- Use random session id on calls to AsyncTestRuntime via python bindings.
- Switch to FuturesUnordered for more efficient wait on computation completion
- Support arbitrary fixedpoint precision pairs from python bindings
- More complete support of tensor types kernels
- AsyncSession is now Send (can be sent between threads)
- Enable use of Shape values on non-Host placement types
- Add ZerosOp on a host placement (create a vector of zeros of a given shape)
- Add Tensorflow Keras support
- Add ReLU activation function (for multilayer perceptron and NN)
- Memory usage optimizations and reduced cloning
- Wellformed compiler pass to statically verify a computation with elk: elk compile -p wellformed comp.moose
- TLS support for gRCP networking
- PyTorch neural networks support
- Elk now supports all three computation formats (textual, msgpack, and bincode)
- Ensures tensors are in “standard layout” (i.e. contiguous & row-major)
- Added local file storage to moose modules (reading csv and numpy data files supported)
- start_server convenance method added to gRPC network plugin
- Significant RAM usage improvements throughout the codebase
- Add Multilayer Perceptron (MLP), a type of neural network to pymoose predictors
PrimDeriveSeed
operator renamed toDeriveSeed
PrimPrfKeyGen
operator renamed toPrfKeyGen
- Added
Host
prefix toSeed
,PrfKey
andUnit
- Unified
MeanOp
and fuseRepFixedpointMeanOp
withRingFixedpointMeanOp
- Using BitVec as the data storage for HostBitTensor
- Replace sodium oxide with thread_rng and blake3
- Fix binary LinearClassifier logic
- Uniform textual format for rendezvous and sync keys
- A bugfix to produce less verbose errors
- Upgraded nom to 7.1 to deal with the floating point parser problems
- Switched base array from ndarray::Array to ndarray:ArcArray
- Deprecated
to_bytes/from_bytes
in favour foto_msgpack/from_msgpack
.
This is the first "stable" release with a complete set of public APIs. Future releases are expected to contain instructions for migration from the previous "stable" version.
- Symbolic compilation support
- Sufficient primitives to execute AES decryption
- Sufficient primitives to execute custom models, including linear regression and XGBoost decision trees.
- NA