ContainersNVIDIADeepSeek-R1NVIDIALlama-3.1-Nemotron-70B-InstructNVIDIAPyTorch
CollectionsNVIDIAOmniverse Kit (FB)NVIDIADeepStream SDKNVIDIAOmniverse Kit App Streaming
ModelsNVIDIAStyleGAN3 pretrained modelsNVIDIAPeopleNetNVIDIATrafficCamNet
ResourcesNVIDIARiva Skills Quick Start
Helm ChartsNVIDIAGPU OperatorNVIDIANVIDIA NIM Operator
NVIDIA
NVIDIA
CUDA Autocomplete
Model
NVIDIA
NVIDIA
CUDA Autocomplete

Fine-tuned Qwen2.5-Coder-7B for CUDA code completion in Nsight Copilot

This model is backed by NVIDIA's Plus Plus (++) Promise
to learn more about the quality of the datasets used to train this model.

Explainability Subcard

intended_domain

Code generation in VSCode / Cursor.

Model Type

Transformer

Intended Users

This model is designed for software developers, CUDA programmers, and AI enthusiasts who want to accelerate their coding workflow with intelligent code completion. Primary users include:

  • CUDA Developers: Engineers building GPU-accelerated applications, high-performance computing (HPC) solutions, or scientific computing applications who need assistance with CUDA-specific syntax and patterns.
  • Software Engineers: General programmers working in languages like Python, C++, and Java who want AI-powered code suggestions to boost productivity.
  • Students and Learners: Individuals learning CUDA programming or general software development who benefit from contextual code completion as a learning aid.
  • AI/ML Engineers: Developers building AI applications who need efficient coding assistance within the VSCode/Cursor development environment.

Output

Types: Several lines of code. Formats: Python, C++, Java, etc.

Describe how the model works:

The model accepts a block of code around the cursor (as left and right context) and predicts the sequence of code that is the best continuation of the code.

Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:

Not Applicable

Technical Limitations & Mitigation:

The model is currently trained with a large distribution of C++, Python, and CUDA code. As the representation of other languages is limited, we recommend fine-tuning the model on more diverse languages and domains.

Verified to have met prescribed NVIDIA quality standards:

Yes

Performance Metrics:

Acceptance rate of the code completion. Offline metrics include LLM-as-a-judge metrics and ROUGE-L score.

Potential Known Risks:

This model may occasionally generate incorrect responses or produce repetitive code. To mitigate this, mechanisms are in place to reduce repetition, and model parameters such as temperature and max_tokens have been carefully configured to minimize these risks.

Licensing:

Use of this model is governed by the NVIDIA Open Model License Agreement.

Additional Information.For Qwen2.5-Coder-7B, Apache License, Version 2.0.