Research Scientist / Engineer – Training Infrastructure

Luma; Redwood City (CA); On-site; USD 210,000 - 260,000; Full time

Job summary

Luma is seeking an engineer to design and optimize distributed training systems for multimodal models across thousands of GPUs. You will tackle advanced parallelism, training stability, and utilization at scale.

You will work with PyTorch, CUDA, NCCL, and Tensor Parallel, building monitoring and tooling for large-scale runs. This role suits someone who has shipped foundation-model training at scale and can improve stability and efficiency on massive clusters.

Qualifications

• Experience in extensive distributed PyTorch training and parallelisms for foundation-models.
• Deep understanding of GPU clusters, networking, and storage systems.
• Familiarity with communication libraries (NCCL, MPI) and distributed-system optimization.

Responsibilities

• Design, implement, and optimize distributed training systems for models across thousands of GPUs.
• Research and implement advanced parallelization (FSDP, Tensor Parallel, Pipeline Parallel, Expert Parallel).
• Build monitoring, visualization, and debugging tools for large-scale training runs.
• Optimize training stability, convergence, and resource utilization across massive clusters.
• Immerse & diagnose training stacks, land a parallelization or stability improvement, and scale tooling for reliability.

Skills

Distributed PyTorch training

GPU clusters & networking

Distributed-system optimization

Tools

NCCL

MPI

Job description

You’ll build the distributed systems that train Luma's large-scale multimodal models across thousands of GPUs, so researchers can focus on innovation on top of reliable, efficient, scalable infrastructure.

This is hard PyTorch, CUDA, and distributed-systems work - advanced parallelism, training stability, and utilization across massive clusters. It fits an engineer who's solved real problems training foundation models at scale. If you haven't worked at the level of FSDP and multi-node training, this is the wrong depth.

What You’ll Own

• Design, implement, and optimize efficient distributed training systems for models across thousands of GPUs.
• Research and implement advanced parallelization (FSDP, Tensor Parallel, Pipeline Parallel, Expert Parallel).
• Build monitoring, visualization, and debugging tools for large-scale training runs.
• Optimize training stability, convergence, and resource utilization across massive clusters.

First 90 Days

• Days 1-30 - Immerse & Diagnose: Learn the current training stack and where stability and utilization hurt at scale.
• Days 30-60 - Ship & Validate: Land a parallelization or stability improvement that measurably helps a real training run.
• Days 60-90 - Scale & Systemize: Build the monitoring and tooling that keeps large runs reliable and efficient.

What You Bring

• Extensive distributed PyTorch training and parallelisms in foundation-model training.
• Deep understanding of GPU clusters, networking, and storage systems.
• Familiarity with communication libraries (NCCL, MPI) and distributed-system optimization.

Nice to Have

• Strong Linux systems administration and scripting.
• Experience managing training runs across 100+ GPUs.
• Experience with containerization, orchestration, and cloud infrastructure.

About Luma: Luma’s mission is to build unified general intelligence that can generate, understand, and operate in the physical world. We believe multimodality is critical for intelligence - the next step beyond language models comes from vision. Luma is an equal opportunity employer.