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-8B-16MLA-16Mamba-SFT, is created by efficiently adapting the pre-trained Llama-3.1-8B-Instruct model conducted post-training on AMD Instinct™ MI300X GPUs. This training approach bypasses the need for costly pre-training from scratch.
Figure 1: 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).
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:
Stage
Action
Description
1. Base Model
Llama-3.1-8B-Instruct
The starting point is a high-quality, pre-trained Transformer model.
2. Initialization
Structured Weight Mapping
Pure Mamba2 and MLA models are initialized from the base model's weights using structured mapping techniques (SVD for MLA, reinterpretation for Mamba2).
3. Refinement
Intermediate Layer Distillation (ILD)
The internal representations of the Mamba2 and MLA models are aligned with the base model's layers on a small dataset to ensure a strong starting point.
4. Composition
SMART Layer Selection
A hybrid architecture is composed using the SMART (Sensitivity Measure-Aware Replacement of Transformer layers) strategy to optimally place each layer type.
5. SFT
End-to-End Knowledge Distillation
The composed hybrid model is fine-tuned via knowledge distillation, using an 8B model as a teacher to transfer rich, pre-trained knowledge.
6. Alignment
Direct Preference Optimization (DPO)
In the final stage, DPO is used to align the model's preferences, with the distilled student model itself serving as the reference model for stability.
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
python
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3from hybrid.hybrid_wrapper import HybridModelWrapper
45checkpoint ="amd/Zebra-Llama-8B-16MLA-16Mamba-SFT"67model = HybridModelWrapper.from_pretrained(checkpoint, torch_dtype=torch.bfloat16).cuda()8tokenizer = AutoTokenizer.from_pretrained(checkpoint)9model.eval()1011# Format the prompt using the chat template12prompt =[{"role":"user","content":"What are the benefits of hybrid language models?"}]13input_ids = tokenizer.apply_chat_template(14 prompt,15 add_generation_prompt=True,16 return_tensors='pt'17).cuda()1819# Generate a response20tokens = model.generate(21 input_ids,22 max_new_tokens=256,23 temperature=0.7,24 do_sample=True,25 eos_token_id=tokenizer.eos_token_id
26)2728print(tokenizer.decode(tokens[0], skip_special_tokens=False))
Model details
Model
KV Size
Param
Index of MLA layers
rkv
rq
drope
dnope
Llama-3.1-8B-Instruct
100%
8.03B
-
-
-
-
-
Zebra-Llama-8B-8MLA-24Mamba
2.73%
8.38B
[0,4,8,12,16,20,25,30]
160
2048
64
64
Zebra-Llama-8B-16MLA-16Mamba
5.47%
8.19B
[0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30]
160
2048
64
64
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.
Tasks
Metric
Llama-3.1-8B-Instruct
Zebra-Llama-8B-8MLA-24M2-SFT
Zebra-Llama-8B-8MLA-24M2-DPO
Zebra-Llama-8B-16MLA-16M2-SFT
Zebra-Llama-8B-16MLA-16M2-DPO
arc_challenge
acc
0.5162±0.0146
0.4923±0.0146
0.5648±0.0145
0.4898±0.0146
0.5717±0.0145
acc_norm
0.5512±0.0145
0.5205±0.0146
0.5887±0.0144
0.5213±0.0146
0.5862±0.0144
arc_easy
acc
0.8186±0.0079
0.8068±0.0081
0.8274±0.0078
0.803±0.0082
0.8312±0.0077
acc_norm
0.7963±0.0083
0.7706±0.0086
0.7917±0.0083
0.7689±0.0086
0.7837±0.0084
hellaswag
acc
0.5913±0.0049
0.5496±0.005
0.5906±0.0049
0.5569±0.005
0.5953±0.0049
acc_norm
0.7919±0.0041
0.7427±0.0044
0.786±0.0041
0.7491±0.0043
0.7927±0.004
mmlu
acc
0.6814±0.1113
0.5567±0.1157
0.546±0.1124
0.5882±0.1212
0.5817±0.1055
- humanities
acc
0.6436±0.0906
0.5048±0.1131
0.4944±0.1132
0.535±0.1277
0.5596±0.09
- other
acc
0.7441±0.1097
0.6283±0.0983
0.6061±0.0982
0.6521±0.1
0.6283±0.0995
- social_sciences
acc
0.7709±0.0621
0.6487±0.0733
0.6344±0.0791
0.6887±0.0752
0.6581±0.0757
- stem
acc
0.5886±0.1252
0.4738±0.1179
0.4776±0.1094
0.5065±0.1161
0.4941±0.1154
openbookqa
acc
0.334±0.0211
0.32±0.0209
0.348±0.0213
0.32±0.0209
0.364±0.0215
acc_norm
0.43±0.0222
0.418±0.0221
0.436±0.0222
0.43±0.0222
0.434±0.0222
piqa
acc
0.7992±0.0093
0.7878±0.0095
0.8025±0.0093
0.7933±0.0094
0.7992±0.0093
acc_norm
0.8085±0.0092
0.7884±0.0095
0.7943±0.0094
0.7927±0.0095
0.8003±0.0093
pubmedqa
acc
0.756±0.0192
0.688±0.0207
0.718±0.0201
0.728±0.0199
0.728±0.0199
race
acc
0.4459±0.0154
0.4077±0.0152
0.4622±0.0154
0.4239±0.0153
0.4928±0.0155
winogrande
acc
0.7427±0.0123
0.6969±0.0129
0.7245±0.0126
0.7032±0.0128
0.7261±0.0125
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}
}
