ByteDance-Seed/cudaLLM-8B

CudaLLM: A Language Model for High-Performance CUDA Kernel Generation

Model Description

cudaLLM-8B is a language model for generating high-performance and syntactically correct CUDA kernels. It is based on the Qwen3-8B model and has undergone a two-stage training process to master the complexities of parallel programming for GPUs.

Performance on KernelBench: | | Bo1 | Bo2 | Bo4 | Bo8 | Bo16 | |---------|-------|-----|-----|-----|------| | Level-1 | 79.75 | 83 | 84 | 86 | 87 | | Level-2 | 67.30 | 70 | 71 | 72 | 73 | | Level-3 | 20.83 | 26 | 30 | 34 | 36 |

Training Procedure

The model was trained using the verl library. The model was trained and evaluated on:

• SFT Dataset: A high-quality dataset of CUDA problem-solution pairs (sft_cuda_llm_r1.parquet), originally generated by DeepSeek R1, DeepSeel Coder-7B, and Qwen2-32B.
• RL Dataset: A refined dataset (rl_cuda_llm_0424.parquet) used to provide performance-based rewards during the RL stage.
• Evaluation Dataset: The model's performance was benchmarked against the KernelBench dataset.

Intended Use and Limitations

Intended Use

The primary use of CudaLLM is to assist developers in writing and optimizing high-performance CUDA kernels. It can be used for:

• Accelerating scientific computing and machine learning workloads.
• As a co-pilot or productivity tool for HPC and CUDA developers.
• Research into AI-driven code generation and optimization.

Limitations and Bias

• Correctness is Not Guaranteed: While trained to produce correct code, the model's output should always be rigorously tested and verified before deployment in production systems.
• Security Risks: The generated code is not guaranteed to be secure. Never run model-generated code from an untrusted source without careful inspection.
• Performance Variability: Kernel performance can vary significantly depending on the target GPU architecture, input data sizes, and compiler version. The generated code may require further manual tuning.
• Specialized Domain: This model is highly specialized for CUDA code generation. Its performance on general-purpose programming tasks or natural language conversation will be limited.