Ladder-residual: parallelism-aware architecture for accelerating large model inference with communication overlapping
- URL: http://arxiv.org/abs/2501.06589v4
- Date: Fri, 07 Feb 2025 08:23:57 GMT
- Title: Ladder-residual: parallelism-aware architecture for accelerating large model inference with communication overlapping
- Authors: Muru Zhang, Mayank Mishra, Zhongzhu Zhou, William Brandon, Jue Wang, Yoon Kim, Jonathan Ragan-Kelley, Shuaiwen Leon Song, Ben Athiwaratkun, Tri Dao,
- Abstract summary: We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models.
Applying Ladder Residual to all its layers can achieve 29% end-to-end wall clock speed up at inference time with TP sharding over 8 devices.
We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline.
- Score: 36.71999572939612
- License:
- Abstract: Large language model inference is both memory-intensive and time-consuming, often requiring distributed algorithms to efficiently scale. Various model parallelism strategies are used in multi-gpu training and inference to partition computation across multiple devices, reducing memory load and computation time. However, using model parallelism necessitates communication of information between GPUs, which has been a major bottleneck and limits the gains obtained by scaling up the number of devices. We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models that enables straightforward overlapping that effectively hides the latency of communication. Our insight is that in addition to systems optimization, one can also redesign the model architecture to decouple communication from computation. While Ladder Residual can allow communication-computation decoupling in conventional parallelism patterns, we focus on Tensor Parallelism in this paper, which is particularly bottlenecked by its heavy communication. For a Transformer model with 70B parameters, applying Ladder Residual to all its layers can achieve 29% end-to-end wall clock speed up at inference time with TP sharding over 8 devices. We refer the resulting Transformer model as the Ladder Transformer. We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline. We also show that it is possible to convert parts of the Llama-3.1 8B model to our Ladder Residual architecture with minimal accuracy degradation by only retraining for 3B tokens. We release our code for training and inference for easier replication of experiments.
Related papers
- Kraken: Inherently Parallel Transformers For Efficient Multi-Device Inference [8.527031391688283]
Kraken is an evolution of the standard Transformer architecture for efficient inference on multi-device systems.
When trained on OpenWebText, Kraken models reach a similar perplexity as standard Transformers.
When tested on the SuperGLUE benchmark, Kraken speeds up Time To First Token by a mean of 35.6% across a range of model sizes.
arXiv Detail & Related papers (2024-08-14T20:24:03Z) - LMUFormer: Low Complexity Yet Powerful Spiking Model With Legendre
Memory Units [5.830814457423021]
Transformer models have demonstrated high accuracy in numerous applications but have high complexity and lack sequential processing capability.
We show how architectural modifications to a recurrent model can help push its performance toward Transformer models.
We present a spiking version of this architecture, which introduces the benefit of states within the patch embedding and channel mixer modules.
arXiv Detail & Related papers (2024-01-20T01:10:18Z) - Retentive Network: A Successor to Transformer for Large Language Models [91.6652200825638]
We propose Retentive Network (RetNet) as a foundation architecture for large language models.
We theoretically derive the connection between recurrence and attention.
Experimental results on language modeling show that RetNet achieves favorable scaling results, parallel training, low-cost deployment, and efficient inference.
arXiv Detail & Related papers (2023-07-17T16:40:01Z) - Winner-Take-All Column Row Sampling for Memory Efficient Adaptation of Language Model [89.8764435351222]
We propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance.
Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones.
arXiv Detail & Related papers (2023-05-24T15:52:08Z) - RWKV: Reinventing RNNs for the Transformer Era [54.716108899349614]
We propose a novel model architecture that combines the efficient parallelizable training of transformers with the efficient inference of RNNs.
We scale our models as large as 14 billion parameters, by far the largest dense RNN ever trained, and find RWKV performs on par with similarly sized Transformers.
arXiv Detail & Related papers (2023-05-22T13:57:41Z) - SWARM Parallelism: Training Large Models Can Be Surprisingly
Communication-Efficient [69.61083127540776]
Deep learning applications benefit from using large models with billions of parameters.
Training these models is notoriously expensive due to the need for specialized HPC clusters.
We consider alternative setups for training large models: using cheap "preemptible" instances or pooling existing resources from multiple regions.
arXiv Detail & Related papers (2023-01-27T18:55:19Z) - LiteTransformerSearch: Training-free On-device Search for Efficient
Autoregressive Language Models [34.673688610935876]
We show that the latency and perplexity pareto-frontier can be found without need for any model training.
We evaluate our method, dubbed Lightweight Transformer Search (LTS), on diverse devices.
We show that the perplexity of Transformer-XL can be achieved with up to 2x lower latency.
arXiv Detail & Related papers (2022-03-04T02:10:43Z) - Layered gradient accumulation and modular pipeline parallelism: fast and
efficient training of large language models [0.0]
We analyse the shortest possible training time for different configurations of distributed training.
We introduce two new methods, textitlayered gradient accumulation and textitmodular pipeline parallelism, which together cut the shortest training time by half.
arXiv Detail & Related papers (2021-06-04T19:21:49Z) - Hardware Beyond Backpropagation: a Photonic Co-Processor for Direct
Feedback Alignment [26.65651157173834]
We present a photonic accelerator for Direct Feedback Alignment, able to compute random projections with trillions of parameters.
This is a significant step towards building scalable hardware, able to go beyond backpropagation.
arXiv Detail & Related papers (2020-12-11T14:20:45Z) - One-step regression and classification with crosspoint resistive memory
arrays [62.997667081978825]
High speed, low energy computing machines are in demand to enable real-time artificial intelligence at the edge.
One-step learning is supported by simulations of the prediction of the cost of a house in Boston and the training of a 2-layer neural network for MNIST digit recognition.
Results are all obtained in one computational step, thanks to the physical, parallel, and analog computing within the crosspoint array.
arXiv Detail & Related papers (2020-05-05T08:00:07Z)
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.