Instructions to use kcxain/KernelZero-CUDA-7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kcxain/KernelZero-CUDA-7B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="kcxain/KernelZero-CUDA-7B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("kcxain/KernelZero-CUDA-7B")
model = AutoModelForCausalLM.from_pretrained("kcxain/KernelZero-CUDA-7B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use kcxain/KernelZero-CUDA-7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kcxain/KernelZero-CUDA-7B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kcxain/KernelZero-CUDA-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kcxain/KernelZero-CUDA-7B

SGLang

How to use kcxain/KernelZero-CUDA-7B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "kcxain/KernelZero-CUDA-7B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kcxain/KernelZero-CUDA-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "kcxain/KernelZero-CUDA-7B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kcxain/KernelZero-CUDA-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use kcxain/KernelZero-CUDA-7B with Docker Model Runner:
```
docker model run hf.co/kcxain/KernelZero-CUDA-7B
```

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.

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

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.