Large time-series models (LTMs) enjoy similar power-law scaling behaviour to LLMs. We just put out a paper (https://lnkd.in/givY528D) establishing power-law like scaling-laws for large time-series models as a function of data, compute, and model size. Similar scaling-laws for LLMs (from the landmark Kaplan et al. paper https://lnkd.in/g9KHYN9u) have provided key guidance in allocating enormous resources for predictable - and eventually breakthrough - performance gains. The demonstration of similarly favourable scaling behaviour for large time-series models provides both a motivation and guide in the pursuit of foundation models for time-series forecasting. Foundation models for time-series are coming (with enough data and compute). Thanks to Thomas Edwards, James Alvey, Benjamin Wandelt and Nam Nguyen for the hard work!
Justin Alsing’s Post
More Relevant Posts
-
Large time-series models (LTMs) are to time-series data what LLMs are to language. LTMs are on the cusp of delivering state-of-the-art forecasting capability across a wide range of domains, changing the way we analyse time-series data forever. Calda AI researchers Justin Alsing and Benjamin Wandelt (with collaborators from Johns Hopkins, University of Amsterdam and Capital One) made an important step forward in the pursuit of LTMs last week, demonstrating favourable performance scaling with model size, data, and compute. Watch this space for breakthrough foundation models for time-series forecasting!
Large time-series models (LTMs) enjoy similar power-law scaling behaviour to LLMs. We just put out a paper (https://lnkd.in/givY528D) establishing power-law like scaling-laws for large time-series models as a function of data, compute, and model size. Similar scaling-laws for LLMs (from the landmark Kaplan et al. paper https://lnkd.in/g9KHYN9u) have provided key guidance in allocating enormous resources for predictable - and eventually breakthrough - performance gains. The demonstration of similarly favourable scaling behaviour for large time-series models provides both a motivation and guide in the pursuit of foundation models for time-series forecasting. Foundation models for time-series are coming (with enough data and compute). Thanks to Thomas Edwards, James Alvey, Benjamin Wandelt and Nam Nguyen for the hard work!
To view or add a comment, sign in
-
-
Co-Founder at Innsyte. AI transformation services. #AITransformation #AIDemocratisation #AIGovernance
💫 💫 💫 Once upon a time, there was a not-so-faraway land of the internet 🕸 , filled with endless streams of data. This land was woven with intricate webs of information, much like a tapestry made of every click, like, purchase, and action that people did in the digital space. 🧙♂️ In this land, the wise wizards of Cleora had a special tool, a magical lens named "Cleora PRO", kept a special place: ✨ app.cleora.ai ✨ This lens wasn't just any ordinary tool; it had the power to see the hidden connections in the great tapestry of data. Every day, the people of the internet land would go about their tasks, leaving trails of data like sparkling stars. These trails, however, were invisible to the common eye and could only be seen and understood through the Cleora PRO lens. 🧙♀️ With a gentle wave of their hands, the wizards could use the Cleora PRO to reveal the hidden patterns in the data. It was like watching a dance of lights, showing how different points in the tapestry were connected. This helped the wizards understand what the people of the internet land might like, what they would do next, and how everything was interconnected in a beautiful, intricate web. The best part about Cleora PRO was that it was not just for the wizards. Anyone in the land – be it a young apprentice, a curious explorer, or even a friendly baker – could use it. All they needed was a special key, which they already had in their pockets: a login to their Google, Microsoft, or Github account. 🗝 With this key, they could step into the world of Cleora, look through the magical lens, and see the wonders of the data tapestry themselves. They could find hidden treasures of knowledge, understand the secret wishes of people, and create new, wonderful experiences in the internet land. And so, every night, as the stars of data twinkled in the vast digital space, the land slept peacefully, knowing that the magical lens of Cleora PRO was there to make sense of the world, bringing understanding and connection to all. And they all lived data-informed ever after. 👉 👉 👉 Register now to see the magic: https://www.cleora.ai/ 🔥 Cleora PRO Beta Launch: Traditionally, graph embeddings required advanced knowledge and significant compute resources. Cleora PRO changes that! Our new self-service SaaS platform simplifies graph embeddings. No need for high-end hardware or deep knowledge of graph theory. 🛠️ Applications: From recommender systems to client segmentation, Cleora's graph embeddings enhance predictive models, enabling you to easily find similar entities by comparing vector distances. Cleora can represent users, products and other entities, facilitating answers to the eternal question: "What will my clients want to do next?" 👉 Start Experimenting Now: Simply export data from your database and log in with your Google, Microsoft, or Github account. Discover graph embeddings with Cleora.ai PRO beta! app.cleora.ai #datascience #ai #bigdata #recommendersystems #graphs
To view or add a comment, sign in
-
Data Mobility For AI | AI Compute | GPU Cloud | AI Cloud Infrastructure Engineering Leader, AI-Ready Data Centers | Cloud,AI/HPC Infra Solutions | Sustainability
Key LLM workloads with docs.ray.io and Anyscale:
Excited to share our end-to-end LLM workflows guide that we’ve used to help our industry customers fine-tune and serve OSS LLMs that outperform closed-source models in quality, performance and cost. Key LLM workloads with docs.ray.io and Anyscale: - 🔢 Preprocess our dataset (filter, schema, etc.) with batch data processing. - 🛠️ Fine-tune our LLMs (ex. Meta Llama 3) with full control (LoRA/full param, compute, loss, etc.) and optimizations (parallelism, mixed precision, flash attn, etc.) with distributed training. - ⚖️ Evaluate our fine-tuned LLMs with batch inference using Ray + vLLM. - 🚀 Serve our LLMs as a production application that can autoscale, swap between LoRA adapters, optimize for latency/throughput, etc. Key Anyscale infra capabilities that keeps these workloads efficient and cost-effective: - ✨ Automatically provision worker nodes (ex. GPUs) based on our workload's needs. They'll spin up, run the workload and then scale back to zero (only pay for compute when needed). - 🔋 Execute workloads (ex. fine-tuning) with commodity hardware (A10s) instead of waiting for inaccessible resources (H100s) with data/model parallelism. - 🔙 Configure spot instance to on-demand fallback (or vice-versa) for cost savings. - 🔄 Swap between multiple LoRA adapters using one base model (optimized with multiplexing). - ⚡️ Autoscale to meet demand and scale back to zero. 🆓 You can run this guide entirely for free on Anyscale (no credit card needed). Instructions in the links below 👇 🔗 Links: - Blog post: https://lnkd.in/gvPQGzjh - GitHub repo: https://lnkd.in/gxzzuFAE - Notebook: https://lnkd.in/gmMxb36y
To view or add a comment, sign in
-
-
This great end to end LLM workflows guide is a gem. Many industry players want to scale various parts of this whole workflow while maintaining control and flexibility offered by open source solutions (e.g. Llama 3). Anyscale and Ray’s generic approach towards building AI infra simplifies the process, making it easier for companies to keep iterating with AI and rapidly move their tailored solutions to production. Anyscale’s platform takes care of a lot of things: Provisioning the approapriate hardware for the workload (both spot and on demand), scaling different parts independently, fast startup time, scale-to-zero, sleek UX with VSCode, jupyter, ray dashboard, distributed logs viewer all integrated in the same environment. By going through this structured walkthrough you can FEEL what it is like to develop on Anyscale without worrying about infrastructure.
Excited to share our end-to-end LLM workflows guide that we’ve used to help our industry customers fine-tune and serve OSS LLMs that outperform closed-source models in quality, performance and cost. Key LLM workloads with docs.ray.io and Anyscale: - 🔢 Preprocess our dataset (filter, schema, etc.) with batch data processing. - 🛠️ Fine-tune our LLMs (ex. Meta Llama 3) with full control (LoRA/full param, compute, loss, etc.) and optimizations (parallelism, mixed precision, flash attn, etc.) with distributed training. - ⚖️ Evaluate our fine-tuned LLMs with batch inference using Ray + vLLM. - 🚀 Serve our LLMs as a production application that can autoscale, swap between LoRA adapters, optimize for latency/throughput, etc. Key Anyscale infra capabilities that keeps these workloads efficient and cost-effective: - ✨ Automatically provision worker nodes (ex. GPUs) based on our workload's needs. They'll spin up, run the workload and then scale back to zero (only pay for compute when needed). - 🔋 Execute workloads (ex. fine-tuning) with commodity hardware (A10s) instead of waiting for inaccessible resources (H100s) with data/model parallelism. - 🔙 Configure spot instance to on-demand fallback (or vice-versa) for cost savings. - 🔄 Swap between multiple LoRA adapters using one base model (optimized with multiplexing). - ⚡️ Autoscale to meet demand and scale back to zero. 🆓 You can run this guide entirely for free on Anyscale (no credit card needed). Instructions in the links below 👇 🔗 Links: - Blog post: https://lnkd.in/gvPQGzjh - GitHub repo: https://lnkd.in/gxzzuFAE - Notebook: https://lnkd.in/gmMxb36y
To view or add a comment, sign in
-
-
This autoscaling functionality that allows you to switch between A10s and H100s as well as on-demand to spot instances is fantastic. Solves so many headaches of LLM training deployment. Way too many training jobs I've run on other platforms hang forever waiting for resources or fail because they don't have these capabilities.
Excited to share our end-to-end LLM workflows guide that we’ve used to help our industry customers fine-tune and serve OSS LLMs that outperform closed-source models in quality, performance and cost. Key LLM workloads with docs.ray.io and Anyscale: - 🔢 Preprocess our dataset (filter, schema, etc.) with batch data processing. - 🛠️ Fine-tune our LLMs (ex. Meta Llama 3) with full control (LoRA/full param, compute, loss, etc.) and optimizations (parallelism, mixed precision, flash attn, etc.) with distributed training. - ⚖️ Evaluate our fine-tuned LLMs with batch inference using Ray + vLLM. - 🚀 Serve our LLMs as a production application that can autoscale, swap between LoRA adapters, optimize for latency/throughput, etc. Key Anyscale infra capabilities that keeps these workloads efficient and cost-effective: - ✨ Automatically provision worker nodes (ex. GPUs) based on our workload's needs. They'll spin up, run the workload and then scale back to zero (only pay for compute when needed). - 🔋 Execute workloads (ex. fine-tuning) with commodity hardware (A10s) instead of waiting for inaccessible resources (H100s) with data/model parallelism. - 🔙 Configure spot instance to on-demand fallback (or vice-versa) for cost savings. - 🔄 Swap between multiple LoRA adapters using one base model (optimized with multiplexing). - ⚡️ Autoscale to meet demand and scale back to zero. 🆓 You can run this guide entirely for free on Anyscale (no credit card needed). Instructions in the links below 👇 🔗 Links: - Blog post: https://lnkd.in/gvPQGzjh - GitHub repo: https://lnkd.in/gxzzuFAE - Notebook: https://lnkd.in/gmMxb36y
To view or add a comment, sign in
-
-
Here's an in-depth look for those interested at how to develop and productionize OSS LLMs that can outperform proprietary models at scale.
Excited to share our end-to-end LLM workflows guide that we’ve used to help our industry customers fine-tune and serve OSS LLMs that outperform closed-source models in quality, performance and cost. Key LLM workloads with docs.ray.io and Anyscale: - 🔢 Preprocess our dataset (filter, schema, etc.) with batch data processing. - 🛠️ Fine-tune our LLMs (ex. Meta Llama 3) with full control (LoRA/full param, compute, loss, etc.) and optimizations (parallelism, mixed precision, flash attn, etc.) with distributed training. - ⚖️ Evaluate our fine-tuned LLMs with batch inference using Ray + vLLM. - 🚀 Serve our LLMs as a production application that can autoscale, swap between LoRA adapters, optimize for latency/throughput, etc. Key Anyscale infra capabilities that keeps these workloads efficient and cost-effective: - ✨ Automatically provision worker nodes (ex. GPUs) based on our workload's needs. They'll spin up, run the workload and then scale back to zero (only pay for compute when needed). - 🔋 Execute workloads (ex. fine-tuning) with commodity hardware (A10s) instead of waiting for inaccessible resources (H100s) with data/model parallelism. - 🔙 Configure spot instance to on-demand fallback (or vice-versa) for cost savings. - 🔄 Swap between multiple LoRA adapters using one base model (optimized with multiplexing). - ⚡️ Autoscale to meet demand and scale back to zero. 🆓 You can run this guide entirely for free on Anyscale (no credit card needed). Instructions in the links below 👇 🔗 Links: - Blog post: https://lnkd.in/gvPQGzjh - GitHub repo: https://lnkd.in/gxzzuFAE - Notebook: https://lnkd.in/gmMxb36y
To view or add a comment, sign in
-
-
Check out our Analysis on LLMs Speed Benchmarked on various inference libraries 🚀
🚀 Just dropped our latest blog: "Exploring LLMs Speed Benchmarks: Independent Analysis" 📊 Dive into our independent analysis comparing the tokens/second of three different LLMs Highlights: - Models Tested: Gemma 7B, Llama-2 7B, and Mistral 7B - Libraries Tested: Text Generation Inference, vLLM, DeepSpeed Mii, CTranslate2, Triton with vLLM Backend, and TensorRT-LLM. - Input Tokens Length Variation : 20 to 5,000 tokens - Output Tokens Length Variation: 100, 200, and 500 tokens - Machine used : A100 on Azure For a comprehensive table of results and key findings, check our full analysis:https://lnkd.in/d9Q7Zq4N Hat Tip to Jack Weissenberger for the suggestion. Curious about other models? Let us know. Below is a graph showcasing the comparative performance.
To view or add a comment, sign in
-
-
AI Product @Amazon | Startup Investor | 2X Linkedin Top Voice for AI, Data Science, Tech, and Innovation | Quantum Computing & Web 3.0 | I use technology to scale and improve compliance processes | Content Creator
Approximate Nearest Neighbor is a powerful algorithm that revolutionizes search efficiency for your large datasets! With vector embeddings being all the craze for LLM use cases, you need an algorithm that will make little to no tradeoffs between speed and accuracy. And ANN is the candidate for this. On 2/5 at 10am PST, Akmal Chaudhri, Esq. will showcase how ANN scales for real-world applications, using AWS services, and you’ll want to join this session! Register for free here: https://lnkd.in/gmvzz8B8 In addition to some code sharing, you should expect to learn: 🔹Key principles of Approximate Nearest Neighbors in vector search. 🔹Techniques to scale vector search for large datasets. 🔹How to integrate ANN with AWS for enhanced performance. 🔹Strategies to balance search accuracy and efficiency. 🔹Insights into real-world applications of ANN in various industries. This webinar is for anyone looking to deploy modern data strategies for hot topic use cases such as LLM applications, and I’m happy to help Matt Brown and Yukthi Dwivedi bring these freebies to the community. See you there. And of course in below image you can notice how DALL-E still struggles to accurately convert text prompts into “text images.” #genai #search #technology #artificialintelligence Click #linkedangle and follow for more content Set the 🔔 notification on my page, don’t miss a post!
To view or add a comment, sign in
-
-
Building RAG-based LLM Applications for Production. An Exceptional Article Written by Goku Mohandas and Philipp Moritz. 📃In this guide, grasp how to construct an RAG-based LLM application, distribute workloads efficiently across compute resources, optimize various performance metrics, seamlessly navigate between OSS and closed LLMs, ensure top-notch scalability and availability, and delve into the impact of techniques such as fine-tuning and prompt engineering on performance. ↗️By the end, you'll have hands-on experience in creating an LLM application, ready to be served using FastAPI and Ray Serve. 🔗Link: [https://lnkd.in/e9yjcz7W] This piece is a must-read for anyone aiming to master the art of developing top-notch LLM applications.👁🗨 --------------------------------------------------------------- Follow me for daily machine learning and LLM content.💡 #machinelearning #llm #largelanguagemodel #deeplearning Anyscale
To view or add a comment, sign in
-
-
The open secret with LLMs right now: no one really knows how to integrate them with traditional computing. For example: - We have devs begging LLMs to generate correct JSON for their APIs - We have RAG - the only real way to give LLMs long-term storage - basically yeeting data-as-text into LLM prompts - We have non-deterministic algorithms breaking unit tests everywhere - We have a ton of latency mismatches, impedance mismatches, and unconstrained outputs Every platform shift builds on the previous platform. And innovators are needed to create that bridge. Luckily, our CTO Paul Blankley knows more about productionizing LLMs than pretty much anybody! Including me. So I'll just share an amazing presentation he gave on connecting LLMs to structured data. This is awesome Paul!
To view or add a comment, sign in
Thanks for sharing! That's interesting and promising. We also propose a work regarding the scaling law for time series forecasting(https://arxiv.org/pdf/2405.15124) at the same time 😁 . We believe these two works are more complementary than substitutive. Moreover, they provide insights into scaling laws in the realm of time series models from different perspectives.