---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
library_name: transformers
pipeline_tag: text-generation
tags:
  - code
  - cuda
  - gpu-kernels
  - kernel-generation
  - kernelbench
  - arxiv:2609.33074
---

# KernelZero-CUDA-7B

KernelZero-CUDA-7B is the CUDA kernel generation checkpoint reported as the final KernelZero CUDA model. It is based on Qwen2.5-Coder-7B-Instruct and trained with the KernelZero pipeline.

## Paper

[KernelZero: Continual Co-Evolution of LLMs for Kernel Generation](https://arxiv.org/abs/2609.33074)

## KernelBench results

The paper reports the following results with 10 sampled candidates per problem.

| Benchmark | pass@1 | pass@5 | pass@10 | fast\_1@1 | fast\_1@10 | fast\_2@1 | fast\_2@10 |
|---|---:|---:|---:|---:|---:|---:|---:|
| Level 1 | 75.8 | 98.87 | 100.0 | 17.6 | 29.0 | 7.2 | 12.0 |
| Level 2 | 69.6 | 93.70 | 97.0 | 2.4 | 12.0 | 1.2 | 6.0 |

## Intended use

This checkpoint generates CUDA extensions from PyTorch modules. Use the same prompt template and validation environment as KernelZero for reproducible evaluation.

## Loading

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "kcxain/KernelZero-CUDA-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
```

## Checkpoint provenance

This release contains the merged Hugging Face inference weights from the final 80-step CUDA Coder followed by the 40-step continuation used for the paper's main CUDA result. Optimizer states and distributed training shards are excluded.