REVIEW 22 cited by
What Matters in Transformers? Not All Attention is Needed
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
What Matters in Transformers? Not All Attention is Needed
read the original abstract
While scaling Transformer-based large language models (LLMs) has demonstrated promising performance across various tasks, it also introduces redundant architectures, posing efficiency challenges for real-world deployment. Despite some recognition of redundancy in LLMs, the variability of redundancy across different architectures in transformers, such as MLP and Attention layers, is under-explored. In this work, we investigate redundancy across different modules within Transformers, including Blocks, MLP, and Attention layers, using a similarity-based metric. Surprisingly, despite the critical role of attention layers in distinguishing transformers from other architectures, we found that a large portion of these layers exhibit excessively high similarity and can be pruned without degrading performance. For instance, Llama-2-70B achieved a 48.4\% speedup with only a 2.4\% performance drop by pruning half of the attention layers. Furthermore, by tracing model checkpoints throughout the training process, we observed that attention layer redundancy is inherent and consistent across training stages. Additionally, we further propose a method that jointly drops Attention and MLP layers, allowing us to more aggressively drop additional layers. For instance, when dropping 31 layers (Attention + MLP), Llama-2-13B still retains 90\% of the performance on the MMLU task. Our work provides valuable insights for future network architecture design. The code is released at: \url{https://github.com/Shwai-He/LLM-Drop}.
Forward citations
Cited by 22 Pith papers
-
SpikeMLLM: Spike-based Multimodal Large Language Models via Modality-Specific Temporal Scales and Temporal Compression
SpikeMLLM is the first spike-based MLLM framework that maintains near-lossless performance under aggressive timestep compression and delivers 9x throughput and 25x power efficiency gains via a custom RTL accelerator.
-
Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?
VLA language backbones show high redundancy on manipulation benchmarks, with half the LLM blocks removable and even two blocks sufficient to recover baseline performance after fine-tuning, unlike vision and action pathways.
-
Beyond FLOPs: Benchmarking Real Inference Acceleration of LLM Pruning under a GEMM-Centric Taxonomy
A GEMM-centric taxonomy and unified benchmark show static depth pruning as the strongest Pareto-optimal baseline for LLM inference acceleration, with the frontier shifting to dynamic depth then static width pruning as...
-
From Layers to Submodules: Rethinking Granularity in Replacement-Based LLM Compression
SubFit enables better LLM compression by fitting residual bypasses to non-contiguously selected submodules, outperforming layer-granularity baselines in accuracy-perplexity trade-offs at 12.5-37.5% sparsity.
-
Structured-Sparse Attention for Entity Tracking with Subquadratic Sequence Complexity
Derives a blockwise resolvent-style attention operator that exploits structured sparsity for subquadratic O(n^{4/3}d) entity tracking while matching dense accuracy.
-
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling
DASH selectively halts stabilized tokens by monitoring layer-wise self-attention deltas, delivering prefill speedups on language and vision tasks without accuracy loss or hardware incompatibility.
-
INCRT: An Incremental Transformer That Determines Its Own Architecture
INCRT grows and prunes transformer attention heads on the fly via a geometric criterion, reaching a provably minimal sufficient architecture that matches or exceeds BERT-base performance on targeted tasks with 3-7 tim...
-
Voltron: Enabling Elastic Multi-Device Execution of LLM Inference for Empowered Edge Intelligence
A framework called Voltron elastically distributes LLM inference across heterogeneous edge devices using layer-wise hybrid parallelism and mixed precision, achieving up to 16.5% higher accuracy than single-device exec...
-
Rethinking Depth Pruning for Vision Transformers: A Heterogeneity-Aware Perspective
Heterogeneity-aware depth pruning of attention and activation layers, guided by a polynomial model-accuracy predictor, delivers up to 1.58× speedup on DeiT-B and 5.19× when combined with width pruning.
-
ScAle: Attention Head Scaling as a Minimal Adapter for Spatial Reasoning in Vision Language Models
ScAle learns scalar coefficients to modulate last-token attention and MLP activations in frozen VLMs, achieving up to 134.1% relative accuracy gains on spatial benchmarks with only 1K parameters.
-
CascadeFormer: Depth-Tapered Transformers Motivated by Gradient Fan-in Asymmetry
CascadeFormer tapers Transformer width with depth based on gradient fan-in asymmetry to match uniform baselines in perplexity while cutting latency 8.6%.
-
Training-Inference Consistent Segmented Execution for Long-Context LLMs
A training-inference consistent segmented execution framework for long-context LLMs matches full-context performance with substantially lower peak memory at very long lengths.
-
Shallow Prefill, Deep Decoding: Efficient Long-Context Inference via Layer-Asymmetric KV Visibility
SPEED uses layer-asymmetric KV visibility to process non-anchor prompt tokens only in lower layers during prefill, achieving near-baseline quality on Llama-3.1-8B with 33% better TTFT and 25% lower active KV memory at...
-
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling
DASH reduces prefill computation in LLMs and vision-language models by halting tokens with low attention-branch residual updates at a chosen layer, achieving 1.7–2× prefill speedups with minimal accuracy loss.
-
Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis
Fixed golden layers for knowledge editing in LLMs can be identified via gradient attribution and generalize across queries and datasets.
-
Resting Neurons, Active Insights: Robustifying Activation Sparsity in LLMs via Spontaneity
SPON adds learnable persistent activation anchors trained via distribution matching to restore LLM accuracy under high activation sparsity by preventing representational distribution shifts.
-
Mixture-of-Control: State-Aware Fine-Tuning for Transformer-based Models
Mixture-of-Control adaptively combines local and global control states in transformer fine-tuning by treating per-block states as experts in a sparse MoE setup to improve cross-block communication while keeping memory...
-
Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity
Cosine similarity poorly predicts performance degradation from layer removal in LLMs, making direct accuracy-drop ablation a more reliable relevance metric.
-
Resting Neurons, Active Insights: Robustifying Activation Sparsity in LLMs via Spontaneity
SPON adds a small set of trainable input-independent activation vectors as representational anchors, trained by distribution matching, to stabilize sparse activation in LLMs and recover performance lost to hidden-stat...
-
RAP: Runtime Adaptive Pruning for LLM Inference
RAP is a reinforcement learning framework for runtime-adaptive pruning of LLMs that jointly optimizes model weights and KV-cache usage under varying memory budgets.
-
NeuronMLP: Efficient LLM Inference via Singular Value Decomposition Compression and Tiling on AWS Trainium
NeuronMLP applies SVD-based compression and Trainium-specific tiling and caching to MLP layers, delivering 1.35x kernel speedup and 1.21x end-to-end inference speedup at 0.05 compression ratio versus AWS NKI baseline.
-
Global-local Spatial-temporal Aware Graph Attention Network for Network Traffic Forecasting
GLSTaGAT is a spatial-temporal graph attention network using data-driven fusion graphs, global-local blocks, node normalization, and a transformer encoder to outperform baselines on real-world network traffic datasets.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.