Zebra-Llama-3B-6MLA-22Mamba-DPO

Zebra-Llama: Towards Extremely Efficient Hybrid Models

Zebra-Llama is a family of hybrid large language models (LLMs) proposed by AMD that composes Multi-head Latent Attention (MLA) and Mamba2 for KV cache compression and computational efficiency. Thus combination achieves Transformer-level accuracy with near-State Space Model (SSM) efficiency. While standard Transformers are limited by the quadratic complexity of self-attention and the large memory footprint of their key-value (KV) cache, Zebra-Llama offers a practical and scalable solution.

This model, Zebra-Llama-3B-6MLA-22Mamba-DPO, is created by efficiently adapting the pre-trained Llama-3.2-3B-Instruct model conducted post-training on AMD Instinct™ MI300X GPUs. This training approach bypasses the need for costly pre-training from scratch.

<div align="center"> <img src="comparison.png" width="570" height="380" style="object-fit: contain;"/> <em><b>Figure 1:</b> Comparing 8B-scale models on average LM Harness score vs. KV cache size. Zebra-Llama (green) matches or exceeds baselines with smaller KV cache and fewer training tokens. Circle and square sizes indicate training tokens (billions for post-training, trillions for pre-training).</em> </div>

Key Takeaways

• Announcing Zebra-Llama, a family of highly efficient 1B, 3B, and 8B hybrid models created by post-training adaptation of existing state-of-the-art Transformers.
• Extreme KV Cache Compression: Zebra-Llama dramatically reduces the KV cache size to 2%-4% of the original Llama model while preserving 100% of its average zero-shot performance on LM Harness tasks.
• Efficient Hybrid Architecture: Zebra-Llama strategically combines Multi-head Latent Attention (MLA) layers, which compress the KV cache, and Mamba2 (SSM) layers, which eliminate the KV cache entirely, to balance memory usage and performance.
• Novel Post-Training Pipeline: Zebra-Llama employs an efficient post-training pipeline featuring refined weight initialization, Intermediate Layer Distillation (ILD) for knowledge transfer, and a sensitivity-aware strategy (SMART) for optimal hybrid composition.

Model Composition Pipeline

The Zebra-Llama models are not trained from scratch. Instead, they are composed from powerful pre-trained Transformers through a lightweight and efficient pipeline. The creation of this model followed these stages:

Training Data

Getting Started

Installation

git clone https://github.com/AMD-AIG-AIMA/AMD-Hybrid-Models.git

Then follow the installation instruction in AMD-AIG-AIMA/AMD-Hybrid-Models repo.

Example Usage

Once the installation completed, we can try the following code for a quick test

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from hybrid.hybrid_wrapper import HybridModelWrapper

checkpoint = "amd/Zebra-Llama-3B-6MLA-22Mamba-DPO"

model = HybridModelWrapper.from_pretrained(checkpoint, torch_dtype=torch.bfloat16).cuda()
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model.eval()

# Format the prompt using the chat template
prompt = [{"role": "user", "content": "What are the benefits of hybrid language models?"}]
input_ids = tokenizer.apply_chat_template(
    prompt,
    add_generation_prompt=True,
    return_tensors='pt'
).cuda()

# Generate a response
tokens = model.generate(
    input_ids, 
    max_new_tokens=256,
    temperature=0.7,
    do_sample=True,
    eos_token_id=tokenizer.eos_token_id
)

print(tokenizer.decode(tokens[0], skip_special_tokens=False))

Model details

Benchmark results

Zebra-Llama was evaluated on the Language Model Harness benchmark for zero-shot tasks and compared against its base model and other post-training methods. The results demonstrate that Zebra-Llama provides a superior balance of performance and efficiency.

Conclusion

Zebra-Llama demonstrates a practical and scalable framework for composing highly efficient hybrid models from existing pre-trained Transformers. By intelligently combining MLA and Mamba2 layers, this approach drastically reduces memory requirements and improves inference throughput while preserving the strong capabilities of the original model. This work highlights the viability of post-training hybridization as a cost-effective and environmentally sustainable alternative to full retraining, paving the way for the deployment of powerful LLMs in resource-constrained environments.

Bias, Risks, and Limitations

• This model is a research artifact and has not been evaluated for safety in production use cases.
• The model's performance is dependent on the quality of its pre-trained base model and the teacher model used during distillation. Its capabilities and biases are inherited from these sources.
• The model may generate content that is factually inaccurate, biased, or otherwise objectionable. Users should be aware of these risks and implement appropriate safeguards for their applications.
• One limitation of this work is the reliance on a strong teacher model for knowledge transfer, which may not always be available. Distillation from a teacher also adds to the resource requirements during the post-training phase.

Citation

If you find this model useful, please consider citing the original paper:

@article{yang2025zebra,
  title={Zebra-Llama: Towards Extremely Efficient Hybrid Models},
  author={Yang, Mingyu and Rezagholizadeh, Mehdi and Li, Guihong and Appia, Vikram and Barsoum, Emad},
  journal={arXiv preprint arXiv:2505.17272},
  year={2025}
}
@article{li2025x,
  title={X-ecomla: Upcycling pre-trained attention into mla for efficient and extreme kv compression},
  author={Li, Guihong and Rezagholizadeh, Mehdi and Yang, Mingyu and Appia, Vikram and Barsoum, Emad},
  journal={arXiv preprint arXiv:2503.11132},
  year={2025}
}