- Видео 37
- Просмотров 26 012
Triton
Добавлен 25 сен 2023
Panel: Hardware Heterogeneity at Meta
Ajit Mathews moderates a panel on how Meta is making Hardware Heterogeneity a reality for AI.
The panelists are Jiaqi Zhai, Roman Levenstein, Jongsoo Park, Xiaodong Wang - they represent ML Engineering, ML Software, ML Hardware, and AI Infra.
Questions address four broad themes: 1) Why Triton, and what are the counter metrics? 2) Triton and its relationship to hardware microarchitecture. 3) Million scale GPU clusters. 4) Looking forward.
The panelists are Jiaqi Zhai, Roman Levenstein, Jongsoo Park, Xiaodong Wang - they represent ML Engineering, ML Software, ML Hardware, and AI Infra.
Questions address four broad themes: 1) Why Triton, and what are the counter metrics? 2) Triton and its relationship to hardware microarchitecture. 3) Million scale GPU clusters. 4) Looking forward.
Просмотров: 331
Видео
Dev Tools: Proton/Interpreter
Просмотров 406Месяц назад
Keren talks to tooling that can help with writing Triton kernels - specifically the Triton interpreter, which is very helpful for debugging things like Illegal Memory Accesses; and Proton which is critical for understanding Triton kernel performance. Slides: bit.ly/proton-interpreter-tutorial
Exo: Let's Build Power Tools for Performance Engineers
Просмотров 180Месяц назад
Gilbert Bernstein shares his work on EXO for Productive Programming of Hardware Accelerators. EXO is a compiler/language design that externalizes parts of the compiler in order to give programmers more control. This is an exciting new collaboration between UW and MIT. Slides: bit.ly/exo-complement
Mosaic GPU: A DSL for Fast Hopper Kernels in Python
Просмотров 398Месяц назад
Adam Paszke talks to the Mosaic GPU kernel authoring project from Google and how it exposes Hopper uarchitectural capabilities. Slides: bit.ly/mosaic-complement
Accelerating the Future: Triton on Blackwell Architecture
Просмотров 383Месяц назад
Pradeep Ramani (NVIDIA) and Thomas Raoux (OpenAI) talk to enhancements to Triton to leverage uarchitectural innovations in Blackwell GPUs. Slides: bit.ly/blackwell-backend
Qualcomm Hexagon NPU Backend for Triton
Просмотров 245Месяц назад
Muthu Baskaran describes the Qualcomm Hexagon NPU for Edge- and Cloud-AI, and talks to Qualcomm's use of Triton for programming it. Slides: bit.ly/qualcomm-backend
Rapid Innovation on AWS Trainium with Triton-Inspired Programming
Просмотров 231Месяц назад
Jonathan Henson describes the AWS Trainium and Inferentia2 architectures, and how Amazon was inspired by Triton to build the Neuron Kernel Interface framework for programming it. Slides: bit.ly/tranium-backend
Triton for Azure MAIA
Просмотров 177Месяц назад
Ian Baird describes Microsoft’s custom AI ASIC, the MAIA-100, and how it leverages Triton for convenient programming. Slides: bit.ly/maia-backend
Pipelining Persistent Kernels
Просмотров 433Месяц назад
Pawel describes how Triton supports pipelining in the context of persistent kernels. (This talk was voted the audience's favorite in an informal poll!) Slides: bit.ly/pipelining-backend
Triton CPU: Its Approach and Early Performance Studies on x86/ARM
Просмотров 474Месяц назад
Minjang Kim (Meta) and Ilya Enkovich (Intel) talk to an exciting new initiative - Triton CPU. Slides: bit.ly/cpu-backend
Triton on AMD GPUs
Просмотров 410Месяц назад
Lei Zhang and Lixun Zhang talk to Triton support for AMD. This talk shows off some very clever optimization techniques around chiplets and also instruction scheduling around L1 and global memory. Slides: bit.ly/amd-backend
Enable Triton on Intel GPUs
Просмотров 345Месяц назад
Ettore Tiotto talks to the state of Triton's support for the Intel GPU. A highlight of this presentation is the custom backend they built, which can serve as a blueprint for similar work. Slides: bit.ly/intel-backend
Keynote: Philippe Tillet (OpenAI)
Просмотров 807Месяц назад
Phil talks to the state-of-the-Triton-Union and his vision for the coming year. Slides: bit.ly/phil-keynote
Keynote: Aparna Ramani (Meta)
Просмотров 354Месяц назад
Aparna makes the business case for Triton at Meta. She leads the Data, Developer, and AI Infrastructure teams at Meta. Her teams are responsible for some of the largest data systems on the planet, Meta's well-known developer environments, as well as AI systems and frameworks - including PyTorch. Slides: bit.ly/aparna-keynote
Compiler Tools: Writing an MLIR Pass
Просмотров 563Месяц назад
Haishan introduces MLIR, and describes how Triton uses MLIR to leverage custom intrinsics. Slides: bit.ly/mlir-tutorial
Triton Conference 2024: Morning Session
Просмотров 1,1 тыс.Месяц назад
Triton Conference 2024: Morning Session
Triton Conference 2024: Afternoon Session
Просмотров 2,4 тыс.Месяц назад
Triton Conference 2024: Afternoon Session
August Triton community meetup 20240806
Просмотров 3943 месяца назад
August Triton community meetup 20240806
Triton May Community meetup 20240507
Просмотров 4426 месяцев назад
Triton May Community meetup 20240507
Triton March Community Meetup 20240402
Просмотров 3347 месяцев назад
Triton March Community Meetup 20240402
Triton Feb community meetup 20240220
Просмотров 4928 месяцев назад
Triton Feb community meetup 20240220
January Triton community meetup 20240124
Просмотров 5199 месяцев назад
January Triton community meetup 20240124
Triton December community meetup 20231213
Просмотров 34611 месяцев назад
Triton December community meetup 20231213
Triton October Community meetup 20231025
Просмотров 279Год назад
Triton October Community meetup 20231025
Triton community meeting August 2023 08 22
Просмотров 194Год назад
Triton community meeting August 2023 08 22
Triton Monthly Community Meetup 2023 07 18
Просмотров 187Год назад
Triton Monthly Community Meetup 2023 07 18
Writing Grouped GEMMs in Triton Nvidia
Просмотров 1,1 тыс.Год назад
Writing Grouped GEMMs in Triton Nvidia
At 5:04, there is a slides for more comprehensive examples, is it anywhere public?
Please refer the slides link here: github.com/triton-lang/triton/blob/main/docs/meetups/dev_conference_2024.md and bit.ly/mlir-tutorial
Hi, I'm working a Triton project for a client. I was wondering if there was a way for me to join in on this community? I've been toying with Triton since I was an undergrad :)
Please join the GPU mode discord server and ask Triton related questions in the #triton channel there. discord.gg/QutVdaFY
Thanks! It is still quite a lot for a relative newcomer to the area, but having a presentation format available walking through this is greatly appreciated - anything that helps build the mental model.
Please refer to the slides here: github.com/triton-lang/triton/blob/main/docs/meetups/dev_conference_2024.md
Hi! The link to the slides in the video description seems to be incorrect. If you could update it, that would be really helpful.
drive.google.com/file/d/1YTGfJq-ccuVPH50r8ERnCrC977-4uzly/view
Please refer the slides link here: github.com/triton-lang/triton/blob/main/docs/meetups/dev_conference_2024.md
"Small, hackable Python core: " hell yeah, call that half a context window
Thank you for great information
Thank you for the information !!
Hi! I love Triton and I just used it on my undergraduate capstone. I want to highlight that sometimes the audio chops off in this video
this is awesome