FFN Fusion: Rethinking Sequential Computation in Large Language Models

• URL: http://arxiv.org/abs/2503.18908v1
• Date: Mon, 24 Mar 2025 17:20:35 GMT
• Title: FFN Fusion: Rethinking Sequential Computation in Large Language Models
• Authors: Akhiad Bercovich, Mohammad Dabbah, Omri Puny, Ido Galil, Amnon Geifman, Yonatan Geifman, Izhak Golan, Ehud Karpas, Itay Levy, Zach Moshe, Najeeb Nabwani, Tomer Ronen, Itamar Schen, Elad Segal, Ido Shahaf, Oren Tropp, Ran Zilberstein, Ran El-Yaniv,
• Abstract summary: We introduce FFN Fusion, an architectural optimization technique that reduces sequential computation in large language models.<n>We develop a principled methodology for identifying and fusing such sequences, transforming them into parallel operations.<n>Applying these techniques to Llama-3.1-405B-Instruct, we create an efficient and soon-to-be publicly available model that achieves a 1.71X speedup in inference latency and 35X lower per-token cost.
• Score: 16.8637819797503
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We introduce FFN Fusion, an architectural optimization technique that reduces sequential computation in large language models by identifying and exploiting natural opportunities for parallelization. Our key insight is that sequences of Feed-Forward Network (FFN) layers, particularly those remaining after the removal of specific attention layers, can often be parallelized with minimal accuracy impact. We develop a principled methodology for identifying and fusing such sequences, transforming them into parallel operations that significantly reduce inference latency while preserving model behavior. Applying these techniques to Llama-3.1-405B-Instruct, we create Llama-Nemotron-Ultra-253B-Base (Ultra-253B-Base), an efficient and soon-to-be publicly available model that achieves a 1.71X speedup in inference latency and 35X lower per-token cost while maintaining strong performance across benchmarks. Through extensive experiments on models from 49B to 253B parameters, we demonstrate that FFN Fusion becomes increasingly effective at larger scales and can complement existing optimization techniques like quantization and pruning. Most intriguingly, we find that even full transformer blocks containing both attention and FFN layers can sometimes be parallelized, suggesting new directions for neural architecture design.

Related papers

• iFlame: Interleaving Full and Linear Attention for Efficient Mesh Generation [49.8026360054331]
  iFlame is a novel transformer-based network architecture for mesh generation.<n>We propose an interleaving autoregressive mesh generation framework that combines the efficiency of linear attention with the expressive power of full attention mechanisms.<n>Our results indicate that the proposed interleaving framework effectively balances computational efficiency and generative performance.
  arXiv  Detail & Related papers  (2025-03-20T19:10:37Z)
• Task-Oriented Feature Compression for Multimodal Understanding via Device-Edge Co-Inference [49.77734021302196]
  We propose a task-oriented feature compression (TOFC) method for multimodal understanding in a device-edge co-inference framework.<n>To enhance compression efficiency, multiple entropy models are adaptively selected based on the characteristics of the visual features.<n>Results show that TOFC achieves up to 60% reduction in data transmission overhead and 50% reduction in system latency.
  arXiv  Detail & Related papers  (2025-03-17T08:37:22Z)
• LESA: Learnable LLM Layer Scaling-Up [57.0510934286449]
  Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive.<n>Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones.<n>We propose textbfLESA, a novel learnable method for depth scaling-up.
  arXiv  Detail & Related papers  (2025-02-19T14:58:48Z)
• LeRF: Learning Resampling Function for Adaptive and Efficient Image Interpolation [64.34935748707673]
  Recent deep neural networks (DNNs) have made impressive progress in performance by introducing learned data priors. We propose a novel method of Learning Resampling (termed LeRF) which takes advantage of both the structural priors learned by DNNs and the locally continuous assumption. LeRF assigns spatially varying resampling functions to input image pixels and learns to predict the shapes of these resampling functions with a neural network.
  arXiv  Detail & Related papers  (2024-07-13T16:09:45Z)
• Building on Efficient Foundations: Effectively Training LLMs with Structured Feedforward Layers [16.253898272659242]
  State-of-the-art results in large language models (LLMs) often rely on scale, which becomes computationally expensive. Our study focuses on transformer-based LLMs, specifically targeting the computationally intensive feedforward networks (FFNs) We show that wide and structured networks can utilize training FLOPs more efficiently, with fewer parameters and lower loss than dense models at their optimal trade-off.
  arXiv  Detail & Related papers  (2024-06-24T08:43:21Z)
• Bypass Back-propagation: Optimization-based Structural Pruning for Large Language Models via Policy Gradient [57.9629676017527]
  We propose an optimization-based structural pruning on Large-Language Models. We learn the pruning masks in a probabilistic space directly by optimizing the loss of the pruned model. Our method operates for 2.7 hours with around 35GB memory for the 13B models on a single A100 GPU.
  arXiv  Detail & Related papers  (2024-06-15T09:31:03Z)
• Lean Attention: Hardware-Aware Scalable Attention Mechanism for the Decode-Phase of Transformers [4.674454841332859]
  Transformer-based models have emerged as one of the most widely used architectures for natural language processing.<n>These huge models are memory hungry and incur significant inference latency even on cutting edge AI-accelerators.<n>We propose LeanAttention, a scalable technique of computing self-attention for the token-generation phase.
  arXiv  Detail & Related papers  (2024-05-17T00:52:39Z)
• Parallelized Spatiotemporal Binding [47.67393266882402]
  We introduce Parallelizable Spatiotemporal Binder or PSB, the first temporally-parallelizable slot learning architecture for sequential inputs. Unlike conventional RNN-based approaches, PSB produces object-centric representations, known as slots, for all time-steps in parallel. Compared to the state-of-the-art, our architecture demonstrates stable training on longer sequences, achieves parallelization that results in a 60% increase in training speed, and yields performance that is on par with or better on unsupervised 2D and 3D object-centric scene decomposition and understanding.
  arXiv  Detail & Related papers  (2024-02-26T23:16:34Z)
• SPEED: Speculative Pipelined Execution for Efficient Decoding [35.45955948053644]
  We propose SPEED, which improves inference efficiency by speculatively executing multiple future tokens in parallel with the current token. For Transformer decoders that employ parameter sharing, the memory operations for the tokens executing in parallel can be amortized. We demonstrate the efficiency of our method in terms of latency reduction relative to model accuracy and demonstrate how speculation allows for training deeper decoders with parameter sharing with minimal runtime overhead.
  arXiv  Detail & Related papers  (2023-10-18T16:07:01Z)
• Scaling Studies for Efficient Parameter Search and Parallelism for Large Language Model Pre-training [2.875838666718042]
  We focus on parallel and distributed machine learning algorithm development, specifically for optimizing the data processing and pre-training of a set of 5 encoder-decoder LLMs. We performed a fine-grained study to quantify the relationships between three ML methods, specifically exploring Microsoft DeepSpeed Zero Redundancy stages.
  arXiv  Detail & Related papers  (2023-10-09T02:22:00Z)
• TransNormerLLM: A Faster and Better Large Language Model with Improved TransNormer [34.790081960470964]
  We present TransNormerLLM, the first linear attention-based Large Language Model (LLM) We make advanced modifications that include positional embedding, linear attention acceleration, gating mechanisms, tensor normalization, and inference acceleration and stabilization. We validate our model design through a series of ablations and train models with sizes of 385M, 1B, and 7B on our self-collected corpus.
  arXiv  Detail & Related papers  (2023-07-27T16:45:33Z)
• Efficiently Scaling Transformer Inference [8.196193683641582]
  We study the problem of efficient generative inference for Transformer models, in one of its most challenging settings. We develop a simple analytical model for inference efficiency to select the best multi-dimensional partitioning techniques optimized for TPU v4 slices. We achieve a low-batch-size latency of 29ms per token during generation (using int8 weight quantization) and a 76% MFU during large-batch-size processing of input tokens.
  arXiv  Detail & Related papers  (2022-11-09T18:50:38Z)
• Belief Propagation Reloaded: Learning BP-Layers for Labeling Problems [83.98774574197613]
  We take one of the simplest inference methods, a truncated max-product Belief propagation, and add what is necessary to make it a proper component of a deep learning model. This BP-Layer can be used as the final or an intermediate block in convolutional neural networks (CNNs) The model is applicable to a range of dense prediction problems, is well-trainable and provides parameter-efficient and robust solutions in stereo, optical flow and semantic segmentation.
  arXiv  Detail & Related papers  (2020-03-13T13:11:35Z)

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.