Zebra-Llama: Towards Extremely Efficient Hybrid Models

• URL: http://arxiv.org/abs/2505.17272v1
• Date: Thu, 22 May 2025 20:39:57 GMT
• Title: Zebra-Llama: Towards Extremely Efficient Hybrid Models
• Authors: Mingyu Yang, Mehdi Rezagholizadeh, Guihong Li, Vikram Appia, Emad Barsoum,
• Abstract summary: We propose a practical and scalable alternative to large language models (LLMs)<n>Zebra-Llama combines State Space Models (SSMs) and Multi-head Latent Attention (MLA) layers to efficiently transfer knowledge from pre-trained Transformers.<n>Zebra-Llama Transformer-level accuracy with near-SSM efficiency using only 7-11B training tokens (compared to trillions of tokens required for pre-training) and an 8B teacher.
• Score: 23.023849840907594
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: With the growing demand for deploying large language models (LLMs) across diverse applications, improving their inference efficiency is crucial for sustainable and democratized access. However, retraining LLMs to meet new user-specific requirements is prohibitively expensive and environmentally unsustainable. In this work, we propose a practical and scalable alternative: composing efficient hybrid language models from existing pre-trained models. Our approach, Zebra-Llama, introduces a family of 1B, 3B, and 8B hybrid models by combining State Space Models (SSMs) and Multi-head Latent Attention (MLA) layers, using a refined initialization and post-training pipeline to efficiently transfer knowledge from pre-trained Transformers. Zebra-Llama achieves Transformer-level accuracy with near-SSM efficiency using only 7-11B training tokens (compared to trillions of tokens required for pre-training) and an 8B teacher. Moreover, Zebra-Llama dramatically reduces KV cache size -down to 3.9%, 2%, and 2.73% of the original for the 1B, 3B, and 8B variants, respectively-while preserving 100%, 100%, and >97% of average zero-shot performance on LM Harness tasks. Compared to models like MambaInLLaMA, X-EcoMLA, Minitron, and Llamba, Zebra-Llama consistently delivers competitive or superior accuracy while using significantly fewer tokens, smaller teachers, and vastly reduced KV cache memory. Notably, Zebra-Llama-8B surpasses Minitron-8B in few-shot accuracy by 7% while using 8x fewer training tokens, over 12x smaller KV cache, and a smaller teacher (8B vs. 15B). It also achieves 2.6x-3.8x higher throughput (tokens/s) than MambaInLlama up to a 32k context length. We will release code and model checkpoints upon acceptance.

Related papers

• MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling [80.48332380100915]
  MiniCPM-SALA is a hybrid model that integrates the high-fidelity long-context modeling of sparse attention with the global efficiency of linear attention.<n>On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens.
  arXiv  Detail & Related papers  (2026-02-12T09:37:05Z)
• Diffusion Language Models are Super Data Learners [61.721441061210896]
  When unique data is limited, diffusion language models (DLMs) consistently surpass autoregressive (AR) models by training for more epochs.<n>We attribute the gains to three compounding factors: (1) any-order modeling, (2) super-dense compute from iterative bidirectional denoising, and (3) built-in Monte Carlo augmentation.
  arXiv  Detail & Related papers  (2025-11-05T08:17:42Z)
• SpikingBrain: Spiking Brain-inspired Large Models [42.41339012839023]
  SpikingBrain is a family of brain-inspired models for efficient long-context training and inference.<n>We develop two models: SpikingBrain-7B, a linear LLM, and SpikingBrain-76B, a hybrid-linear MoE LLM.<n>Our models significantly improve long-sequence training efficiency and deliver inference with (partially) constant memory and event-driven spiking behavior.
  arXiv  Detail & Related papers  (2025-09-05T17:34:00Z)
• OverFill: Two-Stage Models for Efficient Language Model Decoding [68.68408155020568]
  Large language models (LLMs) excel across diverse tasks but face significant deployment challenges due to high inference costs.<n>We propose OverFill, which decouples prefill and decode stages to optimize accuracy-efficiency tradeoffs.<n>Our 3B-to-1B OverFill configuration outperforms 1B pruned models by 83.2%, while the 8B-to-3B configuration improves over 3B pruned models by 79.2% on average.
  arXiv  Detail & Related papers  (2025-08-11T20:07:34Z)
• MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttraining [60.02032710118597]
  We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages.<n>MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed.<n>The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini.
  arXiv  Detail & Related papers  (2025-05-12T14:30:11Z)
• X-EcoMLA: Upcycling Pre-Trained Attention into MLA for Efficient and Extreme KV Compression [23.023849840907594]
  Multi-head latent attention (MLA) is designed to optimize KV cache memory through low-rank key-value joint compression.<n>We show that our proposed method can effectively compress the KV cache while preserving the performance on the benchmarks.
  arXiv  Detail & Related papers  (2025-03-14T06:49:37Z)
• Lillama: Large Language Models Compression via Low-Rank Feature Distillation [8.090496457850852]
  Lillama is a compression method that distills activations with low-rank weights.<n>It compresses Mixtral-8x7B within minutes on a single A100 GPU, removing 10 billion parameters while retaining over 95% of its original performance.<n>It generalizes well to non-transformer architectures, compressing Mamba-3B by 20% while maintaining 99% performance.
  arXiv  Detail & Related papers  (2024-12-21T18:04:01Z)
• Llama 3 Meets MoE: Efficient Upcycling [1.8337958765930928]
  We present an efficient training recipe leveraging pre-trained dense checkpoints, training an 8-Expert Top-2 MoE model from Llama 3-8B with less than $1%$ of typical pre-training compute.<n>Our approach enhances downstream performance on academic benchmarks, achieving a $textbf2%$ improvement in 0-shot accuracy on MMLU.<n>We also integrate online upcycling in NeMo for seamless use of pre-trained weights, enabling cost-effective development of high-capacity MoE models.
  arXiv  Detail & Related papers  (2024-12-13T08:22:19Z)
• The Mamba in the Llama: Distilling and Accelerating Hybrid Models [76.64055251296548]
  We show how to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers with academic GPU resources.<n>The resulting hybrid model achieves performance comparable to the original Transformer in chat benchmarks.<n>We also introduce a hardware-aware speculative decoding algorithm that accelerates the inference speed of Mamba and hybrid models.
  arXiv  Detail & Related papers  (2024-08-27T17:56:11Z)
• Compact Language Models via Pruning and Knowledge Distillation [61.56557874432008]
  Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch. Deriving 8B and 4B models from an already pretrained 15B model using our approach requires up to 40x fewer training tokens per model compared to training from scratch.
  arXiv  Detail & Related papers  (2024-07-19T21:47:57Z)
• An Empirical Study of Mamba-based Language Models [69.74383762508805]
  Selective state-space models (SSMs) like Mamba overcome some shortcomings of Transformers. We present a direct comparison between 8B-context Mamba, Mamba-2, and Transformer models trained on the same datasets. We find that the 8B Mamba-2-Hybrid exceeds the 8B Transformer on all 12 standard tasks.
  arXiv  Detail & Related papers  (2024-06-12T05:25:15Z)
• GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection [133.45193150403537]
  Training Large Language Models (LLMs) presents significant memory challenges due to the growing size of weights and GPU states. In this work, we propose Gradient Low-Rank Projection (GaLore) as a memory-efficient training strategy. Our 8-bit GaLore further reduces memory by up to 82.5% and total training memory by 63.3%, compared to a BF16 baseline.
  arXiv  Detail & Related papers  (2024-03-06T07:29:57Z)
• FP8-LM: Training FP8 Large Language Models [47.17804713425323]
  In this paper, we propose a new FP8 automatic mixed-precision framework for training large language models. Experiment results show that, during the training of GPT-175B model on H100 GPU platform, our FP8 mixed-precision training framework not only achieved a remarkable 39% reduction in real memory usage but also ran 75% faster than the widely adopted BF16 framework.
  arXiv  Detail & Related papers  (2023-10-27T17:59:51Z)
• DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing [117.41016786835452]
  This paper presents a new pre-trained language model, DeBERTaV3, which improves the original DeBERTa model. vanilla embedding sharing in ELECTRA hurts training efficiency and model performance. We propose a new gradient-disentangled embedding sharing method that avoids the tug-of-war dynamics.
  arXiv  Detail & Related papers  (2021-11-18T06:48:00Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.