Model Card for DeciCoder-6B

DeciCoder-6B is a 6 billion parameter decoder-only code completion model trained on the Python, Java, Javascript, Rust, C++, C, and C# subset of Starcoder Training Dataset. The model uses variable Grouped Query Attention and has a context window of 2k tokens. It was trained using a Fill-in-the-Middle training objective. The model's architecture was generated by Deci's proprietary Neural Architecture Search-based technology, AutoNAC.

Model Details

• Developed by: Deci
• Model type: DeciCoder-6B is an auto-regressive language model based on the transformer decoder architecture, using variable Grouped Query Attention.
• Language(s): Python, Java, JavaScript, Rust, C++, C, C#, Go
• License: Model checkpoints are licensed under the Apache 2.0

Documentation

• Blog Post: Introducing DeciCoder-6B: Code LLM Engineered for Accuracy & Cost Efficiency At Scale
• Tutorial: How to Run DeciCoder-6B on Qualcomm Cloud AI 100
• Google Colab Notebook
• Run DeciCoder on AWS DL2q instances using the Qualcomm Cloud AI Platform SDK
• Questions: Feel free to contact us via our Discord Community!

Model Architecture

Parameters	Layers	Heads	Sequence Length	GQA num_key_value_heads
6B	32	32	2k	Variable

• Decoder layer: Variable Grouped Query Attention
• Position Embeddings: Rotary Position Embeddings Su et al., 2021

How to Use

# pip install -q transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "Deci/DeciCoder-6B"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16, trust_remote_code=True).to(device)

inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to(device)
outputs = model.generate(inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

### Attribution

DeciCoder-6B was trained on StarCoder Training Dataset, filtered for
Python, Java, JavaScript, Ruby, RUST, C++, C, and C#. For additional information, please
refer to [https://huggingface.co/datasets/bigcode/starcoderdata](https://huggingface.co/datasets/bigcode/starcoderdata).

Limitations

The model has undergone training with source code from Python, Java, JavaScript, RUST, C++, C, and C#, and Go. While the primary language in the source is English, it does contain other languages. Therefore, the model can produce code snippets given some context. However, there is no assurance that the resulting code will function as expected. It might be suboptimal, contain bugs, or even exploits.

Evaluation

Below are DeciCoder-6B's pass@1 on MultiPL HumanEval scores

Python	JavaScript	Java	C++	C#	Rust	Go
33.3%	29.3%	30.3%	29.93%	20.31%	20.5%	77.47%

Runtime Benchmarks

Inference Tool	Hardware	Prompt Length	Generation Length	Throughput (tokens/sec)
Qualcomm Cloud AI 100 SDK	Qualcomm Cloud AI 100	1024	1024	531.3

• Measured for maximal batch size on the device

How to Cite

Please cite this model using this format.

@misc{DeciFoundationModels,
title = {DeciCoder-6B},
author = {DeciAI Research Team},
year = {2024}
url={[https://huggingface.co/deci/decicoder-6B](https://huggingface.co/deci/decicoder-6B)},
}