Scaling Laws for Fine-Grained Mixture of Experts
read the original abstract
Mixture of Experts (MoE) models have emerged as a primary solution for reducing the computational cost of Large Language Models. In this work, we analyze their scaling properties, incorporating an expanded range of variables. Specifically, we introduce a new hyperparameter, granularity, whose adjustment enables precise control over the size of the experts. Building on this, we establish scaling laws for fine-grained MoE, taking into account the number of training tokens, model size, and granularity. Leveraging these laws, we derive the optimal training configuration for a given computational budget. Our findings not only show that MoE models consistently outperform dense Transformers but also highlight that the efficiency gap between dense and MoE models widens as we scale up the model size and training budget. Furthermore, we demonstrate that the common practice of setting the size of experts in MoE to mirror the feed-forward layer is not optimal at almost any computational budget.
This paper has not been read by Pith yet.
Forward citations
Cited by 26 Pith papers
-
How to Scale Mixture-of-Experts: From muP to the Maximally Scale-Stable Parameterization
The authors derive a Maximally Scale-Stable Parameterization (MSSP) for MoE models that achieves robust learning-rate transfer and monotonic performance gains with scale across co-scaling regimes of width, experts, an...
-
Scaling Laws for Mixture Pretraining Under Data Constraints
Repetition-aware scaling laws show scarce target data in pretraining mixtures can be repeated 15-20 times optimally, with the best count depending on data size, compute, and model scale.
-
Geometric Asymmetry in MoE Specialization: Functional Decorrelation and Representational Overlap
MoE experts in pretrained Transformers exhibit functional decorrelation with near-zero Jacobian alignment yet occupy partially overlapping representation subspaces, with routing sparsity modulating the geometry.
-
Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts
Expert upcycling duplicates experts in an existing MoE checkpoint and continues pre-training to match fixed-size baseline performance with 32% less compute.
-
Sparsely gated tiny linear experts
Sgatlin replaces transformer FF layers with sparse single linear neurons, improving perplexity across compute budgets and enabling direct interpretation of semantically clustered circuits for factual recall.
-
NUCLEUS-MoE: Unified Model of Pool Boiling for Liquid Cooling
NUCLEUS-MoE is a single neural network architecture that models saturated and subcooled pool boiling for dielectrics, refrigerants, and cryogens with generalization to new fluids.
-
MobileMoE: Scaling On-Device Mixture of Experts
MobileMoE introduces on-device MoE LLMs that match dense models with 2-4x fewer FLOPs and provide efficient smartphone inference.
-
Scaling Laws for Mixture Pretraining Under Data Constraints
Empirical study shows mixture pretraining tolerates higher target data repetition than single-source training, with a new repetition-aware scaling law enabling principled mixture selection based on data size, compute,...
-
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
DECO is a sparse MoE architecture with ReLU-based routing, learnable expert scaling, and NormSiLU activation that matches dense Transformer performance at 20% expert activation and delivers 2.93x speedup on Jetson AGX Orin.
-
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
DECO sparse MoE matches dense Transformer performance at 20% expert activation with a 3x hardware inference speedup.
-
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
DECO matches dense model performance at 20% expert activation via ReLU-based routing with learnable scaling and the NormSiLU activation, plus a 3x real-hardware speedup.
-
UniPool: A Globally Shared Expert Pool for Mixture-of-Experts
A shared global expert pool in MoE improves validation loss over per-layer experts and allows sublinear expert-parameter growth with depth.
-
A Limit Theory of Foundation Models: A Mathematical Approach to Understanding Emergent Intelligence and Scaling Laws
Emergent intelligence is recast as the existence of the limit of performance E(N,P,K) as N,P,K to infinity, with necessary and sufficient conditions derived via nonlinear Lipschitz operator theory and scaling laws obt...
-
Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts
Expert upcycling expands MoE models by duplicating experts and continuing pre-training, matching baseline performance while saving 32% GPU hours in 7B-13B experiments.
-
Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts
BAR trains independent domain experts via separate mid-training, SFT, and RL pipelines then composes them with a MoE router to match monolithic retraining performance at lower cost and without catastrophic forgetting.
-
Generalization and Scaling Laws for Mixture-of-Experts Transformers
A covering-number bound and ERM analysis for MoE Transformers under manifold data produces generalization and scaling laws that treat active capacity like dense networks while adding routing overhead.
-
Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs
A conditional scaling law fitted on over 200 models from 80M to 3B parameters identifies architectures that deliver up to 2.1% higher accuracy and 42% higher inference throughput than LLaMA-3.2 under the same training budget.
-
ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution
ShinkaEvolve improves sample efficiency in LLM-driven program evolution via parent sampling, code novelty rejection-sampling, and bandit LLM ensemble selection, achieving new SOTA circle packing with 150 samples and g...
-
LoRA-Mixer: Coordinate Modular LoRA Experts Through Serial Attention Routing
LoRA-Mixer routes modular LoRA experts into attention projection matrices with an adaptive Routing Specialization Loss to improve multi-task performance while using fewer trainable parameters than prior LoRA-MoE methods.
-
Mixture-of-Control: State-Aware Fine-Tuning for Transformer-based Models
Mixture-of-Control adaptively combines local and global control states in transformer fine-tuning by treating per-block states as experts in a sparse MoE setup to improve cross-block communication while keeping memory...
-
Sakana Fugu Technical Report
Sakana Fugu trains LLM orchestrators using fine-tuning, evolutionary algorithms, and RL to build query-adaptive multi-agent scaffolds, claiming SOTA results on benchmarks including SWE-Bench Pro and GPQA-Diamond.
-
Dense vs Sparse Pretraining at Tiny Scale: Active-Parameter vs Total-Parameter Matching
At tiny scale, MoE transformers lower validation loss versus dense models when active parameters match but raise it when total stored parameters match.
-
A Limit Theory of Foundation Models: A Mathematical Approach to Understanding Emergent Intelligence and Scaling Laws
Emergent intelligence corresponds to the limit of a performance function E(N,P,K) as N, P, K go to infinity, originating from a parameter-limit architecture whose existence is governed by Lipschitz conditions, with sc...
-
Mellum2 Technical Report
Mellum 2 is a 12B MoE model with 2.5B active parameters, trained on 10.6T tokens with MoE, GQA, SWA, and MTP, then post-trained into Instruct and Thinking variants, claimed competitive with 4B-14B models at 2.5B compute.
-
A Limit Theory of Foundation Models: A Mathematical Approach to Understanding Emergent Intelligence and Scaling Laws
Formalizes emergent intelligence in foundation models as the limit of E(N,P,K) as N,P,K approach infinity, proves existence conditions via nonlinear Lipschitz operators, and derives scaling laws from covering numbers.
-
There Will Be a Scientific Theory of Deep Learning
A mechanics of the learning process is emerging in deep learning theory, characterized by dynamics, coarse statistics, and falsifiable predictions across idealized settings, limits, laws, hyperparameters, and universa...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.