Helen Byrne
VP, Applied Ai @Graphcore
Signup · Get unlimited contacts
WORK HISTORY
VP, Applied Ai @Graphcore
GB
I lead and mentor a team of ML engineers focused on pushing the boundaries of efficient AI—from kernel-level to system-level optimization. What we do:• Optimize the latest AI models for our stack: identify bottlenecks, improve compute efficiency and accuracy.• Research and prototype new techniques to scale and accelerate training and inference across modalities and architectures.• Build software that ensures the best out-of-the-box experience for AI developers—from kernels to model deployment.Some things that I care more about.1. Quantization for efficient training and inference. • The challenges and tricks to resolve instabilities with FP8/FP4 and beyond. • The limits of precision for different parts of model storage, comms, compute. • Obtaining optimal utilisation of hardware in training i.e. MFU > 50%+ • Utility of custom non-native formats in inference e.g. INT3 / NF4 / … • Quantization solutions for extreme outliers. The many dimensions of Sparsity. Including: • Efficiency for MoEs, as the most widely-used axis of sparsity. • Sparsity of attention/KV cache. 3.(Other) techniques for efficient autoregressive inference including: • Efficient kv cache management and paging. • Improvements to attention mechanism for decode. • Challenges (and how to solve them) for long-context. • Scheduling and batching strategies. Scaling workloads to of nodes: • Parallelisation strategies. • Identifying/solving bottlenecks inc efficient overlap collectives/compute.5. Next gen model architectures with improved computational complexity vs attention. Kernel optimisation e.g. for efficient attention variants. The SW stack components that enable the best user experience for AI model training and deployment.Strategic:• Guide strategy on near and long-term priorities—spanning SW roadmaps, HW design, product direction.• Present at conferences, customer forums, and developer meetups.
EDUCATION
Universitat Politècnica de Catalunya
Master's degree, Artificial Intelligence
University of Bristol
Bachelor of Science (BSc), Mathematics
Sheffield Hallam University
PGCE (Teach First), Education
SKILLS
ABOUT HELEN BYRNE
I lead the Applied AI organisation at Graphcore - a team of ML Engineers - developing and optimising AI models tailored to our specialised hardware. We build reference applications and develop and publish novel ideas with a focus on efficient compute - making models run faster and consume less power!Some things I care about more than others:1. Quantization for efficient training and inference.2. The many dimensions of Sparsity.3.(Broadly) techniques for efficient autoregressive inference.4. Scaling ML workloads to of nodes. Next generation model architectures with improved computational complexity vs attention. Kernel optimisation e.g. for efficient attention variants. The software stack components that enable the best out-of-the-box experience for AI model training and deployment. Previously, I led AI Field Engineering - a customer-facing AI organisation, and worked in AI Research, tackling problems in distributed machine learning. My background is in Mathematics and I have a Master’s degree in Artificial Intelligence.
This profile is compiled from publicly available professional sources. Unifers is not affiliated with or endorsed by LinkedIn. Request removal of this profile.