Swapnil Gandhi

Swapnil Gandhi, Christos Kozyrakis

As large language models scale, training them requires thousands of GPUs over extended durations--making frequent failures an inevitable reality. While checkpointing remains the primary fault-tolerance mechanism, existing methods fall short when applied to Mixture-of-Experts (MoE) models. Due to their substantially larger training state, MoE models exacerbate checkpointing overheads, often causing costly stalls or prolonged recovery that severely degrade training efficiency. We present MoEvement, a distributed, in-memory checkpointing system tailored for MoE models. MoEvement is built on three key ideas: (1) sparse checkpointing, which incrementally snapshots subsets of experts across iterations to reduce overhead; (2) a sparse-to-dense checkpoint conversion mechanism that incrementally reconstructs consistent dense checkpoints from sparse snapshots; and (3) upstream logging of activations and gradients at pipeline-stage boundaries, enabling localized recovery without re-executing unaffected workers. Evaluations across diverse MoE models with up to 64 experts show that MoEvement reduces checkpointing overhead by up to $4\times$ and recovery overhead by up to $31\times$ compared to state-of-the-art approaches, sustaining ETTR $\ge 0.94$ even under frequent failures (MTBF as low as 10 minutes) and delivering up to $8\times$ overall training speedup, all without compromising synchronous training semantics. Overall, MoEvement offers a robust and scalable fault-tolerance solution for the next generation of sparsely activated models.

Jungwoo Kim, Rubens Lacouture, Genghan Zhang et al.

Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models (LLMs). However, the execution characteristics of MoE inference are changing rapidly and increasingly mismatch the assumptions underlying existing Processing-in-Memory (PIM) systems. Prior PIM systems for LLMs rely on static rules to offload memory-bound operations to PIM, without accounting for the combined effects of load imbalance and inter-GPU communication. Meanwhile, modern MoE models activate fewer experts out of increasingly many, creating a bimodal expert distribution: a small set of experts receives many tokens, while a long tail of experts receives only one or a few. We identify a trend in modern MoE models toward increasingly bimodal token-to-expert distributions, quantify the resulting disparity in arithmetic intensity across experts, and show that this disparity dramatically reduces the efficiency of state-of-the-art PIM systems for LLMs. To address this problem, we propose a scheduler for serving MoE models on multi-GPU systems with attached HBM-PIM stacks. Our scheduler partitions expert execution between GPU and PIM based on runtime token-to-expert distributions, while jointly considering interconnect overhead, memory bandwidth, GPU throughput, and PIM throughput. Moreover, we propose Sieve, a runtime framework that employs the scheduler to coordinate execution across GPUs and their attached HBM-PIM stacks. Sieve overlaps GPU computation, PIM computation, and intra- and inter-device communication while preserving cross-device dependencies induced by expert parallelism. Sieve is evaluated on our cycle-accurate simulator based on Ramulator 2.0. Compared to state-of-the-art PIM systems for MoE, Sieve improves both throughput and interactivity by 1.3x, 1.3x, and 1.6x on Qwen3.5-397B-A17B, GPT-OSS-120B, and Qwen3-30B-A3B, respectively.

Swapnil Gandhi, Siva Hari, William J. Dally et al.

Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the branch externality and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce TAPER, a per-step admission controller that treats extra branches as opportunistic work, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, TAPER improves goodput by $1.77\times$ over IRP-Off and by $1.48\times$ over IRP-Eager, while maintaining over $95\%$ SLO attainment.

Swapnil Gandhi, Mark Zhao, Athinagoras Skiadopoulos et al.

Training large Deep Neural Network (DNN) models requires thousands of GPUs over the course of several days or weeks. At this scale, failures are frequent and can have a big impact on training throughput. Utilizing spare GPU servers to mitigate performance loss becomes increasingly costly as model sizes grow. ReCycle is a system designed for efficient DNN training in the presence of failures, without relying on spare servers. It exploits the inherent functional redundancy in distributed training systems -- where servers across data-parallel groups store the same model parameters -- and pipeline schedule bubbles within each data-parallel group. When servers fails, ReCycle dynamically re-routes micro-batches to data-parallel peers, allowing for uninterrupted training despite multiple failures. However, this re-routing can create imbalances across pipeline stages, leading to reduced training throughput. To address this, ReCycle introduces two key optimizations that ensure re-routed micro-batches are processed within the original pipeline schedule's bubbles. First, it decouples the backward pass into two phases: one for computing gradients for the input and another for calculating gradients for the parameters. Second, it avoids synchronization across pipeline stages by staggering the optimizer step. Together, these optimizations enable adaptive pipeline schedules that minimize or even eliminate training throughput degradation during failures. We describe a prototype for ReCycle and show that it achieves high training throughput under multiple failures, outperforming recent proposals for fault-tolerant training such as Oobleck and Bamboo by up to $1.46\times$ and $1.64\times$, respectively.

Athinagoras Skiadopoulos, Mark Zhao, Swapnil Gandhi et al.

Mixture-of-Experts (MoE) models have become a widely-adopted solution to continue scaling model sizes without a corresponding linear increase in compute. During MoE model training, each input token is dynamically routed to a subset of experts -- sparsely-activated feed-forward networks -- within each transformer layer. The distribution of tokens assigned to each expert varies widely and rapidly over the course of training. To handle the wide load imbalance across experts, current systems are forced to either drop tokens assigned to popular experts, degrading convergence, or frequently rebalance resources allocated to each expert based on popularity, incurring high state migration overheads. To break this performance-accuracy tradeoff, we introduce SYMI, an adaptive MoE training system. The key insight of SYMI is to decouple the placement of expert parameters from their large optimizer state. SYMI statically partitions the optimizer of each expert across all training nodes. Meanwhile, SYMI dynamically adjusts the placement of expert parameters by repurposing existing weight updates, avoiding migration overheads. In doing so, SYMI right-sizes the GPU resources allocated to each expert, on a per-iteration basis, with minimal overhead. Compared to state-of-the-art MoE training systems, DeepSpeed and FlexMoE, SYMI is able to achieve a 30.5% and 25.9% faster time-to-convergence, respectively.