SkipPipe: Partial and Reordered Pipelining Framework for Training LLMs in Heterogeneous Networks

TL;DR

SkipPipe addresses the bottleneck of sequential pipeline execution in distributed LLM training on heterogeneous networks by introducing partial and reordered pipelining. It combines node-to-stage allocation with a MAPF-based path scheduler to skip and reorder stages for individual microbatches, subject to convergence and throughput constraints. Empirical results across LLaMa models (500M–8B) show up to 55% reductions in iteration time while preserving convergence, and paradoxically improved robustness to layer omissions during inference, enabling early-exit and fault-tolerant strategies. This work offers a practical framework for cost-effective, scalable LLM training and robust inference in heterogeneous compute environments, with code available at https://github.com/gensyn-ai/skippipe.

Abstract

Data and pipeline parallelism are ubiquitous for training of Large Language Models (LLM) on distributed nodes. Driven by the need for cost-effective training, recent work explores efficient communication arrangement for end to end training. Motivated by LLM's resistance to layer skipping and layer reordering, in this paper, we explore stage (several consecutive layers) skipping in pipeline training, and challenge the conventional practice of sequential pipeline execution. We derive convergence and throughput constraints (guidelines) for pipelining with skipping and swapping pipeline stages. Based on these constraints, we propose SkipPipe, the first partial pipeline framework to reduce the end-to-end training time for LLMs while preserving the convergence. The core of SkipPipe is a path scheduling algorithm that optimizes the paths for individual microbatches and reduces idle time (due to microbatch collisions) on the distributed nodes, complying with the given stage skipping ratio. We extensively evaluate SkipPipe on LLaMa models from 500M to 8B parameters on up to 20 nodes. Our results show that SkipPipe reduces training iteration time by up to $55\%$ compared to full pipeline. Our partial pipeline training also improves resistance to layer omission during inference, experiencing a drop in perplexity of only $7\%$ when running only half the model. Our code is available at https://github.com/gensyn-ai/skippipe.

Figures (6)

• Figure 1: An example of partial pipeline parallelism scheduling where each colored (solid or dashed) path represents a different microbatch. Each node in stage 0 sends out 2 microbatches, the first in solid, the second in dashed. Green backgrounds show the forward pass, while light orange - the backwards pass. For better visualisation, we omit the loss and deembedding computations. We place arrows to show the prioritisation of the microbatches from forward to backward pass within the same node. An example of a collision can be seen on node 7 during the forward pass, which subsequently delays the execution of the solid blue microbatch because of the dashed yellow microbatch.
• Figure 2: Convergence of LLaMa-30M model. The validation loss is calculated for the whole model for every 50th iteration.
• Figure 3: Time per iteration with different strategies. We analyse four schedulers with two skip percentages (25% and 33%) and three token numbers (4K, 8K and 16K). SkipPipe is compared with: DT-FM$^*$ representing the compensated results for the baseline with no skips, DT-FM-skip uses node arrangement of DT-FM and skips $k\%$ with additional constraints (see Appendix \ref{['qrps']}), SkipPipe (no TC2) is our scheduler without TC2.
• Figure 4: Convergence of validation loss with $33\%$ skip rate, $25\%$ skip rate, and $0\%$ skips (full model).
• Figure 5: The validation loss of LLaMa-7B under % of random skipping in pipeline training.