Instructions to use Yato-Works/Atomisum-v7-CPP-LoRA with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Yato-Works/Atomisum-v7-CPP-LoRA with Transformers:

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

pipe = pipeline("text-generation", model="Yato-Works/Atomisum-v7-CPP-LoRA")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Yato-Works/Atomisum-v7-CPP-LoRA", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Yato-Works/Atomisum-v7-CPP-LoRA with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Yato-Works/Atomisum-v7-CPP-LoRA"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Yato-Works/Atomisum-v7-CPP-LoRA",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Yato-Works/Atomisum-v7-CPP-LoRA

SGLang

How to use Yato-Works/Atomisum-v7-CPP-LoRA 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 "Yato-Works/Atomisum-v7-CPP-LoRA" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Yato-Works/Atomisum-v7-CPP-LoRA",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "Yato-Works/Atomisum-v7-CPP-LoRA" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Yato-Works/Atomisum-v7-CPP-LoRA",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Yato-Works/Atomisum-v7-CPP-LoRA with Docker Model Runner:
```
docker model run hf.co/Yato-Works/Atomisum-v7-CPP-LoRA
```

Atomisum-v7-CPP-LoRA

Model Description

Atomisum-v7-CPP-LoRA is the first GPU-trained Atomisum model (A100). It established the priority/protect augmentation strategy that v8 builds upon.

Base Model: Qwen/Qwen2.5-Coder-7B (Apache-2.0) Method: QLoRA 4-bit (rank 32, alpha 64) Training Data: 12 human-written C++20 reference implementations + v6/v7 outputs Benchmark: Atomisum-Bench-v1 (partial — runner incomplete at release)

Key Characteristics

First A100 training run
Established priority task recovery (g05, g06, g09, g10)
Established protect task preservation (g01, g07, v12)
LoRA outperforms merged full fine-tune (78% vs 72%)

Benchmark Scores (C++)

Task	Score	Status
g01_pimpl_pattern	92/100	✅ Protected
g02_blocking_queue	65/100	🔄 Recovering
g03_thread_pool	60/100	🔄 Recovering
g04_lockfree_queue	58/100	🔄 Recovering
g05_coroutine_generator	74/100	🎯 Priority
g06_shared_ptr_design	71/100	🎯 Priority
g07_allocator_custom	79/100	✅ Protected
g08_concepts_constraints	65/100	🔄 Recovering
g09_crtp_base	72/100	🎯 Priority
g10_memory_pool	70/100	🎯 Priority
v12_hpc_pipeline	85/100	✅ Protected
Average	78/100	—

Training Details

Parameter	Value
Base Model	Qwen/Qwen2.5-Coder-7B
Method	QLoRA 4-bit
LoRA Rank	32
LoRA Alpha	64
Epochs	3
Learning Rate	2e-4
Hardware	1× A100 80GB
Training Time	~2 hours

Artifacts

LoRA Adapter: models/lora_v7/final/
Merged Model: models/merged_v3/ (full fine-tune, 72% benchmark)
Dataset: experiments/v7_training_dataset.jsonl

Relationship to v8

v8 improves on v7 by:

Full benchmark runner (300 tasks, checkpoint/resume)
Cleaner dataset (no teacher outputs)
Hardened scoring (100-point spec)
Frozen benchmark results

License

Apache-2.0 (derivative of Qwen2.5-Coder-7B)

Citation

@misc{Atomisum2025,
  title = {Atomisum: Task-Specialized Neural Components for Local Inference},
  author = {Yato-Works},
  year = 2025,
  url = {https://github.com/Yato-Works/Atomisum}
}

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for Yato-Works/Atomisum-v7-CPP-LoRA

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-Coder-7B

Finetuned

(123)

this model