Ling-Coder-lite-GPTQ-Int8

Introduction

Ling-Coder-Lite is a MoE LLM provided and open-sourced by InclusionAI, which has 16.8B parameters with 2.75B activated parameters. This model demonstrates state-of-the-art performance on 12 coding benchmarks, while simultaneously offering competitive latency and throughput compared to code LLMs of similar size. In addition to open-sourcing the model itself, we also release a substantial amount of code-related data, including synthetic QA, SFT and DPO datasets. More details are described in the technique report Ling-Coder-TR.

This repo contains the GPTQ-quantized 8-bit Ling-Coder-lite model which can be served using vLLM.

Model Downloads

You can download the following table to see the various parameters for your use case. If you are located in mainland China, we also provide the model on modelscope.cn to speed up the download process.

Model	#Total Params	#Activated Params	Context Length	Download
Ling-Coder-lite-base	16.8B	2.75B	16K	🤗 HuggingFace
Ling-Coder-lite	16.8B	2.75B	16K	🤗 HuggingFace
Ling-Coder-lite-GPTQ-Int8	16.8B	2.75B	16K	🤗 HuggingFace

Dataset Downloads

Model	Samples	Download
Ling-Coder-SyntheticQA	24M	🤗 HuggingFace
Ling-Coder-SFT	5M	🤗 HuggingFace
Ling-Coder-DPO	250K	🤗 HuggingFace

Evaluation

Detailed evaluation results are reported in our technical report Ling-Coder-TR.

For the quantized model deployed via vLLM, preliminary evaluation results are presented below:

Benchmark	Scores
HumanEval-Python	88.41
MBPP-Python	73.28
EvalPlus-HumanEval+	85.37
EvalPlus-MBPP+	73.28

Quickstart

vLLM

Requirement: vllm==0.6.3.post1.

Patch ling_gptq.patch onto vLLM by executing:

patch -p1 < ling_gptq.patch -d $(python -c "from importlib.util import find_spec; print(find_spec('vllm').submodule_search_locations[0])")

from vllm import LLM
from vllm.sampling_params import SamplingParams
from transformers import AutoTokenizer

model_name = "inclusionAI/Ling-Coder-lite-GPTQ-Int8"

model = LLM(model_name, trust_remote_code=True, gpu_memory_utilization=0.80, max_model_len=4096)

tokenizer = AutoTokenizer.from_pretrained(
    model_name, 
    trust_remote_code=True
)

prompt = "Write a quick sort algorithm in python."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

sample_params = SamplingParams(max_tokens=1024, ignore_eos=False)
outputs = model.generate(text, sampling_params=sample_params, prompt_token_ids=None)

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

Note: Ling Coder Lite GTPQ int8 only borrows the inference code of DeepSeek MoE in vLLM. The model itself is not related to DeepSeek.

Deployment

Please refer to Github

License

This code repository is licensed under the MIT License.

Citation

@misc{codefuse2025samplemattersleveragingmixtureofexperts,
      title={Every Sample Matters: Leveraging Mixture-of-Experts and High-Quality Data for Efficient and Accurate Code LLM}, 
      author={Codefuse and Ling Team},
      year={2025},
      eprint={2503.17793},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.17793}, 
}