Instructions to use arthu1/VEG-Transformer-Transplant with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use arthu1/VEG-Transformer-Transplant with Transformers:

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

pipe = pipeline("text-generation", model="arthu1/VEG-Transformer-Transplant")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("arthu1/VEG-Transformer-Transplant")
model = AutoModelForCausalLM.from_pretrained("arthu1/VEG-Transformer-Transplant", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use arthu1/VEG-Transformer-Transplant with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "arthu1/VEG-Transformer-Transplant"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "arthu1/VEG-Transformer-Transplant",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/arthu1/VEG-Transformer-Transplant

SGLang

How to use arthu1/VEG-Transformer-Transplant 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 "arthu1/VEG-Transformer-Transplant" \
    --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": "arthu1/VEG-Transformer-Transplant",
		"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 "arthu1/VEG-Transformer-Transplant" \
        --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": "arthu1/VEG-Transformer-Transplant",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use arthu1/VEG-Transformer-Transplant with Docker Model Runner:
```
docker model run hf.co/arthu1/VEG-Transformer-Transplant
```

VEG Transformer Transplant

Experimental structural transplant. This is not equivalent to the VEG checkpoint.

A standard Transformer cannot directly express VEG's phase-coherence routing, elementwise sigmoid route gate, causal 1/sqrt(t+1) normalization, or selective recurrent field. This artifact maps the tensors that have a Llama analogue into a normal decoder-only Transformer. It is useful as a compatibility baseline and conversion reference, not as a preserved-quality conversion.

Model

Architecture: LlamaForCausalLM
Parameters: 84,042,816
Vocabulary: 256 bytes
Context: 512
Hidden size: 832
Layers: 10
Attention heads: 13, head dimension 64
Intermediate size: 2240
Precision: FP32

Mapping

emb -> model.embed_tokens.weight
Wq/Wk -> duplicated 64-dimensional q/k heads
Wv * c -> self_attn.v_proj.weight
diag(gs) @ Wo -> self_attn.o_proj.weight
W1/W3/W2 -> SwiGLU projections
g1/g2/gf -> RMSNorm gains
Wout -> lm_head.weight

Dropped or approximated VEG behavior is listed in conversion_report.json. Expect degraded quality relative to the source model.

Use

pip install -r requirements.txt

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained('.', torch_dtype=torch.float32)
tok = AutoTokenizer.from_pretrained('.')
inputs = tok('The capital of France is', return_tensors='pt')
out = model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.8)
print(tok.decode(out[0], skip_special_tokens=False))

The tokenizer is byte-level and has no BOS/EOS token. Generation should use an explicit max_new_tokens or a custom stopping criterion.

Downloads last month: 131

Safetensors

Model size

84M params

Tensor type

F32