Microservice architecture has become the de-facto standard for developing cloud-native applications, yet its complex inter-service dependencies make systems highly fragile to cascading failures. Resilience testing, which validates system behavior by injecting various faults, is therefore critical to improve the robustness of target systems. We present FaultWeave, a practical and effective resilience testing framework with failure diagnosis capability for microservice applications. Based on the small scope hypothesis, Fault-Weave designs an efficient fault space exploration technique which incrementally explores fault combinations up to a bounded depth, taking full advantage of previous fault injection results to speed up test execution and reduce redundant test scenarios. This incremental strategy naturally identifies Minimal Failure Sets (MFS)—the smallest fault combinations that trigger resilience failures—which provide structured differential profiles for LLM-assisted failure diagnosis. Industrial deployment on an enterprise-level cloud-native application (containing 512 microservices) at a large power grid company in China over three months discovered 237 resilience vulnerabilities. The evaluation demonstrates significant improvements in testing efficiency and discovered resilience failures compared to existing manual practices.
Feed-forward networks (FFNs) dominate memory traffic and computation in large language model (LLM) inference, making them a primary target for activation sparsification. However, existing training-free methods suffer substantial model-quality degradation at high sparsity due to limitations in their channel-selection strategies. We observe that the SwiGLU intermediate state provides a highly effective channel-selection signal, but obtaining it requires costly dense computation. To address this, we present \emph{Prox}, a two-stage training-free framework for sparse SwiGLU FFNs. Prox hinges on the key insight: sparse execution requires only the channel mask induced by the intermediate state, which can be constructed from the magnitude ranking of its entries rather than their exact values. Specifically, Stage 1 uses input sparsity and quantized proxy weights to construct a shared mask; Stage 2 computes the selected channels exactly, enabling sparse execution of all three projections. Across ten LLMs from six model families, Prox outperforms training-free baselines at all sparsity levels, achieves up to a $1.99\times$ end-to-end decoding speedup at 70\% FFN sparsity, and is compatible with quantization and sparse attention.
Jinyi Liu, Wei Chen, Pengyu Chen et al.· 0 citations
The first systematic empirical study of defects introduced during this stage of deep learning compilers in TorchDynamo, the default DLC frontend for PyTorch 2, the most popular DL framework is conducted, using a domain-knowledge-enhanced LLM-aided methodology.