Senior Machine Learning Infrastructure Engineer (Precision, Diagnostics & Hardware)

Veeda AI; Toronto; On-site; CAD 130,000 - 185,000; Full time

Job summary

Veeda AI is building the next generation of multimodal foundation world models for Physical AI. We seek engineers to design and optimize distributed training systems across large GPU clusters, handling FP16/BF16/FP8 precision and debugging complex stability issues.

You will implement fault-detection, performance profiling, and resilient checkpointing to keep researchers productive in a fast-moving environment. Collaboration across teams is essential for success.

Qualifications

• Bachelor's degree or equivalent hands-on experience in Computer Science, Computer Engineering, or a related technical field.
• Experience with deep learning training frameworks (e.g., PyTorch) and distributed training paradigms (FSDP, Megatron-LM, DeepSpeed, Tensor Parallelism, Pipeline Parallelism).
• Proven experience in numerical precision analysis, low-precision training (BF16/FP8), and debugging complex loss divergence/stability issues in massive training runs.
• Strong root-cause analysis skills for hardware/software interaction bugs, including stuck CUDA kernels, NCCL timeouts, GPU hardware faults, and silent training corruptions.

Responsibilities

• Design, optimize, and maintain high-throughput distributed training systems across large-scale GPU clusters for multi-modal foundation models.
• Debug, diagnose, and resolve subtle numerical instability issues (underflow/overflow, loss spikes, gradient explosion, and mixed-precision divergence) in FP16, BF16, FP8, and custom quantization schemes.
• Build advanced fault-detection mechanisms and automated diagnostics to rapidly pinpoint and isolate silent data corruption (SDC), hardware hang/deadlock, memory leaks, and card-freeze issues during large training runs.
• Profile distributed communication bottlenecks, memory usage, and kernel execution to improve overall FLOPS utilization across multi-node, multi-GPU training jobs.
• Develop resilient checkpointing systems, rapid fault-recovery pipelines, and execution telemetry to keep researcher productivity high and hardware downtime minimal.

Skills

Python

C++/CUDA

PyTorch

Distributed training

Megatron-LM

DeepSpeed

Tensor Parallelism

Pipeline Parallelism

NCCL

Low-level GPU

CUDA kernels

Education

Bachelor's degree in CS/CE

Tools

Triton kernels

NCCL

DeepSpeed

Megatron-LM

PyTorch

Job description

About Us

Veeda AI is building the next generation of multimodal foundation world models for Physical AI. We're a small, fast-moving team of engineers and researchers from leading AI labs, tackling some of the most challenging problems at the intersection of AI, robotics, and embodied intelligence. If you're excited about pushing the boundaries of what's possible with Physical AI, you'll have the opportunity to make an outsized impact from day one.

Responsibilities

• Distributed Training Systems & Scalability: Design, optimize, and maintain high-throughput distributed training systems across large-scale GPU clusters for multi-modal foundation models.

• Precision & Numerical Stability: Debug, diagnose, and resolve subtle numerical instability issues (underflow/overflow, loss spikes, gradient explosion, and mixed-precision divergence) in FP16, BF16, FP8, and custom quantization schemes.

• Fault Diagnostics & Recovery: Build advanced fault-detection mechanisms and automated diagnostics to rapidly pinpoint and isolate silent data corruption (SDC), hardware hang/deadlock, memory leaks, and "card-freeze" issues during large training runs.

• Performance Profiling & Optimization: Profile distributed communication bottlenecks, memory usage, and kernel execution to improve overall FLOPS utilization across multi-node, multi-GPU training jobs.

• Developer Tooling & Infrastructure: Develop resilient checkpointing systems, rapid fault-recovery pipelines, and execution telemetry to keep researcher productivity high and hardware downtime minimal.

Requirements

• You have a Bachelor's degree or equivalent hands-on experience in Computer Science, Computer Engineering, or a related technical field.

• You have deep hands-on experience with deep learning training frameworks (e.g., PyTorch) and distributed training paradigms (FSDP, Megatron-LM, DeepSpeed, Tensor Parallelism, Pipeline Parallelism).

• You have proven experience in numerical precision analysis, low-precision training (BF16/FP8), and debugging complex loss divergence/stability issues in massive training runs.

• You have strong root-cause analysis skills for hardware/software interaction bugs, including stuck CUDA kernels, NCCL timeouts, GPU hardware faults, and silent training corruptions.

• You have strong programming skills in Python and C++/CUDA, with a deep understanding of low-level GPU architectures and memory hierarchies.

Nice to Have

• You have experience running or porting large-scale training workloads on AMD GPUs (ROCm platform) or Google TPUs (JAX/XLA stack).

• You have contributed to low-level training infrastructure, custom CUDA/Triton kernels, or distributed training open-source projects.

• You have built resilient fault-tolerant training frameworks with dynamic node re-queueing and rapid checkpointing/saving mechanisms.