Shared systems for research and compute.

MPI one-sided communication (RMA) lets a process write into a peer's memory window directly. mpi-rma puts a typed Rust API over rsmpi and adds a ring transport: payloads go into pre-allocated slots, polling reads local memory, and no sender waits for the receiver.

Safe mode gates the sender with a cumulative ack counter, and a send that would overwrite an unacknowledged slot is refused with Error::Full. Raw mode never refuses: it overwrites and reports the gap through lost(). Slots carry a sequence, length and CRC32, so loss and corruption are observable rather than silent.

Published under MIT. The repository's measurements were taken on a single node; the README warns the advantage may not hold across nodes.

trame is the transport and execution layer of the NERVE neuromorphic runtime. It defines one contract — launches, deployments, frames, lanes, processes and steps — and chooses its implementation when the crate is built: a single process with no transport, an MPI backend, or one CUDA warp per rank.

Because the choice is a Cargo feature, callers and verification fixtures are written once against the same functions, macros and bounds. The device backend is opt-in, so an ordinary build links no CUDA stack.

The contract is a model, not a proof. The repository's conformance run exercises a claim-by-backend table and records only the coverage it exercised, so the implementation qualification is narrower than the contract text.