Represents an infrastructure incident.

Multi-method anomaly detection engine combining: 1. Statistical methods (Z-score, IQR) 2. Moving average deviation 3. Rate-of-change

Configuration for LLM backends, read from environment variables.

Custom Gymnasium environment for GPU workload scheduling. State space: - Per-node: GPU utilization, memory usage, CPU usage, network

Unified LLM client supporting multiple backends. Attempts providers in priority order, falls back gracefully. Capabilities: - Infe

A single remediation action with command and risk assessment.

============================================================================ Models ==================================================================

Specialized anomaly detector for GPU workloads. Combines statistical methods (Z-score, IQR, EMA, RoC) with deep learning models (LSTM Autoenc

============================================================================ Priority Queue with Preemption ==========================================

Trains and manages the RL scheduling model.

Complete incident analysis result.

LLM-powered operations agent for intelligent incident management. Uses LLM for root cause analysis and remediation suggestions, falls bac

Threshold detection + LLM-powered security analysis.

GPUType defines supported GPU accelerator types

Multi-factor scoring + LLM reasoning for scheduling decisions.

WorkloadType defines AI workload categories

CloudProviderType defines supported cloud providers

Dynamic cost analysis + LLM-powered recommendations.

Production Gymnasium environment for GPU workload scheduling. Observation (continuous, Box): Per-node (N nodes x 6 features):

RuntimeType defines supported Wasm runtimes

============================================================================ WASI Sandbox Isolation ==================================================

============================================================================ Metadata carrier — adapts gRPC metadata to TextMapCarrier ===============

AgentType defines the type of AI agent

AlertSeverity defines alert severity levels

NodeTier defines the compute tier

Proximal Policy Optimization trainer for GPU scheduling. PPO is on-policy, stable, and well-suited for: - High-dimensional continuous ac

============================================================================ Multi-Language Support ==================================================

Soft Actor-Critic trainer for GPU scheduling. SAC is off-policy, sample-efficient, and well-suited for: - Continuous action spaces -

Configuration for the tracing subsystem.

Custom SB3 callback for logging scheduling-specific metrics.

============================================================================ Admission Control Chain =================================================

============================================================================ Error Budget Policy =====================================================

============================================================================ Circuit Breaker — Three-state pattern (Closed → Open → HalfOpen → Closed)

Manages fine-tuning jobs on cloud LLM providers. Supports OpenAI fine-tuning API and Alibaba DashScope.

ClusterStatus represents cluster health state

Production-grade ensemble combining statistical and deep learning methods. Voting strategy: - Each method produces a score in [0, 1]

Generates fine-tuning datasets from domain knowledge. Domains: - scheduling: GPU workload placement decisions with reasoning

EdgeNodeStatus represents edge node health

EnforcementMode defines how a policy is enforced

EventType classifies real-time events.

Compare base model vs fine-tuned model on domain-specific tasks.

End-to-end fine-tuning orchestrator. Usage: pipeline = FineTuningPipeline() # Prepare domain datasets pipeline.prepa

============================================================================ GGUF: llama.cpp Quantization Formats Supports Q2_K, Q3_K_S/M/L, Q4_K_S/M,

LSTM Autoencoder for multivariate time-series anomaly detection. Training: online learning via sliding window reconstruction. Inference:

Local LoRA/QLoRA fine-tuning using HuggingFace Transformers + PEFT. Enables domain adaptation of open-source models (Qwen, LLaMA, Mistral)

============================================================================ OPA / Gatekeeper Policy Engine ==========================================

============================================================================ Models ==================================================================

SwapState represents the state of a hot-swap operation

Transformer-based anomaly detection with reconstruction + attention scoring. Advantages over LSTM: - Better at capturing long-range depe

WorkloadStatus represents Wasm workload lifecycle

Adapter for extracting trace context from ASGI/Starlette headers.

LSTM decoder that reconstructs the time window from a latent vector.

LSTM encoder that compresses a time window into a latent vector.

Sinusoidal positional encoding for transformer input.

Transformer encoder with reconstruction head for anomaly detection.

ABACConfig configures the ABAC engine.

ABACDecision is the result of policy evaluation.

============================================================================ ABAC Engine =============================================================

============================================================================ ABAC Policy Model =======================================================

============================================================================ ABAC Request Context ====================================================

============================================================================ ACK API Types ===========================================================

ACKNode represents a node from DescribeClusterNodes API

ACKNodePage wraps the paginated node list response

============================================================================ AKS API Types ===========================================================

AKSCredentials holds the kubeconfig credential block

AKSNodePool represents an AKS agent pool (node group)

APIErrorBody is the inner error object.

============================================================================ Unified API Error Response ==============================================

============================================================================ API Key Model ===========================================================

============================================================================ API Response Types ======================================================

============================================================================ AWQ: Activation-Aware Weight Quantization Reference: Lin et al., "AWQ: Ac

AWQResult contains AWQ quantization output metrics.

============================================================================ AWS EKS Client (Official SDK v2) ========================================

============================================================================= AWSProvider implements Provider for Amazon Web Services (EKS/SageMaker)

============================================================================ Cross-AZ Deployment Strategy ============================================

AccessRecord tracks an access event for pattern learning

ActionItem tracks follow-up tasks from a retrospective.

ActionMatch defines the actions to match.

============================================================================ Adaptive TTL — hot keys live longer, cold keys expire faster ============

AdaptiveTTLManager tracks access patterns and computes optimal TTLs.

AdmissionAuditEntry records an admission evaluation for audit

AdmissionChain manages an ordered set of admission webhooks

AdmissionDecision is the result of evaluating an admission request

============================================================================ Admission Extension Hub — Unified Entry Point ===========================

AdmissionRequest represents a resource admission request

AdmissionResponse represents the response from admission evaluation