ML Performance Engineer - Distributed Training & Inference

Veeda AI; Zürich; On-site; CHF 120,000 - 160,000; Full time

Job summary

Veeda AI in Zürich, Switzerland is building the next generation of multimodal world models for Physical AI. The role focuses on optimizing multi-node video world model training and inference, managing tensor, context, and expert parallelism across PyTorch FSDP2 and Megatron-Core, not default options.

You will profile end-to-end latency, GPU utilization, memory pressure, and kernel efficiency using PyTorch Profiler, Nsight Systems, and torch.utils.benchmark, then improve throughput via

Qualifications

• Hands-on experience with PyTorch and large-scale parallelism stacks on multi-node clusters.

Responsibilities

• Own step time and model FLOPs utilization for multi-node video world model training.
• Profile end-to-end latency, GPU utilization, memory pressure, and kernel efficiency for inference throughput.
• Tuning CUDA and Triton kernels to improve throughput and scaling across GPUs.
• Develop and integrate fault-tolerant and asynchronous checkpointing.
• Tune NCCL collectives and overlap across NVLink domains.
• Diagnose silent data corruption and stuck kernels, enabling fast recovery.

Skills

Python

C++/CUDA

PyTorch

FSDP2

Megatron-Core

TorchTitan

DeepSpeed

Nsight Systems

torch.compile

CUDA Graphs

Education

Bachelor's degree in Computer Science/Engineering or related field

Tools

Nsight Systems

PyTorch Profiler

Triton

CuTe-DSL

Torch.compile

CUDA

Job description

Veeda AI in Zürich, Switzerland is building the next generation of multimodal world models for Physical AI. The role focuses on optimizing multi-node video world model training and inference, managing tensor, context, and expert parallelism across PyTorch FSDP2 and Megatron-Core, not default options.

You will profile end-to-end latency, GPU utilization, memory pressure, and kernel efficiency using PyTorch Profiler, Nsight Systems, and torch.utils.benchmark, then improve throughput via