X-EcoMLA is an efficient KV cache compression technique for large language models (LLMs) proposed by AMD that upcycles transformer blocks into Multi-head Latent Attention (MLA) for extreme KV cache compression and computational efficiency. Instead of training a MLA model from scratch, the proposed X-EcoMLA first initializes the MLA weights based on Singular Value Decomposition (SVD) of the existing transformer weights, followed by lightweight pre-training or post-training distillation. This model, X-EcoMLA-1B8B-fixed-kv64-DPO, is created by efficiently adapting the pre-trained Llama-3.2-1B-Instruct model conducted post-training on AMD Instinct™ MI300X GPUs. This training approach bypasses the need for costly pre-training from scratch.

Key Takeaways

Announcing X-EcoMLA, an efficient approach to upcycle existing transformer blocks into MLA.
Extreme KV Cache Compression: X-EcoMLA dramatically reduces the KV cache size by 6.4x - 10.6x with only 3.6B - 7B training tokens, while preserving almost 100% of its average zero-shot performance on LM Harness tasks.
Novel SVD Initialization: X-EcoMLA employs an efficient SVD-based weight initialization which dramatically improves the training efficiency and model performance.

Model Composition Pipeline

The X-EcoMLA 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-3.2-1B-Instruct	The starting point is a high-quality, pre-trained Transformer model.
2. Initialization	Structured Weight Mapping	MLA models are initialized from the base model's weights using SVD.
3. SFT	End-to-End Knowledge Distillation	The initialized model is fine-tuned via knowledge distillation.
4. 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.

Training Data

Stage	Dataset	License
SFT	https://huggingface.co/datasets/teknium/OpenHermes-2.5	Refer source materials
SFT	https://huggingface.co/datasets/tomg-group-umd/GenQA	CC BY-NC 4.0
SFT	https://huggingface.co/datasets/BAAI/Infinity-Instruct	CC BY-SA 4.0
DPO	https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized	MIT
DPO	https://huggingface.co/datasets/HuggingFaceH4/orca_dpo_pairs	MIT
DPO	https://huggingface.co/datasets/JunxiongWang/llama3-ultrafeedback-armorm	MIT

Getting Started

Installation

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

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

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from mla.hybrid_wrapper import MLATransformerHybridModelWrapper

checkpoint = "amd/X-EcoMLA-1B8B-fixed-kv64-DPO"

model = MLATransformerHybridModelWrapper.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 Evaluation:

python benchmark/llm_eval/lm_harness_eval.py \
    --model mla_hybrid \
    --model_args pretrained="amd/X-EcoMLA-1B8B-fixed-kv64-DPO" \
    --tasks mmlu,hellaswag,piqa,arc_easy,arc_challenge,winogrande,openbookqa,pubmedqa,race \
    --num_fewshot 0 --device cuda --batch_size 16

Model details

Model	KV Size	Target Model	Teacher Model	Training Tokens	Pre-/Post-Training	r_kv	r_q	d_rope	d_nope
X-EcoMLA-1B1B-fixed-kv512-DPO	53.1%	Llama-3.2-1B-Instruct	Llama-3.2-1B-Instruct	7B	Post	512	864	32	32
X-EcoMLA-1B1B-dynamic-0.95-DPO	54.7%	Llama-3.2-1B-Instruct	Llama-3.2-1B-Instruct	7B	Post	0.95	0.95	32	32
X-EcoMLA-1B8B-fixed-kv64-DPO	9.4%	Llama-3.2-1B-Instruct	Llama-3.1-8B-Instruct	7B	Post	64	1424	32	32
X-EcoMLA-3B3B-fixed-kv816-DPO	43%	Llama-3.2-3B-Instruct	Llama-3.2-3B-Instruct	7B	Post	816	1536	64	64
X-EcoMLA-3B3B-dynamic-0.95-DPO	43%	Llama-3.2-3B-Instruct	Llama-3.2-3B-Instruct	7B	Post	0.95	0.95	64	64
X-EcoMLA-SmolLM-1.7B-fixed-kv480-Pretrain	12.5%	SmolLM-1.7B	-	6B	Pre	480	2048	32	32
X-EcoMLA-SmolLM-1.7B1.7B-fixed-kv480-Pretrain	12.5%	SmolLM-1.7B	SmolLM-1.7B	6B	Pre	480	2048	32	32
X-EcoMLA-SmolLM-1.7B1.7B-fixed-kv480-DPO	12.5%	SmolLM-1.7B-Instruct	SmolLM-1.7B-Instruct	7B	Post	480	2048	32	32

Benchmark results

X-EcoMLA 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-3.2-1B-Instruct	X-EcoMLA-1B1B-fixed-kv512-DPO	X-EcoMLA-1B1B-dynamic-0.95-DPO	X-EcoMLA-1B8B-fixed-kv64-DPO
arc_challenge	acc	0.3575 (±0.0140)	0.3643 (±0.0141)	0.3686 (±0.0141)	0.3729 (±0.0141)
	acc_norm	0.3797 (±0.0142)	0.3993 (±0.0143)	0.4121 (±0.0144)	0.3985 (±0.0143)
arc_easy	acc	0.6843 (±0.0095)	0.6873 (±0.0095)	0.6932 (±0.0095)	0.7256 (±0.0092)
	acc_norm	0.6351 (±0.0099)	0.6389 (±0.0099)	0.6486 (±0.0098)	0.6713 (±0.0096)
hellaswag	acc	0.4506 (±0.0050)	0.4483 (±0.0050)	0.4459 (±0.0050)	0.4398 (±0.0050)
	acc_norm	0.6077 (±0.0049)	0.6073 (±0.0049)	0.6096 (±0.0049)	0.5845 (±0.0049)
mmlu	acc	0.4609 (±0.0918)	0.4239 (±0.0785)	0.4286 (±0.0809)	0.3851 (±0.0684)
- humanities	acc	0.4397 (±0.0763)	0.4064 (±0.0663)	0.4013 (±0.0733)	0.3609 (±0.0565)
- other	acc	0.5204 (±0.0868)	0.4583 (±0.0760)	0.4747 (±0.0774)	0.4052 (±0.0632)
- social_sciences	acc	0.5109 (±0.0843)	0.4686 (±0.0735)	0.4729 (±0.0734)	0.4277 (±0.0676)
- stem	acc	0.3850 (±0.0900)	0.3723 (±0.0818)	0.3806 (±0.0798)	0.3600 (±0.0768)
openbookqa	acc	0.2440 (±0.0192)	0.2560 (±0.0195)	0.2660 (±0.0198)	0.2600 (±0.0196)
	acc_norm	0.3500 (±0.0214)	0.3780 (±0.0217)	0.3760 (±0.0217)	0.3740 (±0.0217)
piqa	acc	0.7405 (±0.0102)	0.7443 (±0.0102)	0.7301 (±0.0104)	0.7334 (±0.0103)
	acc_norm	0.7437 (±0.0102)	0.7492 (±0.0101)	0.7443 (±0.0102)	0.7383 (±0.0103)
pubmedqa	acc	0.6020 (±0.0219)	0.5880 (±0.0220)	0.5860 (±0.0220)	0.5800 (±0.0221)
race	acc	0.3809 (±0.0150)	0.4077 (±0.0152)	0.3923 (±0.0151)	0.3981 (±0.0151)
winogrande	acc	0.5967 (±0.0138)	0.6054 (±0.0137)	0.5833 (±0.0139)	0.5927 (±0.0138)

Conclusion

X-EcoMLA demonstrates an efficient technique to upcycle pre-trained Transformers into MLA modules to compress KV cache. 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{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}
}