WARD: Runtime Workload-Adaptive Vision TRansformer Framework for Dependable Edge AI

Mahdi Taheri; Pramit Kumar Bhaduri; Mohammad Masoumi; Ali Mahani

Source-linked AI summary

WARD: Runtime Workload-Adaptive Vision TRansformer Framework for Dependable Edge AI

Mahdi Taheri, Pramit Kumar Bhaduri, Mohammad Masoumi, Ali Mahani

arXiv:2609.17556v1Jul 20, 2026cs.ARcs.AIcs.LG

TL;DR

Edge AI must adapt to changing power, reliability, and input distributions, but existing fault-tolerance and continual-learning approaches do not jointly address these conditions or practical accelerator deployment. WARD combines isolated ViT subnetworks, reliability-aware continual learning, and dynamic operating modes, and its FPGA implementation achieves low failure rates with minimal hardware overhead.

Problem
Existing fault-tolerant methods assume static conditions, while continual-learning methods neglect concurrent hardware faults and uninterrupted inference during online adaptation.
Method
WARD combines channel-wise subnetwork partitioning, reliability-aware continual learning, and four dynamically scheduled operating modes on a lightweight FPGA-based accelerator.
Results
1.79% network-level failure rate and less than 5% hardware overhead, with runtime mode switching in 50 clock cycles.
Takeaways & Limitations
WARD provides a lightweight framework for uninterrupted, runtime-adaptive reliability management on programmable edge AI accelerators.

Abstract

from arXiv · show

Edge-deployed AI operate under dynamically changing power budgets, reliability requirements, and input distributions, requiring continuous adaptation. Such conditions arise in long-running edge AI applications, including autonomous systems, industrial monitoring, and satellite onboard intelligence. Existing fault-tolerant methods assume static operating conditions, whereas continual learning techniques neglect concurrent hardware faults during online adaptation. Moreover, the practical deployment of runtime-adaptive reliability frameworks on programmable AI accelerators remains largely unexplored. This paper presents WARD, a runtime-adaptive Vision Transformer framework that combines channel-wise subnetwork partitioning, reliability-aware continual learning, and dynamic operating-mode scheduling to jointly optimize performance, fault tolerance, and adaptation according to runtime conditions. Two physically isolated subnetworks execute under four operating modes (i.e. Full-Precision Mode, Low-Power Mode, High-Reliability Mode, and Adaptive Mode) that dynamically adjust computational cost and reliability while ensuring uninterrupted inference for real-time requirements. To validate the practical deployability of the proposed framework, WARD is implemented on a lightweight FPGA-based accelerator extended with runtime hardware support for mode scheduling and resource management. Experimental results demonstrate that the proposed split architecture achieves a network-level failure rate of only 1.79% under high Bit Error Rates. The hardware implementation incurs less than 5% area overhead and supports runtime mode transitions within few clock cycles, demonstrating that adaptive reliability management can be integrated into programmable edge AI accelerators with negligible implementation overhead.

I. INTRODUCTION

Mission-critical edge AI must adapt continuously to changing power, radiation, reliability, and input conditions without interrupting inference. WARD addresses this need through isolated subnetworks, runtime operating modes, reliability-aware continual learning, and lightweight accelerator support.

I. INTRODUCTION: Mission-critical edge applications require uninterrupted real-time inference while adapting behavior to changing power, environmental, and reliability conditions.Satellite deployments additionally face changing power availability, radiation intensity, and environmental conditions.
I. INTRODUCTION: Existing fault-tolerance methods use fixed protection overhead or static reliability assumptions, while continual learning commonly assumes inference can be interrupted and hardware remains reliable.These limitations leave concurrent adaptation and hardware-fault containment insufficiently addressed.
I. INTRODUCTION: WARD partitions a pretrained Vision Transformer into two physically isolated subnetworks operating across four runtime modes to balance cost, reliability, and adaptation.The modes are Full-Precision, Low-Power, High-Reliability, and Adaptive.
I. INTRODUCTION: WARD combines channel-wise partitioning, reliability-aware continual learning, and dynamic scheduling to preserve inference while adapting to runtime conditions.Its continual-learning strategy updates empirically fault-resilient parameter regions while preserving reliability-critical parameters.
I. INTRODUCTION: WARD extends a lightweight FPGA accelerator with runtime support for mode scheduling and resource allocation, demonstrating practical deployment with lightweight architectural changes.The accelerator implementation is intended to support runtime-adaptive inference beyond software simulation.

II. BACKGROUND

Vision Transformers process images as patch-token sequences whose tightly coupled residual computation can propagate hardware faults through later layers. Continual-learning methods address distribution shift but generally assume uninterruptedly reliable execution, motivating fault-aware runtime adaptation.

A. Vision Transformer: A Vision Transformer divides an image into patches, embeds them as tokens, processes them through attention and MLP encoder blocks, and classifies the final CLS token.Layer normalization and residual connections surround the encoder sub-operations.
A. Vision Transformer: Strictly additive residual connections allow corrupted attention or MLP weights to inject errors into the shared token stream and propagate them through subsequent blocks.These properties make fault isolation and recovery difficult.
B. Continual Learning: Continual learning updates models for new data distributions while mitigating catastrophic forgetting through regularization, rehearsal, distillation, parameter isolation, or test-time adaptation.Examples include EWC, Synaptic Intelligence, iCaRL, and Learning without Forgetting.
B. Continual Learning: Existing continual-learning methods assume adaptation can proceed without violating inference timing and that hardware execution remains reliable during updates.A transient fault in a parameter or gradient update can contaminate the learned model.
B. Continual Learning: Prior hardware-reliability research uses fault injection, vulnerability characterization, selective protection, and reduced-redundancy mechanisms, but does not establish the complete runtime-adaptive deployment addressed here.The cited prior work includes accelerator-level propagation studies and selective hardening for Vision Transformers.

A. Overview of WARD

WARD continuously adjusts Vision Transformer execution to changing power, reliability, and input conditions while maintaining uninterrupted inference. It uses channel-wise partitioned, physically independent subnetworks and lightweight runtime support for mode scheduling and resource management.

A. Overview of WARD: WARD dynamically selects among four operating modes using available computational budget, hardware reliability, and observed input distribution.The controller evaluates these runtime signals continuously to satisfy current system requirements.
A. Overview of WARD: WARD integrates with programmable AI accelerators through lightweight runtime support for mode scheduling and resource management rather than a new accelerator architecture.The implementation demonstrates integration through minimal architectural modifications.
A. Overview of WARD: Independent forward paths and parameter storage enable inference, redundancy, and continual adaptation without interrupting the primary task.Physical separation also allows selective activation, adaptation, or replication according to the selected operating mode.
A. Overview of WARD: Channel-wise partitioning creates two independent subnetworks that preserve the original model's representational capacity through joint fine-tuning.Each subnetwork uses embedding dimension S = D/2, and both are fine-tuned on the target task to maintain complementary features.
A. Overview of WARD: Q, K, and V projections are separated before partitioning to avoid mixing attention heads and violating attention consistency.The resulting subnet-specific QKV matrices use corresponding channel-wise submatrices.

C. Runtime Operating Modes

WARD's four operating modes trade inference accuracy, computational cost, and fault tolerance according to runtime conditions. High-Reliability mode uses isolated dual execution and numerical validation to preserve inference when one path is corrupted.

C. Runtime Operating Modes: WARD's controller selects modes continuously instead of using a fixed pipeline, balancing computational efficiency, reliability, and online adaptation.Selection is based on available power, hardware reliability, and model behaviour.
C. Runtime Operating Modes: Full-Precision mode runs the original Vision Transformer for highest accuracy under benign conditions but provides no redundancy or fault isolation.A fault affecting any model parameter directly propagates to the final prediction.
C. Runtime Operating Modes: Low-Power mode runs one subnetwork with embedding dimension S = D/2, reducing self-attention and MLP computation for power-constrained conditions.Because only one path is active, hardware faults directly affect the inference result.
C. Runtime Operating Modes: High-Reliability mode executes both subnetworks concurrently and averages their logits under fault-free operation.The mode exploits complementary feature representations learned during joint training.
C. Runtime Operating Modes: A hardware fault in one subnetwork does not interrupt inference while the second execution path remains operational.Physical isolation provides fault containment without complete model replication.

4) Adaptive Mode:

Adaptive Mode preserves uninterrupted inference while one physically independent subnetwork learns online from high-confidence pseudo-labels and fault-resilient parameters. Runtime monitoring combines power, hardware reliability, numerical validity, and subnetwork disagreement to support adaptation and fault detection.

4) Adaptive Mode:: Adaptive Mode updates one subnetwork online with high-confidence pseudo-labels while the other continues inference, preserving uninterrupted real-time operation.Updates are restricted to fault-resilient parameters, leaving reliability-critical parameters unchanged during adaptation.
4) Adaptive Mode:: Online adaptation is restricted to the fault-resilient parameter subset identified through offline statistical fault injection analysis.The vulnerability-critical subset remains unchanged while other parameters compensate for distribution shift.
4) Adaptive Mode:: The runtime controller balances computational efficiency, reliability, and adaptation using available power, hardware reliability, and model behaviour.Power is estimated from solar generation and battery state-of-charge, while ECC events report elevated radiation activity.
4) Adaptive Mode:: WARD detects runtime faults by checking non-finite outputs and measuring prediction disagreement between the two physically independent subnetworks.Disagreement is accumulated over a rolling observation window to suppress transient variations.
4) Adaptive Mode:: A non-finite output immediately sets disagreement to one, allowing direct same-step fault detection without waiting for the rolling window to fill.The rolling disagreement threshold is determined from healthy baseline disagreement.

3) Mode Selection:

WARD selects the next operating mode from power availability, hardware reliability, runtime fault status, and consecutive fault-free execution. Power constraints take priority, while adaptive learning begins only after stable fault-free operation.

3) Mode Selection:: The controller evaluates environmental conditions and fault status to select the operating mode for the next inference cycle.The selection inputs include power availability, ECC reliability status, fault detection, and consecutive fault-free steps.
3) Mode Selection:: Power availability receives highest priority, confirmed faults trigger recovery, and elevated radiation activates High-Reliability mode.Adaptive Mode is entered only after a sufficiently long fault-free observation period.
3) Mode Selection:: Adaptive Mode begins only after the system remains fault-free long enough to establish stable operating conditions.This policy delays online learning until the required consecutive fault-free execution count is reached.

4) Recovery:

WARD recovers from hardware faults by switching inference to a healthy subnetwork, restoring the affected path, and updating only eligible parameters. Its PERUN implementation maps the four modes onto lightweight runtime control without changing the computational datapath.

4) Recovery:: After a confirmed fault, WARD redirects inference to the healthy subnetwork, restores the corrupted path from a clean checkpoint, and updates only plastic parameters.Recovery uses confidence-filtered pseudo-labels generated by the healthy subnetwork when sufficient samples are available.
E. Hardware Realization: PERUN maps Full-Precision, Low-Power, High-Reliability, and Adaptive modes directly onto a programmable FPGA accelerator through runtime control support.The implementation targets runtime overhead and deployability rather than accelerator performance itself.
E. Hardware Realization: The Rocket processor manages PERUN through interrupts, control-register updates, and mode-specific instruction and connectivity loading over the AXI interface.This software-assisted management enables rapid operating-mode transitions with lightweight processor intervention.
E. Hardware Realization: The four hardware execution modes preserve their software roles: one path minimizes workload, two paths support voting and monitoring, and background learning continues during inference.Adaptive execution maintains inference on one path while the second performs continual learning.
E. Hardware Realization: WARD requires only lightweight runtime extensions for scheduling, memory allocation, and synchronization, leaving the processing elements and computational datapath unchanged.The runtime-layer design supports integration across a broad class of programmable AI accelerators.

B. Runtime Operating Modes

The four operating modes trade inference accuracy, computational cost, latency, and fault tolerance according to deployment conditions. Full-Precision maximizes accuracy, Low-Power reduces workload, and High-Reliability preserves performance under an injected fault through voting and recovery.

B. Runtime Operating Modes:: The evaluation compares the four modes across accuracy, computational cost, latency, and fault-handling capability under varying deployment conditions.Table I summarizes mode performance, while Table II summarizes fault handling and Figure 6 illustrates accuracy-computation trade-offs.
B. Runtime Operating Modes: Full-Precision mode achieves 96.11% clean Top-1 accuracy and 94.83% shifted accuracy, serving as the highest-accuracy reference at the highest computational cost.It provides no runtime fault tolerance because faults in the single deployed model propagate directly to inference output.
2) Low-Power Mode:: 0.25 relative FLOPs and 84.3 ms latency make Low-Power mode the lowest-workload option, while clean and shifted accuracies remain 90.63% and 85.02%.The active embedding dimension falls from 192 to 96, reducing the active parameter budget to 26% of the original model.
3) High-Reliability Mode:: High-Reliability mode produces 92.80% clean and 86.48% shifted voted accuracy while using approximately half the original computational workload and 167.2 ms latency.Both subnetworks execute concurrently, enabling output voting and runtime fault monitoring.
3) High-Reliability Mode:: 92.17% voted accuracy remains after a single weight fault reduces subnet A to 11.06%, and recovery reaches 92.67% after checkpoint restoration and adaptation.The disagreement monitor detects the fault after seven execution steps and initiates recovery.

4) Adaptive Mode:

Adaptive Mode maintains uninterrupted inference while one isolated subnet adapts online, and WARD’s partition preserves the original network’s fault behavior and storage isolation.

4) Adaptive Mode:: 92.46% clean and 85.76% shifted accuracy are achieved after online adaptation while preserving LP’s 84.3 ms serving latency.Adaptive Mode uses subnet A for inference and subnet B for confidence-filtered adaptation with a 0.88 pseudo-label threshold.
4) Adaptive Mode:: 89.81% output accuracy is maintained during a hardware fault, and recovery restores accuracy to 92.39%.Inference redirects to the healthy path while the affected subnet is restored and re-adapted; Adaptive Mode activates after 25 fault-free steps.
Statistical Fault Injection Validation:: WARD’s network-level failure rate is 1.79% with a 95% confidence interval of [0.23%, 3.34%], overlapping the original Vision Transformer’s reported interval.This supports preservation of intrinsic fault behavior despite halving the embedding dimension.
Statistical Fault Injection Validation:: The final Layer Normalization is the most vulnerable component at 7.4%, followed by encoder block 7 at 6.4%, while subnet B remains byte-identical after faults in subnet A.The layer-wise distribution resembles the original model, and byte-level checks verify physical storage isolation.

2) Reliability-Aware Parameter Protection:

WARD protects reliability-critical parameters identified offline while restricting monitoring to dominant fault locations, preserving adaptation capacity and reducing runtime assessment cost.

2) Reliability-Aware Parameter Protection:: 92.17%–92.98% post-recovery accuracy across protected fractions from 5% to 50% varies by only 0.81%.This is smaller than the 3.31% accuracy difference between Full-Precision and High-Reliability modes.
2) Reliability-Aware Parameter Protection:: A protected fraction near 10% peaks at 0.8446, while the runtime operating point of 18.5% remains within 0.44% of that peak.The vulnerability-guided prefix list freezes 266,218 of 1,436,074 parameters per subnet without explicit hyperparameter search.
2) Reliability-Aware Parameter Protection:: Monitoring exponent most-significant bits in the twelve highest-vulnerability parameter groups reduces candidate fault locations from 45,954,368 to 149,568, a 307× reduction.The reduced space covers the dominant hardware failure mechanism and approximately 0.33% of the original fault space.
Random Bit-Flip Fault Model:: Under uniform random bit flips, High-Reliability and Adaptive modes each record 0% failure across 50 injections, versus 2% for Full-Precision and 4% for Low-Power.The result extends fault-containment evidence beyond targeted exponent most-significant-bit injections.

D. Runtime Adaptation Controller

The runtime controller transitions among operating modes in response to power, radiation, and fault conditions, while the accelerator implementation adds lightweight support without changing baseline timing.

D. Runtime Adaptation Controller: During eclipse, Low-Power mode runs uninterrupted with 90.83% mean evaluated accuracy over steps 30–59, while Adaptive Mode averages 93.16% during steps 24–29.Mode transitions follow solar generation and the 25-step fault-free stability threshold.
D. Runtime Adaptation Controller: A fault injected at step 100 is contained and recovered within one simulation step, while High-Reliability execution resumes without inference interruption.The controller escalates at elevated radiation, uses subnet B for serving during recovery, and all six state-machine validation conditions pass.
E. Hardware Evaluation: High-Reliability execution averages 1.96× overhead, while Full-Precision, Low-Power, and Adaptive modes incur no redundancy penalty.The measured High-Reliability range is 1.79×–2.14×, closely matching duplicate-execution cost.
E. Hardware Evaluation: Compared with static DMR’s more than 100% area overhead and HMR-NEureka’s approximately 9%, WARD requires less than 5% additional hardware.The implementation also preserves OpenCL compatibility and dynamically switches operating modes.
E. Hardware Evaluation: Runtime support adds below 5% hardware overhead, completes mode transitions within 50 cycles, and preserves the accelerator’s approximately 240 MHz maximum operating frequency.The added scheduler, dual-port parameter memory, masking logic, and comparator contribute only lightweight resource increases.

VI. CONCLUSION

WARD combines channel-wise partitioning, runtime mode scheduling, and reliability-aware continual learning for dependable Vision Transformer deployment on programmable edge accelerators. It reports adaptive accuracy, low-power execution, fault-space reduction, and low implementation overhead.

VI. CONCLUSION: WARD combines channel-wise network partitioning, runtime mode scheduling, and reliability-aware continual learning for uninterrupted inference under changing conditions.The framework targets programmable edge AI accelerators and balances accuracy, computational cost, and reliability at runtime.
VI. CONCLUSION: WARD achieves up to 96.11% classification accuracy, 0.25× serving FLOPs in Low-Power mode, and a 1.79% network-level failure rate.The reported results summarize performance, computational cost, and fault tolerance across the framework.
VI. CONCLUSION: The framework reduces the monitored fault space by 307× and supports runtime mode switching in 50 clock cycles with less than 5% hardware overhead.These results support the paper’s conclusion that WARD is a lightweight solution for runtime-adaptive dependable edge AI deployment.