performance
Mixture-of-Experts Model Routing Efficiency Analyser
Analyses MoE model routing patterns to identify load imbalances, expert utilisation, and gating network inefficiencies.
mixture-of-experts routing-analysis load-balancing
prompt
# MoE Routing Efficiency Analysis You are an expert in Mixture-of-Experts (MoE) model architectures. Analyse the routing patterns and performance metrics below to identify inefficiencies and optimisation opportunities. ## Model Configuration **Number of experts:** [number] **Routing strategy:** [top-k/switch/hash-based/other] **Expert capacity:** [tokens per expert] **Load balancing method:** [auxiliary loss/random routing/other] ## Routing Metrics ``` [paste routing logs/metrics here] ``` ## Performance Data ``` [paste performance metrics: throughput, latency, memory usage] ``` ## Analysis Requirements ### Expert Utilisation - Calculate expert load distribution across the input batch - Identify underutilised and overloaded experts - Measure routing entropy and diversity ### Gating Network Performance - Analyse confidence scores for expert selection - Detect routing collapse or expert death - Evaluate load balancing loss effectiveness ### Efficiency Metrics - Compute FLOPs utilisation vs theoretical maximum - Measure communication overhead between experts - Calculate memory efficiency per expert ### Bottleneck Identification - Identify capacity bottlenecks causing token dropping - Detect routing hotspots - Analyse gradient flow to expert parameters ## Recommendations Provide specific optimisation strategies: - Routing algorithm adjustments - Expert capacity rebalancing - Load balancing hyperparameter tuning - Architecture modifications Include quantified impact estimates and implementation complexity for each recommendation.
Use this when profiling MoE models like Mixtral, GLaM, or Switch Transformer to optimise expert utilisation and reduce routing overhead. Works with Claude, GPT-4, and Gemini for analysing routing logs and performance telemetry.