PaperlayerView exact revision on arXiv

Source-linked AI summary

SCORAS-MoE: Joint Compression and Resource-Adaptive Deployment of MoE-VLMs in LEO Satellite Networks

Tong Quan, Yuanlong Wan, Huasen He, Yunpeng Hou, Shuangwu Chen, Xiaofeng Jiang, Jian Yang

arXiv:2609.36763v1cs.NI

TL;DR

Limited onboard memory and energy, together with changing satellite resources, make MoE-VLM compression and deployment challenging. SCORAS-MoE combines sensitivity-guided compression with adaptive shard placement, improving accuracy under aggressive compression and balancing service quality with energy use.

  • Problem

    Limited satellite memory and energy require compressing large MoE-VLMs, while changing energy and communication conditions make deployment decisions time-sensitive.

    Source: §I · p. 1

  • Method

    SCORAS-MoE uses routed-output perturbation to allocate expert ranks, evaluates compressed accuracy–energy profiles, then selects profile compositions and shard placements online.

    Source: §I · pp. 1–2

  • Results

    69.9% mean accuracy at β = 0.3 versus 66.2% for uniform rank allocation; the fixed-profile solver reports 5.3% higher high-load throughput than Greedy.

    Source: §I · p. 2, §VII · p. 12

  • Takeaways & Limitations

    Adaptive profile selection further improves the quality–energy balance over fixed-profile deployment in the evaluated satellite inference setting.

    Source: §VII · p. 12

Abstract

from arXiv · show

Deploying large vision-language models (VLMs) onboard satellites enables onboard data processing and reduces raw data downlink. However, onboard inference faces two resource challenges. Limited onboard memory and energy require model compression and distributed deployment. Dynamic resource availability requires fast deployment decisions as illumination, battery levels, and communication conditions change. We present SCORAS-MoE, a joint compression and deployment framework for mixture-of-experts (MoE) VLMs in low Earth orbit (LEO) satellite networks. To address limited resources, SCORAS-MoE measures the perturbation of the routed MoE output caused by low-rank approximation, assigns higher ranks to more sensitive experts, and distributes compressed model shards across satellites for cooperative inference. The compressed models yield profiles of measured accuracy and inference energy. To adapt to dynamic resources, the online scheduler selects profile compositions and shard placements in each slot. For each candidate composition, it reduces placement to a minimum-cost assignment problem solved by the Hungarian algorithm, while enumerating the compositions yields the optimal deployment for the current-slot objective. Experiments on Qwen3-VL-30B-A3B-Instruct show that allocating ranks based on output perturbation is particularly effective under aggressive compression, with an absolute gain of $3.7\%$ in mean accuracy over uniform rank allocation when expert projections retain $30\%$ of their original parameters. The fixed-profile scheduler achieves higher throughput with fewer service switches and lower battery impact than the evaluated proximal policy optimization (PPO) and evolutionary baselines, with respective speedups of $8.7\times$ and $183.5\times$. Adaptive profile selection further improves the balance between service quality and energy use.

I. INTRODUCTION

Onboard MoE-VLM inference in LEO satellites must contend with limited resources and changing energy and communication conditions. SCORAS-MoE links sensitivity-guided compression to adaptive distributed deployment, addressing a gap between model compression and online placement.

I. INTRODUCTION

  • Limited memory, computing capacity, and energy motivate compression, while models that exceed one satellite’s capacity require shards to be distributed for cooperative inference.Dynamic orbital energy and communication conditions also require fast decisions about model profiles and shard placement.Source: §I · p. 1

A. MoE Compression and Compression Sensitivity

  • MoE compression work uses pruning, quantization, and low-rank approximation, but structural and routing signals alone do not directly measure functional distortion from a specific approximation.Output-space distortion provides a functional sensitivity signal; SCORAS-MoE applies a forward-only, routing-aware probe to low-rank expert compression and builds deployable profiles.Source: §I · p. 1, A. MoE Compression and Compression… · p. 2

I. INTRODUCTION

  • SCORAS-MoE measures routed-output perturbation to allocate higher ranks to sensitive experts, then uses evaluated accuracy–energy profiles for online profile selection and shard placement.For each candidate profile composition, placement is solved as a minimum-cost assignment; comparing compositions yields the current-slot optimum.Source: §I · p. 2
  • At β = 0.3, SCORAS-MoE achieves a mean accuracy of 69.9%, compared with 66.2% for uniform rank allocation.The introduction also reports 5.3% higher high-load throughput than Greedy for the common β = 0.4 profile, with reduced switching and battery impact.Source: §I · p. 2

B. Distributed Inference and Satellite Service Scheduling

  • Satellite scheduling and distributed inference studies address offloading, partitioning, energy, and mobility, but generally do not jointly model complete inference pipelines whose shards use the same profile.Learning-based policies also tend to focus on tasks, missions, or offloading rather than profile-specific pipeline composition and placement.Source: B. Distributed Inference and Satellite… · p. 3

III. SYSTEM MODEL AND PROBLEM FORMULATION

The system model represents cooperative MoE-VLM inference as profile-specific pipelines whose shards are placed across a changing LEO constellation. Profiles connect measured model quality and energy use to per-slot deployment decisions.

B. Satellite Network and Resource Dynamics

  • A complete model may exceed an individual satellite’s compute, memory, and energy capacity, so inference pipelines execute cooperatively across a Walker-Delta LEO constellation.The constellation supports onboard sensing and ground-terminal requests through satellite–ground links.Source: B. Satellite Network and Resource… · p. 3
  • The horizon is divided into equal-length slots, with satellite positions, link capacities, illumination, and battery states treated as constant within each slot and updated between slots.Orbital propagation updates geometry, links, and illumination; battery energy evolves between slots.Source: B. Satellite Network and Resource… · p. 3
  • Only satellites meeting both satellite–ground and inter-satellite link requirements may host inference shards.This communication-feasibility restriction limits the eligible placement set in each slot.Source: B. Satellite Network and Resource… · p. 3

A. SCORAS-MoE Framework

  • The offline stage compresses and evaluates models at multiple expert-projection retention ratios, producing a finite set of profiles with distinct parameter counts, accuracy, and inference-energy costs.The scheduler observes satellite states and demand each slot, chooses a profile composition, and assigns its required shards.Source: C. Sharded Pipeline Inference · pp. 3–4, A. SCORAS-MoE Framework · p. 3

C. Sharded Pipeline Inference

  • Each profile’s model is partitioned at MoE-layer granularity into K required shards, and one replica of every shard forms a complete inference pipeline.Active replica counts must match across a profile’s shards; each communication-feasible satellite can host at most one active shard instance.Source: C. Sharded Pipeline Inference · p. 4
  • Each profile records its expert-projection retention ratio, measured mean accuracy, and measured slot-level inference energy for a representative shard.The representative-shard energy is used as a common estimate because shards have almost identical tensor dimensions and the same number of MoE layers.Source: C. Sharded Pipeline Inference · p. 4
  • Satellite service state records whether a satellite is idle or hosts a particular shard under a profile; changes between consecutive states count as service switches.The definition includes activation, deactivation, shard replacement, and profile replacement.Source: C. Sharded Pipeline Inference · p. 4

D. Energy Model

The energy model tracks each satellite’s battery as inference load and harvested solar energy change over time. It accounts for charging and discharging limits and requires placements to remain energy-feasible.

D. Energy Model

  • Battery energy evolves from the balance between solar energy harvested along the orbit and energy consumed by the inference service.The inference load depends on whether the satellite is idle or active and, when active, on its hosted shard.Source: D. Energy Model · p. 4
  • The model distinguishes beginning-of-slot battery energy, illumination, battery capacity, harvested solar power, and idle power for each satellite.The energy model uses these quantities to calculate the subsystem load and allocated solar energy.Source: D. Energy Model · p. 4
  • Charging is constrained by battery capacity and maximum charging rate, while a load deficit is supplied by the battery subject to a maximum discharge rate.A placement is energy-feasible only when the modeled battery constraints are satisfied.Source: D. Energy Model · pp. 4–5
  • Charging and discharging do not occur in the same slot.Source: D. Energy Model · p. 5

E. Joint Optimization Formulation

The joint formulation evaluates service quality, unmet demand, inference-induced battery impact, and switching within a per-slot utility. A causal policy makes decisions from each slot’s observed state, while the long-horizon offline problem is NP-hard.

E. Joint Optimization Formulation

  • The per-slot utility weights service quality against penalties for unmet demand, battery impact, and switching.The corresponding weights are nonnegative; switches are counted only between consecutive slots, with the initial switch count set to zero.Source: E. Joint Optimization Formulation · p. 5
  • Service quality combines the number of delivered pipelines with each profile’s measured mean accuracy, while unmet demand and aggregate pipeline capacity constrain service provision.The formulation also tracks the battery impact of inference and the number of service switches.Source: E. Joint Optimization Formulation · p. 5
  • A causal scheduling policy maps the state observed at each slot’s beginning to profile-specific pipeline counts and shard-placement decisions.Within a slot, composition determines accuracy and energy demand, while placement determines feasibility and switching; deployments update states for later decisions.Source: E. Joint Optimization Formulation · p. 5
  • The deterministic offline version of the long-horizon problem is NP-hard, since a restricted single-satellite case contains cardinality-constrained unbounded knapsack.SCORAS-MoE therefore optimizes deployment at the beginning of each slot using the current state.Source: E. Joint Optimization Formulation · p. 5

IV. COMPRESSION GUIDED BY OUTPUT PERTURBATION

SCORAS-MoE compresses MoE expert projections with low-rank approximations and allocates ranks according to how strongly each expert’s approximation perturbs the routed layer output.

B. Routing-Aware Output Perturbation

A. MoE Expert Compression

  • Low-rank factorization replaces each expert projection matrix with two factors, reducing its parameter count from doutdin to r(dout + din).The approximations use activation-aware reconstruction on calibration inputs.Source: A. MoE Expert Compression · p. 6

B. Routing-Aware Output Perturbation

  • The routing-aware probe uses cached calibration activations, expert selections, and routing weights to measure each expert’s effect on the routed layer output.Experts differ in both their contribution to that output and their response to approximation.Source: B. Routing-Aware Output Perturbation · p. 6

C. Rank Allocation and Deployment Profiles

For each target retention ratio, SCORAS-MoE distributes a layer’s rank budget across experts using normalized perturbation scores, then evaluates the resulting models to form accuracy–energy profiles.

C. Rank Allocation and Deployment Profiles

  • Each expert receives a minimum rank, and the remaining layer rank is distributed according to normalized probe scores.The allocation is rounded and clipped to feasible ranks, then applied to both gate–up and down projections.Source: C. Rank Allocation and Deployment… · pp. 6–7
  • Calibration caches projection inputs, expert selections, and routing weights, and reuses them across target retention ratios.The rank-allocation procedure also uses a rank-floor coefficient and stabilization constant.Source: C. Rank Allocation and Deployment… · p. 6
  • The offline procedure evaluates each compressed model for mean accuracy and measures representative-shard inference energy to create deployment profiles.The profile records the target parameter retention ratio, measured accuracy, and measured energy.Source: C. Rank Allocation and Deployment… · pp. 6–7

V. FAST RESOURCE-ADAPTIVE DEPLOYMENT

SCORAS-MoE adapts deployment each slot by enumerating profile compositions and assigning their required shards to communication-feasible satellites at minimum battery-impact and switching cost.

A. Current-Slot State and Profile Composition

  • The scheduler enumerates profile-count vectors for complete pipelines, with every shard in an individual pipeline using the same profile.It excludes compositions whose required shard instances exceed the number of communication-feasible satellites.Source: A. Current-Slot State and Profile… · p. 7

B. Minimum-Cost Shard Assignment

  • For each profile-count vector, the scheduler assigns required shards to satellites by minimizing battery impact and switching cost.Battery feasibility and assignment cost depend on satellite energy, illumination, communication feasibility, and prior placement.Source: B. Minimum-Cost Shard Assignment · pp. 7–8, A. Current-Slot State and Profile… · p. 7
  • The assignment cost accounts for activating, deactivating, replacing, or retaining satellite shard and profile states.Retaining a complete profile–shard state cancels its assumed deactivation cost; activating a previously idle satellite incurs an adjustment.Source: B. Minimum-Cost Shard Assignment · p. 8
  • For each composition, a rectangular linear sum-assignment problem maps required shard instances to satellites, and the Hungarian algorithm solves it.Each satellite can host at most one instance, and energy-infeasible assignment edges receive a prohibitive cost.Source: B. Minimum-Cost Shard Assignment · p. 8

C. Profile Selection and Per-Slot Optimality

For the observed state, the scheduler scores the minimum-cost assignment for each feasible profile composition and selects the composition with the highest current-slot utility.

C. Profile Selection and Per-Slot Optimality

  • The scheduler selects the profile-count vector whose minimum-cost assignment maximizes current-slot utility.The utility compares service benefit against deployment cost for each feasible composition.Source: C. Profile Selection and Per-Slot… · p. 8
  • The per-slot solution is globally optimal for the stated complete-pipeline model and single-slot objective.For each profile-count vector, Hungarian matching minimizes placement cost while service-quality and unmet-demand terms remain fixed.Source: C. Profile Selection and Per-Slot… · p. 8
  • The matching work is bounded by the number of profile compositions and the cost of solving each assignment; the number of profiles is small in practice.The passage gives an O(Ct^3) upper bound for each assignment.Source: C. Profile Selection and Per-Slot… · p. 8
  • The online procedure uses demand, satellite battery and illumination states, communication feasibility, previous service states, and offline profiles to build candidate deployments.It returns the selected shard placement and updates battery and service states for the next slot.Source: C. Profile Selection and Per-Slot… · p. 8

VI. PERFORMANCE EVALUATION

The evaluation examines compression accuracy, deployment performance and decision time, adaptive profile composition, and deployment ablations. Algorithm 2 selects a per-slot profile composition and shard placement from demand, satellite states, and available profiles.

VI. PERFORMANCE EVALUATION

  • The evaluation covers compression accuracy, fixed-profile deployment performance and decision time, adaptive profile composition, and deployment ablations.These evaluations assess support for inference under limited and time-varying satellite resources.Source: §VI · pp. 8–9
  • Algorithm 2 takes current demand, satellite states, and the profile set, and returns a profile composition and profile-aware shard placement.It initializes an empty deployment option and iterates over nonempty candidate compositions.Source: §VI · p. 9
  • The procedure skips compositions requiring more shards than candidate satellites and rejects assignments containing energy-infeasible edges.It decodes the selected assignment into placement, executes the deployment, and updates satellite states for the next slot.Source: §VI · p. 9

1) Model and Compression Configuration:

The evaluation compares rank-allocation methods and deployment settings for compressed MoE models. Output-perturbation allocation has its clearest mean-accuracy advantage at the lowest retention ratio, within a setup that measures profile accuracy and energy for resource-adaptive deployment.

1) Model and Compression Configuration:

  • The 48-layer model is divided into eight balanced six-layer shards; four allocation methods share calibration inputs, nominal layer-wise total rank, minimum rank, and full-model evaluation.Accuracy is measured on TextVQA, DocVQA, AI2D, OCRBench, and RealWorldQA; the uncompressed mean is 80.7%.Source: 1) Model and Compression Configuration · p. 9

2) Satellite and Communication Configuration:

  • The satellite experiment uses 60 Walker-Delta satellites over 24 five-minute slots, with 5–9 requested pipelines; four profiles pair mean accuracies of 69.9%, 76.1%, 80.7%, and 82.7% with measured shard energies.Per-shard slot energies are 2662.8, 3146.0, 3922.8, and 4976.8 J for β = 0.3, 0.4, 0.6, and 0.8, respectively; the inference subsystem receives an 8 Wh battery budget and up to 25 W harvested solar power when illuminated.Source: 2) Satellite and Communication… · p. 9

3) Baselines and Metrics:

  • Fixed-profile deployment compares Greedy, evolutionary, and PPO baselines under shared workloads and feasibility rules; reported measures include pipeline completion, demand fulfillment, normalized DoD, switches, and solve time.Joint compression–deployment also compares online selection against four fixed-β baselines; profile-selection measures add accuracy-weighted throughput, pipeline-weighted mean accuracy, and profile counts.Source: 3) Baselines and Metrics · pp. 9–10

B. Compression Accuracy at Low Parameter Retention Ratios

  • At β = 0.3, mean accuracy is 69.9% for SCORAS-MoE versus 66.2% for Uniform, 67.4% for Energy, and 66.9% for Frequency; its gains are largest at low retention.SCORAS-MoE has the best mean score at β = 0.3, β = 0.4, and β = 0.8; at β = 0.6, Energy reaches 80.8% versus 80.7% for SCORAS-MoE.Source: B. Compression Accuracy at Low… · p. 10, Figure 2 · p. 10

C. Dynamic Deployment with Fixed β = 0.4

With a fixed β = 0.4 profile, SCORAS-MoE improves high-load throughput over Greedy while sharply reducing switches and scheduling faster than the evaluated PPO and evolutionary baselines. Adaptive profile selection improves the quality–energy balance, and ablations show the roles of communication filtering, switching cost, and unmet-demand penalties.

C. Dynamic Deployment with Fixed β = 0.4

  • At nreq = 8–9, SCORAS-MoE completes 160 pipelines versus Greedy’s 152; across nreq = 5–9, it reduces service switches by 76.2–82.8% relative to Greedy.At nreq = 9, inference-induced DoD falls from 58.68 to 53.88, an 8.2% reduction.Source: C. Dynamic Deployment with Fixed β = 0.4 · p. 10, Figure 3 · p. 11
  • 183.5× and 8.7× are the fixed-profile solver’s speedups over the evolutionary algorithm and PPO, respectively.Across 24 slots averaged over nreq = 5–9, the respective runtimes are 0.155 s, 28.517 s, and 1.352 s; PPO stops early when it misses the required improvement over Greedy within five episodes.Source: C. Dynamic Deployment with Fixed β = 0.4 · p. 10

D. Joint Benefits of Compression and Adaptive Deployment

  • At nreq = 7, adaptive SCORAS-MoE completes 164 pipelines versus 151 for Fixed-β = 0.8, raises accuracy-weighted throughput by 0.8%, and reduces unmet demand from 17 to 4 pipelines.At nreq = 5, it retains 97.2% of Fixed-β = 0.8 accuracy-weighted throughput while reducing inference-induced DoD by 70.1%.Source: D. Joint Benefits of Compression and… · p. 10, Figure 4 · p. 11
  • At nreq = 7, the selected mixture uses 75 pipelines at β = 0.3, 86 at β = 0.8, and 3 at β = 0.4; at nreq = 8–9, completed pipelines, accuracy-weighted throughput, and DoD remain unchanged while unmet demand rises to 28 and 52.Adaptive profile enumeration averages 1.021 s over 24 slots and five loads.Source: D. Joint Benefits of Compression and… · p. 10

E. Ablation Study

  • Removing communication filtering reduces completed pipelines from 120–160 to 110–151; removing switching cost raises service switches by 165.1–246.8%, while removing the unmet-demand penalty lowers completed pipelines to 107–111.Without filtering, all 60 satellites enter the candidate set, including those that fail communication requirements; without the unmet-demand penalty, unmet demand rises to 13–105.Source: E. Ablation Study · pp. 11–12, Table III · p. 12

VII. CONCLUSION

  • SCORAS-MoE combines offline output-perturbation rank allocation with online profile enumeration and minimum-cost shard assignment to achieve the per-slot optimum; future work targets heterogeneous platforms and latency-aware deployment in larger constellations.Under β = 0.3, mean accuracy is 69.9% versus 66.2% for uniform rank allocation.Source: §VII · p. 12

Related papers

Loading 2609.36763v1…