Member of Technical Staff, Distributed Training Performance
Hark · San Jose · HQ
Frontier labonsite$180k–$450k
At a glance
Highlights
- Own MFU and step time for dense and mixture-of-experts pretraining runs.
- Build seamless real-time multimodal intelligence across text, audio and beyond.
- Reduce lost training work with faster checkpointing, restarts and observability.
Ideal candidate
A performance-obsessed systems engineer who loves squeezing maximum throughput from massive GPU runs and partnering with researchers to keep large-scale training fast and reliable.
Requirements
- Experience developing distributed training systems and working with internals of Megatron-LM, NeMo, or comparable framework
- Strong understanding of data, tensor, pipeline, sequence, context, and expert parallelism
- Experience using GPU profilers to analyze kernel execution, memory, communication and data-loading stalls
- Working knowledge of mixed-precision training including BF16 and FP8 and diagnosing numerical instability
- Strong Python and C++/CUDA engineering skills building reliable tools for researchers
- Experience developing CUDA or Triton kernels or using CUTLASS for attention, normalization or MoE
- Record of measurably improving MFU or throughput on large-scale training jobs
Lab profile
About Hark
Building proactive, personalized multimodal AI with next-generation hardware for human-machine interaction.
Known for
- Proactive personalized intelligence
- Multimodal speech, text and vision
- Persistent memory
- Next-generation AI hardware
Posted · Last seen