Distributed Training Engineer for Protocol Learning

Openkrill; United States; Remoto; USD 150.000 - 210.000; Tempo pieno

Vantaggi offerti da questo lavoro

Remote-first culture

Visa sponsorship & relocation support

Equity opportunities

Descrizione del lavoro

Pluralis Research is seeking a Research Engineer to build the training system for Protocol Learning, scaling large models across heterogeneous hardware and regional deployments. You will work on distributed pretraining, fault tolerance, and performance optimizations in a remote-friendly setup.

You will contribute to monitoring and tooling to ensure model quality across hundreds of devices, with emphasis on production-quality Python and robust parallelism in PyTorch ecosystems.

Competenze

• Experience with distributed training across many devices using PyTorch and parallelism.
• Production-quality Python and robust concurrency handling.
• Shipped systems, research code, open-source work, or serious personal projects.
• Mission alignment with Protocol Learning and the envisaged third path for sovereign AI.

Mansioni

• Distributed pretraining: Implement and optimize model-parallel training across heterogeneous GPUs.
• Performance optimization to reduce communication while preserving convergence.
• Elasticity and fault tolerance: handle node churn, robust checkpointing and recovery.
• Run instrumentation: build monitoring for throughput, bottlenecks, and model quality across devices.

Conoscenze

Distributed training

PyTorch

FSDP

DeepSpeed

Megatron

Parallelism

Strumenti

Python

Descrizione del lavoro

Pluralis Research is seeking a Research Engineer to build the training system for Protocol Learning, scaling large models across heterogeneous hardware and regional deployments. You will work on distributed pretraining, fault tolerance, and performance optimizations in a remote-friendly setup.

You will contribute to monitoring and tooling to ensure model quality across hundreds of devices, with emphasis on production-quality Python and robust parallelism in PyTorch ecosystems.