Module hardware_aware

Structs§

AdaptiveTuner: Adaptive tuner for dynamic optimization
HardwareAwareOptimizer: Hardware-aware optimizer that adapts to different platforms
HardwareOptimizationConfig: Hardware-specific optimization configuration
HardwarePerformanceStats: Hardware performance statistics
HardwareStepReport: Outcome of a single call to OptimizationState::step.
OptimizationState: Current optimization state: parameters plus everything needed to move them.
PerformanceProfiler: Performance profiler for hardware-specific metrics
ResourceMonitor: Resource monitor for real-time hardware monitoring
TunableParameter: A parameter the tuner is allowed to move, and the range it may move it in.
TuningObservation: What the evaluator reports back for one candidate parameter set.
TuningOutcome: Result of one AdaptiveTuner::tune call.
TuningRecord: Tuning record for adaptive optimization

AllReduceAlgorithm: All-reduce algorithms
CommunicationStrategy: Communication strategies for distributed training
GPUArchitecture: GPU architectures
HardwareOptimizerKind: Optimizer families the hardware analysis can recommend.
HardwarePlatform: Hardware platform types
MemoryStrategy: Memory optimization strategies
ParallelizationStrategy: Parallelization strategies
PartitionStrategy: Model partitioning strategies
PrecisionStrategy: Precision strategies for different hardware
QuantizationSupport: Quantization support levels
SIMDSupport: SIMD instruction set support
TPUVersion: TPU versions
TuningStrategy: Tuning strategies

DEFAULT_BASE_LEARNING_RATE: Learning rate used when the caller does not supply one.

tuned_value_as_f64: Read a tuned parameter back as an f64, for callers that need to map it onto an integer configuration field such as a batch size.