Expand description
Meganeura: graph-optimized neural network framework on blade-graphics.
Models are defined as declarative computation graphs, optimized with greedy rewrites by default (or optional equality saturation via egglog), and compiled to static GPU dispatch sequences — no manual CUDA-graphing needed.
Re-exports§
pub use data::DataLoader;pub use data::MnistDataset;pub use graph::DType;pub use graph::Graph;pub use graph::NodeId;pub use graph::TensorType;pub use load::nnef::NnefError;pub use load::nnef::NnefModel;pub use load::nnef::load_nnef;pub use load::onnx::OnnxError;pub use load::onnx::OnnxModel;pub use load::onnx::load_onnx;pub use load::onnx::load_onnx_bytes;pub use optimize::ExtractionCost;pub use optimize::OptimizeConfig;pub use optimize::OptimizeMode;pub use optimize::OptimizeReport;pub use runtime::CoopPolicy;pub use runtime::DebugStepReport;pub use runtime::DeviceMemoryStats;pub use runtime::DispatchAnomaly;pub use runtime::ExternalBindError;pub use runtime::ExternalSlot;pub use runtime::GpuOptions;pub use runtime::MemorySummary;pub use runtime::ReadNodeError;pub use runtime::Session;pub use runtime::SessionOptions;pub use runtime::TuneOutcome;pub use runtime::init_gpu_context;pub use runtime::init_gpu_context_with;pub use train::EpochStats;pub use train::LossHistory;pub use train::MetricCallback;pub use train::Mode;pub use train::Optimizer;pub use train::SessionConfig;pub use train::StepMetrics;pub use train::TrainConfig;pub use train::TrainHistory;pub use train::Trainer;pub use train::build;pub use train::build_inference_session;pub use train::build_session;pub use train::build_session_unoptimized;pub use train::compile_training_graph;pub use tune::MatmulTile;pub use tune::TuneClass;pub use tune::TuneConv2d;pub use tune::TuneDecision;pub use tune::TuneError;pub use tune::TuneOptions;pub use tune::TunePhaseTimes;pub use tune::TunePreparationTimes;pub use tune::TuneQualificationTimes;pub use tune::TuneReport;pub use tune::TuneScope;pub use tune::TuneScratchStats;pub use tune::TuneScratchUsage;pub use tune::TuneStaging;pub use tune::TuneStagingReuse;
Modules§
- config
- Central registry of every
MEGANEURA_*environment variable — and the ONLY place that reads them. - data
- Data loading utilities for training.
- eager
- Eager evaluation: inspect any node of a graph while you build it.
- graph
- load
- Model loading from standard interchange formats.
- models
- nn
- High-level neural network building blocks.
- optimize
- Graph optimization with deterministic greedy rewrites by default and optional equality saturation through egglog. The current local rule set reaches the same useful forms with greedy rewriting at much lower build cost. Equality saturation retains alternatives for extraction using an expression-size or estimated tensor-traffic objective.
- profiler
- Profiling infrastructure producing Perfetto binary traces (
.pftrace). - runtime
- train
- tune
- Bounded, opt-in kernel selection. See
crate::Session::tune_with.
Structs§
- Compile
Options - Options controlling graph → execution-plan compilation.
- Coop
Caps - Generate a Flash Attention 2 forward kernel with BQ>1 multi-query tiling.
- Tuning
Knobs - Performance-tuning knobs that shape generated kernels and dispatch
geometry. A knob is data: it lives in
CompileOptions, is stamped into the compiledExecutionPlan(geometry and generated WGSL must agree), and participates in the plan-cache fingerprint automatically. Defaults come from capability-signature heuristics plusMEGANEURA_FLASH_*env overrides; a session-build tuner can substitute measured values instead.