Play Krillion, Wordle the LinkedIn Puzzles and other daily games? Use Bunch to share your scores with all your friends!
We've been building a central solution for sharing your game scores with your family, friends and coworkers check it out at https://t.co/5L0MyRCaZf
Over the past few weeks, I've been working on optimizing neural text-to-speech inference using adaptive compute and caching with dots.tts at @uwwatai.
Last week, our team, Adaptive-Compute Inference for Real-Time TTS, led by @saikodakalla19, gave a talk @UWaterloo about some of our findings.
One of our biggest findings: 98.5% of traced repeated-operation time was spent inside the audio refinement model.
Here's what I worked on and what we discovered! ↓
Our research identified two promising directions:
Adaptive NFE, reducing unnecessary DiT evaluations by allocating compute based on individual audio patches.
Extending existing caching mechanisms while maintaining correctness across changing refinement schedules.
My biggest takeaway was that inference optimization starts with understanding what the model actually computes, what changes between evaluations and what can safely be reused.
Research repo:
https://t.co/xh9LxlYQ52
Over the past few weeks, I've been working on optimizing neural text-to-speech inference using adaptive compute and caching with dots.tts at @uwwatai.
Last week, our team, Adaptive-Compute Inference for Real-Time TTS, led by @saikodakalla19, gave a talk @UWaterloo about some of our findings.
One of our biggest findings: 98.5% of traced repeated-operation time was spent inside the audio refinement model.
Here's what I worked on and what we discovered! ↓
I also investigated dots.tts's existing KV-cache implementation, comparing EagerDiTRunner with CachedDiTRunner.
I found that the optimized path already caches historical attention K/V states, but timestep-conditioned AdaLN introduces an important constraint: cached representations aren't necessarily valid when the NFE schedule changes.
In our T4 benchmark at 10 NFE, optimized inference achieved:
• Eager: 8.06s
• Optimized: 2.97s
• 2.72× faster end-to-end
This measures the complete optimized implementation, not KV caching alone.
Hello everyone! WATai is hosting our first Builders Night next Wednesday.
If you're working on a project, have something that interests you, or just have some ideas you'd like to share, don't hesitate to join us! Looking forward to seeing everyone.
Link: https://t.co/8I0XfKtJXY
Inspired by the power of @Shopify, we built an Agentic AI pipeline to help you repair your items as fast as possible without installing anything on IOS using App Clips.
We just won $5000 at Hack Canada for this, check it out here: https://t.co/OBqD9guYoI
🤵 we built @cursor_ai for fashion
> made it in 24hrs @ nexhacks (cmu)
> got noticed by vcs and businesss
> now we’re building it for production as an ai stylist
@4derekwang@YTExo1k @dishes_io