Sieve dynamically schedules MoE experts across GPU and PIM hardware to handle bimodal token distributions, achieving 1.3x to 1.6x gains in throughput and interactivity over static prior PIM systems on three large models.
super hub Canonical reference
GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
Canonical reference. 78% of citing Pith papers cite this work as background.
107
Pith papers citing it
Background
78% of classified citations
abstract
Neural network scaling has been critical for improving the model quality in many real-world machine learning applications with vast amounts of training data and compute. Although this trend of scaling is affirmed to be a sure-fire approach for better model quality, there are challenges on the path such as the computation cost, ease of programming, and efficient implementation on parallel devices. GShard is a module composed of a set of lightweight annotation APIs and an extension to the XLA compiler. It provides an elegant way to express a wide range of parallel computation patterns with minimal changes to the existing model code. GShard enabled us to scale up multilingual neural machine translation Transformer model with Sparsely-Gated Mixture-of-Experts beyond 600 billion parameters using automatic sharding. We demonstrate that such a giant model can efficiently be trained on 2048 TPU v3 accelerators in 4 days to achieve far superior quality for translation from 100 languages to English compared to the prior art.
hub tools
citation-role summary
background 21
method 4
dataset 2
citation-polarity summary
claims ledger
- abstract Neural network scaling has been critical for improving the model quality in many real-world machine learning applications with vast amounts of training data and compute. Although this trend of scaling is affirmed to be a sure-fire approach for better model quality, there are challenges on the path such as the computation cost, ease of programming, and efficient implementation on parallel devices. GShard is a module composed of a set of lightweight annotation APIs and an extension to the XLA compiler. It provides an elegant way to express a wide range of parallel computation patterns with minim
authors
co-cited works
representative citing papers
The Pile is a newly constructed 825 GiB dataset from 22 diverse sources that enables language models to achieve better performance on academic, professional, and cross-domain tasks than models trained on Common Crawl variants.
Frontier is a new discrete-event simulator for disaggregated LLM serving that incorporates co-location, PDD, AFD, and optimizations, achieving under 4% throughput error and large reductions in latency prediction error versus prior simulators.
Routers in SMoE models form geometric alignments with their experts through shared gradient directions, enabling effective specialization that auxiliary load-balancing losses tend to disrupt.
MISA routes to a small subset of indexer heads via block statistics, matching full DSA performance on LongBench with 4-8x fewer heads and 3.82x speedup while recovering over 92% of selected tokens.
Standard top-k routers in MoE language models often select suboptimal routes for difficult tokens, and updating only the final router layer raises pass@K on AIME and HMMT benchmarks across multiple models.
BatMIL uses hybrid hyperbolic-Euclidean geometry, an S4 state-space backbone, and chunk-level mixture-of-experts to outperform prior multiple-instance learning methods on seven whole-slide image datasets across six cancers.
Approximate multipliers degrade MoE and dense DNNs at different rates; ResNet-20 recovers fully after retraining while VGG models often fail at aggressive approximations except Cluster MoE, and Hard MoE can outperform dense on ViT under cost-matched aggressive approximation.
Coral cuts multi-LLM serving costs by up to 2.79x and raises goodput by up to 2.39x on heterogeneous GPUs through adaptive joint optimization and a lossless two-stage decomposition that solves quickly.
MoE-Prefill achieves 1.35-1.59x higher throughput for prefill-only MoE serving by using asynchronous expert parallelism to overlap weight AllGather with computation and prefix-aware routing with true-FLOPs tracking.
A new SFT framework for MoE models combines bias-driven sparsification with gated condenser experts to retain long-tailed expert information, outperforming DenseMixer and ESFT by over 2.5% on math reasoning and commonsense QA benchmarks.
FEPLB reduces token and GEMM stragglers in MoE training by 50-70% using nearly free Copy Engine communication on Hopper architecture.
Expert upcycling duplicates experts in an existing MoE checkpoint and continues pre-training to match fixed-size baseline performance with 32% less compute.
DepthVAR adaptively allocates per-token computational depth in VAR models using a cyclic rotated scheduler and dynamic layer masking to achieve 2.3-3.1x inference speedup with minimal quality loss.
A mixture-of-experts transformer foundation model pretrained on diverse SEM images enables generalization across materials and outperforms SOTA on unsupervised defocus-to-focus restoration.
PathMoE constrains expert paths in MoE models by sharing router parameters across layer blocks, yielding more concentrated paths, better performance on perplexity and tasks, and no need for auxiliary losses.
Split-MoPE integrates split learning with predefined-expert routing to maximize usable data in vertical federated learning under sample misalignment, delivering state-of-the-art accuracy in one communication round plus built-in robustness and per-sample contribution scores.
TWNM framework equips audio-language models with spatial scene analysis via FOA simulation and metadata-grounded training, reaching 70.8% accuracy on a new ASA benchmark.
Loss-Free Balancing keeps expert loads balanced in MoE models by dynamically adjusting routing-score biases based on recent usage, avoiding auxiliary-loss interference and yielding better performance.
LLM.int8() performs 8-bit inference for transformers up to 175B parameters with no accuracy loss by combining vector-wise quantization for most features with 16-bit mixed-precision handling of systematic outlier dimensions.
Switch Transformers use top-1 expert routing in a Mixture of Experts setup to scale to trillion-parameter language models with constant compute and up to 4x speedup over T5-XXL.
Complete-muE combines active-width μP and activated-expert scaling to transfer hyperparameters across dense FFN, dense MoE, and sparse MoE while covering changes in experts, capacity, width, depth, batch size, and duration.
NASiC fuses CAM-based expert selection and multibit CIM computation in 3D NAND into one cycle for MoE LLM inference, claiming 4-114.8x performance and 3.9-70x energy efficiency gains over prior designs with high accuracy.
MultiWrite is a new many-to-many transmission semantic that uses multicast principles to eliminate redundant packets in collective operations, delivering up to 33% lower latency for AllGather and AlltoAll on Ascend NPUs.
citing papers explorer
-
Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts
Loss-Free Balancing keeps expert loads balanced in MoE models by dynamically adjusting routing-score biases based on recent usage, avoiding auxiliary-loss interference and yielding better performance.
-
Lynx: Enabling Efficient MoE Inference through Dynamic Batch-Aware Expert Selection
Lynx exploits training-induced batch-level expert activation skews via AffinityBinning to reduce invoked experts per batch, delivering up to 1.30x throughput with under 1% accuracy loss across four model families.
-
Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models
MoT decouples non-embedding parameters by modality in transformers to match dense multi-modal performance with roughly one-third to one-half the FLOPs.
-
$\pi_0$: A Vision-Language-Action Flow Model for General Robot Control
π₀ is a vision-language-action flow model trained on diverse multi-platform robot data that supports zero-shot task performance, language instruction following, and efficient fine-tuning for dexterous tasks.
-
LongVILA: Scaling Long-Context Visual Language Models for Long Videos
LongVILA scales visual-language models from 8 to 2048 video frames with 99.8% needle-in-a-haystack accuracy using long-context extension, supervised fine-tuning, and multi-modal sequence parallelism on up to 256 GPUs.
-
Mixture-of-Depths: Dynamically allocating compute in transformer-based language models
Mixture-of-Depths enables transformers to dynamically allocate compute by routing only the top-k tokens through each layer's full computations, matching baseline performance with a fraction of the FLOPs per forward pass and up to 50% faster sampling.
-
MoE-LLaVA: Mixture of Experts for Large Vision-Language Models
MoE-LLaVA applies mixture-of-experts sparsity to LVLMs via MoE-Tuning, delivering LLaVA-1.5-7B level visual understanding and better hallucination resistance with only ~3B active parameters.
-
Test-Time Alignment via Hypothesis Reweighting
HyRe personalizes reward models at test time by reweighting an ensemble of heads trained on aggregate preferences, using few target examples to outperform uniform averaging and prior methods on RewardBench and 32 tasks.
-
Mixtral of Experts
Mixtral 8x7B is a sparse MoE LLM activating 2 of 8 experts per layer that matches or exceeds Llama 2 70B and GPT-3.5 on benchmarks while using only 13B active parameters.
-
Qwen2.5 Technical Report
Qwen2.5 LLMs scale pre-training data to 18 trillion tokens and apply multistage reinforcement learning, achieving competitive performance on benchmarks with models up to 5 times larger.
-
A Survey on Efficient Inference for Large Language Models
The paper surveys techniques to speed up and reduce the resource needs of LLM inference, organized by data-level, model-level, and system-level changes, with comparative experiments on representative methods.