Published my first kernel to go the last mile to optimize LTX-2.3 from @Lightricks! torch.compile cuDNN attn already gave a 1.42x boost. W/ the custom kernel added, I got 1.52x on a GB10 🔥 This was my systematic exploration of a simple agentic kernel dev workflow. More 👇

When this kernel was patched in, it delivered the finale! Some visible and consistent gains _without_ any regression in the outputs.

It takes a special kind of commitment to provide this kind of direction to the readers. Part II of the profiling posts is up, diligently brought to you by @ariG23498. Check it out at hf [dot] co / blog / torch-mlp-fusion 🔥

I have to clear some stuff up. When I was contacted to be included in this article, "Agentic AI" didn't sound like the sole focus of it. This is what I was told: > We’re currently working on a listicle featuring some of the top developers and AI builders from India who are making an impact through open-source, AI, infrastructure, and agentic AI projects. I want to clarify that I am definitely not an Agentic Software Developer nor did I pioneer anything that's happening in the field. I thought they were reaching out to folks who had experience in the other stuff too (e.g., open-source) and not JUST agentic stuff. So, I respectfully request @Analyticsindiam to take my name down from the article.

I like how `kernels` is getting into projects over time in an organic manner. If you work with custom kernels (be it the packaging or be it the consumption), consider giving the project a try 🤗 And let us know your feedback on the repo!