</td> <td>4102</td> <td>1.56</td> <td>1757</td> <td>1.90</td> <td>2143</td> <td>1.48</td> <td>1665</td> <td>0.44</td> <td>493</td> <td>0.34</td> <td>380</td> <td>0.22</td> <td>245</td> <td>0.05</td> <td>55</td> </tr> <tr> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8</th> <td>1.76</td> <td>5.89</td> <td>6625</td> <td>2.94</td> <td>3307</td> <td>3.36</td> <td>3775</td> <td>2.59</td> <td>2916</td> <td>0.74</td> <td>828</td> <td>0.53</td> <td>601</td> <td>0.35</td> <td>398</td> <td>0.11</td> <td>120</td> </tr> <tr> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <td>1.48</td> <td>4.91</td> <td>5528</td> <td>2.01</td> <td>2259</td> <td>2.03</td> <td>2280</td> <td>1.12</td> <td>1255</td> <td>1.11</td> <td>1251</td> <td>0.76</td> <td>852</td> <td>0.24</td> <td>267</td> <td>0.07</td> <td>81</td> </tr> <tr> <th rowspan="3" valign="top">A100x4</th> <th>deepseek-ai/DeepSeek-R1-Distill-Llama-70B</th> <td>---</td> <td>10.41</td> <td>5235</td> <td>5.10</td> <td>2565</td> <td>5.50</td> <td>2766</td> <td>4.36</td> <td>2193</td> <td>1.49</td> <td>751</td> <td>1.21</td> <td>607</td> <td>0.89</td> <td>447</td> <td>0.19</td> <td>98</td> </tr> <tr> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8</th> <td>1.63</td> <td>18.11</td> <td>9103</td> <td>8.90</td> <td>4477</td> <td>9.41</td> <td>4730</td> <td>7.42</td> <td>3731</td> <td>2.44</td> <td>1229</td> <td>1.89</td> <td>948</td> <td>1.26</td> <td>631</td> <td>0.30</td> <td>149</td> </tr> <tr> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <td>1.12</td> <td>12.63</td> <td>6353</td> <td>5.32</td> <td>2673</td> <td>5.58</td> <td>2804</td> <td>4.27</td> <td>2144</td> <td>2.30</td> <td>1158</td> <td>1.45</td> <td>729</td> <td>0.76</td> <td>381</td> <td>0.22</td> <td>110</td> </tr> <tr> <th rowspan="3" valign="top">H100x4</th> <th>deepseek-ai/DeepSeek-R1-Distill-Llama-70B</th> <td>---</td> <td>14.04</td> <td>2113</td> <td>10.85</td> <td>1634</td> <td>12.25</td> <td>1844</td> <td>9.93</td> <td>1494</td> <td>3.68</td> <td>554</td> <td>2.82</td> <td>425</td> <td>1.81</td> <td>273</td> <td>0.35</td> <td>52</td> </tr> <tr> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic</th> <td>1.78</td> <td>41.44</td> <td>6236</td> <td>19.64</td> <td>2956</td> <td>21.03</td> <td>3166</td> <td>16.72</td> <td>2516</td> <td>6.01</td> <td>904</td> <td>4.46</td> <td>672</td> <td>2.55</td> <td>383</td> <td>0.49</td> <td>74</td> </tr> <tr> <th>neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16</th> <td>1.45</td> <td>36.61</td> <td>5509</td> <td>15.12</td> <td>2275</td> <td>16.24</td> <td>2443</td> <td>13.22</td> <td>1990</td> <td>5.48</td> <td>825</td> <td>3.01</td> <td>453</td> <td>2.07</td> <td>312</td> <td>0.43</td> <td>64</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 (observed on 2/18/2025).