Overview
An MPI coordination scaffold in C++17 with a REST monitoring dashboard on port 8080. It ranks processes, reduces gradients across ranks, and broadcasts from rank 0. The learning is not implemented: batch gradients are index-derived, the parameter update is a no-op, and the data is randomly generated.
Architecture
Three layers: DistributedTrainer owns the MPI collectives and the training loop, the benchmark files hold the process coordination, and the CLI parses arguments and aggregates metrics. There is no tensor library and no model.
Capabilities
- MPI rank coordination across three-rank test runs
- Gradient all-reduce and rank-0 broadcast
- REST dashboard and metrics endpoint on port 8080
- Kubernetes manifests and a Helm chart
Installation
git clone https://github.com/coccinella-labs/ml