Uncertainty estimation at scale with functime, Polars and conformal predictions
functime is a modern time-series forecasting library to generate predictions for thousands of time series at once, while never leaving your laptop. Thanks to Polars' powerful query engine, feature extraction and cross-validation are 1-2 orders of magnitude faster. Plus, functime offers a best-of-the-class set of diagnostic tools to further streamline your workflow. In this talk, we'll learn how to use functime to analyse your model and generate blazingly fast prediction intervals using EnBPI, a state-of-the-art conformal prediction framework that is also available in other popular Python packages.
Rust is easy
An intro talk that starts from Python-like code inside a Rust program and iteratively fixes compiler errors to reach valid Rust, showing how Rust catches mistakes early and why its constraints can improve code quality.
LLM Inference Arithmetics: the Theory behind Model Serving
Have you ever asked yourself how parameters for an LLM are counted, or wondered why Gemma 2B is actually closer to a 3B model? You have no clue about what a KV-Cache is? (And, before you ask: no, it's not a Redis fork.) Do you want to find out how much GPU VRAM you need to run your model smoothly? If your answer to any of these questions was "yes", or you have another doubt about inference with LLMs - such as batching, or time-to-first-token - this talk is for you. Well, except for the Redis part.
Is the great data frame showdown finally over? Enter: polars
A deep dive into Polars at PyCon Italia 2023, discussing lazy and eager APIs and where Polars beats or differs from pandas for real-world data workloads.
I know Polars is fast, but my data pipelines are written in pandas!
We all know it by now: Polars is blazingly fast. Yet my pipelines are all written in pandas, and it will just take too much time to rewrite them in Polars... won't it? Turns out, it takes less than thirty minutes to tame this new arctic beast!
How I used Polars to build built functime, a next gen ML forecasting library
Everybody knows Polars revolutionised the dataframe landscape, yet fewer realise that machine learning is next. Thanks to its extreme speed, we can speed up feature engineering by 1-2 orders of magnitude. The true gains, however, span across the whole ML lifecycle, with significantly faster batch inference and effortless scaling (no PySpark required!). Add a best-of-the-class set of tools for feature extraction, model evaluation and diagnostic visualisations and you'll get functime: a next-generation library for ML forecasting. Though time-series practitioners are the primary audience, there's something for all data scientists. It's not just forecasting: it's about building the next generation of machine learning libraries.
functime: a next generation ML forecasting library powered by Polars
Polars is mature, production ready, intuitive to write and pleasant to read. And it is fast. Thanks to Rust and Rayon, you can achieve speeds greater than numba's. If you combine it with top-of-the-class evaluation methods, not only can you get speedups of about 1-2x order of magnitude in feature engineering and cross-validation, but also dramatically improve your development workflow. That's what we set out to demonstrate with functime. We chose to write a time-series library first, because forecasting can be a costly undertaking, with significant problems of scale. Making predictions with big panel datasets usually required fitting thousands of univariate models, one at a time, using distributed systems. On the other hand, functime unlocks an efficient forecasting workflow, from your laptop. This talk is a hands-on demonstration for forecasting practitioners and data scientists alike. It will showcase how to build clean and performant forecasting pipelines with rich feature-engineering capabilities, enabling a seamless and more efficient modelling workflow. Nevertheless, the principles behind functime can be grasped by every machine learning practitioner: forecasting is just a use-case to show off Polars' potential. With Polars, we can improve the current state of machine learning modelling and raise the ceiling for what reasonable scale means.
Embeddings, Transformers, RLHF: Three key ideas to understand ChatGPT
ChatGPT has become a groundbreaking tool, transforming how professionals in various industries work. However, while many articles focus on "the 30 prompts everybody needs to know", they often overlook the underlying technology of ChatGPT. To truly understand ChatGPT, it's important to comprehend three key concepts: Embeddings (how LLMs convert words and phrases into numerical values), Transformers (deep-learning modules that capture semantic connections), and RLHF (Reinforcement Learning with Human Feedback) to align models with intended purposes and ethical standards. The talk also covers the four primary steps involved in building and training a GPT-like model, plus limitations and strengths of current generative AI models and actionable insights for safe adoption.
Build your first Python package with Rust
Are you tired of that colleague of yours that boasts about how great of a package manager cargo is, and how Python packaging ecosystem sucks in comparison? Or do you want to know how Python libraries with a Rust core like ruff and polars are built? You've come to the right place: we will guide you trough how to use PyO3 and maturin to package and publish your first, blazingly fast, rust-powered Python package!
OpenAI o1/o3: the New Scaling Laws for LLMs that can reason
With o1, OpenAI ushered a new era of LLMs: reasoning capabilities. This new breed of models broadened the concept of scaling laws, shifting focus from **train-time** to **test-time** (or inference-time) compute. How do these models work? What do we think their architectures look like, and what data do we use to train them? And finally - and perhaps more importantly: how expensive can they get, and what can we use them for?
From OpenAI to DeepSeek: New Scaling Laws for LLMs that can Reason
With o1, OpenAI ushered a new era: LLMs with reasoning capabilities. This new breed of models broadened the concept of scaling laws, shifting focus from train-time to inference-time compute. But how do these models work? What do we think their architectures look like, and what data do we use to train them? And finally - and perhaps more importantly: how expensive can they get, and what can we use them for?
Bigger models or more data? The new scaling laws for LLMs
The incredibly famous Chinchilla paper changed the way we train LLMs. The authors - including the current Mistral CEO - outlined the scaling laws to maximise your model performance under a compute budget, balancing the number of parameters and training tokens.
Today, these heuristics are in jeopardy. LLaMA-3, for one, is trained on an unreasonable amount of tokens of text - but this is why it's so good. How much data do we actually need to train LLMs? This talk will shed light on the latest trends in model training and perhaps suggest newer scaling laws.
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top