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. |
Similar Papers
Memory-Efficient Structured Backpropagation for On-Device LLM Fine-Tuning (2026.acl-industry)
Copied to clipboard
| 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. |
Memory-Efficient Fine-Tuning of Transformers via Token Selection (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning require caching of intermediate activations to update weights during the backward pass. |
| Approach: | They develop a method to reduce memory usage in fine-tuning of transformers by backpropagating through just a subset of input tokens. |
| Outcome: | The proposed method reduces memory usage and memory footprint on large transformer models . it can be easily combined with existing methods like LoRA, reducing memory cost . |
SparseGrad: A Selective Method for Efficient Fine-tuning of MLP Layers (2024.emnlp-main)
Copied to clipboard
Viktoriia Chekalina, Anna Rudenko, Gleb Mezentsev, Aleksandr Mikhalev, Alexander Panchenko, Ivan Oseledets
| Challenge: | High-performance methods for parameter-efficient fine-tuning (PEFT) typically work with Attention blocks and overlook dense MLP blocks, which contain about half of the model parameters. |
| Approach: | They propose a selective PEFT method that performs well on MLP blocks by converting layer gradients into a sparse structure and reducing the number of updated parameters. |
| Outcome: | The proposed method outperforms LoRA and MeProp, robust state-of-the-art PEFT approaches. |
Jump to Conclusions: Short-Cutting Transformers with Linear Transformations (2024.lrec-main)
Copied to clipboard
| Challenge: | Transformer-based language models create hidden representations of inputs at every layer, but only use final-layer representations for prediction. |
| Approach: | They propose a method for casting hidden representations as final representations, bypassing transformer computation in-between. |
| Outcome: | The proposed method produces more accurate predictions from hidden layers across various model scales, architectures, and data distributions. |
Backward Lens: Projecting Language Model Gradients into the Vocabulary Space (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent interpretability methods project weights and hidden states obtained from the forward pass to the models’ vocabularies, helping to uncover how information flows within LMs. |
| Approach: | They propose to cast a gradient matrix as a low-rank linear combination of forward and backward passes’ inputs and then to project these gradients into vocabulary items. |
| Outcome: | The proposed method can be cast as a low-rank linear combination of forward and backward passes’ inputs and project these gradients into vocabulary items. |
Rethinking Network Pruning – under the Pre-train and Fine-tune Paradigm (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing pruning results on benchmark transformers, such as BERT, are not as remarkable as those of convolutional neural networks. |
| Approach: | They propose to apply a knowledge-aware pruning process to transformer-based pre-trained language models to reduce model size and model weight. |
| Outcome: | The proposed pruning method outperforms the leading competitors with a 20-times weight/FLOPs compression and neglectable loss in prediction accuracy. |
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. |
FiRST: Finetuning Router-Selective Transformers for Input-Adaptive Latency Reduction (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to improve latency via skipping layers have limitations . fiRST is a model-agnostic framework that reduces inference latency while maintaining quality . |
| Approach: | They propose a model-agnostic framework that skips transformer layers during decoding . it is fully compatible with KV caching, enabling faster decoding while maintaining quality . |
| Outcome: | a new framework reduces inference latency by using layer-specific routers to skip transformer layers during decoding. |
Context Compression for Auto-regressive Transformers with Sentinel Tokens (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing Transformer-based LLMs have limited performance due to complexity of attention module . key-value cache is the major memory footprint and inference latency problem . |
| Approach: | They propose a plug-and-play approach that incrementally compresses token activation into compact ones . they also profile the benefit of context compression on improving the system throughout . |
| Outcome: | The proposed approach reduces memory footprint and inference latency by compressing tokens into compact ones. |
FBS: Modeling Native Parallel Reading inside a Transformer (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing acceleration methods largely patch the autoregressive pipeline and miss core human-reading ingredients. |
| Approach: | They propose a trainable loop that injects a causal loop into Transformers via a 'parafoveal' approach. |
| Outcome: | The proposed model improves quality-efficiency trade-off without increasing parameters . ablations show the three modules are complementary . |