Challenge: Existing approaches for fine-tuning large language models require a trade-off between exact gradients with high memory and low memory with noisy estimates (MeZO).
Approach: They propose a method which derivates gradients from LoRA's low-rank structure and manually deriving backward passes to exploit the low-level structure.
Outcome: The proposed method reduces peak memory from 361MB to 136MB for Qwen2.5-0.5B, enabling fine-tuning scenarios previously infeasible on memory-constrained devices.

Similar Papers

Memory-Efficient Backpropagation for Fine-Tuning LLMs on Resource-Constrained Mobile Devices (2025.emnlp-industry)

Copied to clipboard

Challenge: Existing work on memory-efficient on-device fine-tuning of large language models with backpropagation has focused on approximating gradients with zeroth-order optimization (ZO).
Approach: They propose a memory-efficient implementation of backpropagation on mobile devices that allows flexible trade-offs between memory usage and compute time while converging faster.
Outcome: The proposed method can fine-tune LLMs with backpropagation using less than 1GB of memory while achieving better performance than the baseline.
Forward Knows Efficient Backward Path: Saliency-Guided Memory-Efficient Fine-tuning of Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: a number of fine-tuning approaches are available to improve performance of large language models.
Approach: They propose a memory-efficient method to minimize memory associated with cached intermediate activations.
Outcome: The proposed method minimizes memory associated with cached intermediate activations while preserving accuracy.
META-LORA: Memory-Efficient Sample Reweighting for Fine-Tuning Large Language Models (2025.coling-main)

Copied to clipboard

Challenge: Supervised fine-tuning (SFT) is widely adopted for tailoring large language models (LLMs) to specific downstream tasks.
Approach: They propose a memory-efficient method for automatic sample reweighting that learns to re-weight fine-tuning samples by minimizing the loss on a small, high-quality validation set.
Outcome: Meta-LoRA learns to reweight fine-tuning samples by minimizing the loss on a small, high-quality validation set through an end-to-end bi-level optimization framework based on meta-learning.
Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models (2024.acl-long)

Copied to clipboard

Challenge: Existing methods to finetun large language models (LLMs) only update a small number of trainable parameters, or attempt to reduce the memory footprint during the training phase of the finetune process.
Approach: They propose quantized side tuing (QST) which quantizes an LLM’s model weights into 4-bit to reduce the memory footprint of the original weights.
Outcome: The proposed method reduces the memory footprint of the model weights, optimizer states, and intermediate activations while reducing the memory requirements.
Acceleration of Backpropagation in Linear Layers of Transformer Models Based on Gradient Structure (2026.eacl-srw)

Copied to clipboard

Challenge: a sparsity-exploiting backward pass is a memory-efficient way to accelerate LLM fine-tuning.
Approach: They propose a method that exploits padding-induced gradient sparsity to accelerate backward computation.
Outcome: The proposed method achieves a backward pass speedup of 2.15x on GLUE and 1.99x on reasoning benchmarks while maintaining memory usage identical to the regular PyTorch fine-tuning.
LoRAPrune: Structured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning (2024.findings-acl)

Copied to clipboard

Challenge: Low-rank adaption (LoRA) is a low-level pruning method that can be expensive and slow to deploy.
Approach: They propose a low-rank adaption pruning framework that provides an accurate structured pruned model in a memory-efficient manner.
Outcome: The proposed pruning framework reduces perplexity and memory usage by 52.6% on LLaMA and T5 models while reducing memory usage.
MobiZO: Enabling Efficient LLM Fine-Tuning at the Edge via Inference Engines (2025.emnlp-main)

Copied to clipboard

Challenge: Large Language Models (LLMs) are currently pre-trained and fine-tuned on large cloud servers . fine-timing on resource-constrained edge devices presents significant memory and computational demands .
Approach: They propose a resource-efficient fine-tuning framework for LLMs specifically designed for edge devices.
Outcome: Experiments show that MobiZO achieves substantial runtime speedups and memory savings while improving fine-tuning accuracy.
LaMDA: Large Model Fine-Tuning via Spectrally Decomposed Low-Dimensional Adaptation (2024.findings-emnlp)

Copied to clipboard

Challenge: Low-rank adaptation (LoRA) fine-tunes large language models due to its significant reduction in trainable parameters, but its backward updates require storing high-dimensional intermediate activations and optimizer states, requiring high peak GPU memory.
Approach: They propose a low-dimensional adaptation approach to fine-tune large language models which freezes a first projection matrix while introducing a lower-dimensional trainable square matrix.
Outcome: The proposed approach reduces trainable parameters and peak GPU memory footprint while preserving low-dimensional trainable square matrix.
Continual Gradient Low-Rank Projection Fine-Tuning for LLMs (2025.acl-long)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) offers efficiency but constrains the model’s ability to learn new tasks and transfer knowledge due to its low-rank nature and reliance on explicit parameter constraints.
Approach: They propose a training strategy that synergistically combines full and low-rank parameters and jointly updating within a unified low-ranked gradient subspace.
Outcome: Extensive experiments on continual learning benchmarks show that GORP improves performance compared to state-of-the-art approaches.
Bypass Back-propagation: Optimization-based Structural Pruning for Large Language Models via Policy Gradient (2025.acl-long)

Copied to clipboard

Challenge: Recent pruning methods rely on heuristically hand-crafted metrics, leading to suboptimal performance.
Approach: They propose a method that optimizes pruning masks by minimizing back-propagation . they learn an underlying Bernoulli distribution to sample binary pruning mask samples .
Outcome: The proposed method is able to support global and heterogeneous pruning without back-propagation.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations