I spent the morning reading all 244 pages. About 180 of them are getting zero coverage. Here are the findings that deserve more attention from anyone building with AI esp the psych evals and p-hacking.
I am unsure as to why this is still not a native EKS feature. I understand that there can be a cascading node failure but that can be easily detected and fixed.
Karpenter version that you mention is also pretty recent. I am just surprised as to why this is not a standard thing.
In this post, we’ll break down what NCCL does, why it’s critical for multi-GPU training, and how to tackle one of its common challenges – the dreaded “watchdog timeout” error.
Sounds interesting! I love all these edge experiments. But as long as there is architecture dependent code for models, I feel these edge experiments can't fully express their strong suit.
You try to run something and Voila you need Ampere or Hopper or Laplace for flash attnt.
GPU workloads cannot be tested on local machines and need access to specialised hardware and specific types of GPU machines and production deployments need Docker containerised files. This abstraction connects your machine via SSH to an ec2 instance and enables hot reloading Docker containers with GPU enabled. Makes it very easy to test out containerised GPU apps.
At Tensorfuse we have been rewriting and restructuring our docs as they have become a huge dump of how to guides rather than being something that is navigable and understandable. After a lot of peer discussion and reading, we are now trying the diataxis approach and it seems to be working with our beta users for now. Sharing it here as some people might be looking to make their documentation more navigable.
I’ve been working on Python-based backend development for about three years now in various forms. I primarily use Django and FastAPI, although I initially started with Flask. However, during my backend work, I frequently encountered the terms ASGI and WSGI. For example, one of my Django deployment scripts included references to asgi_app and wsgi_app, and used gunicorn to deploy these apps. Although I initially dismissed these terms as implementation details, I now find myself needing to support both ASGI and WSGI apps for my company tensorfuse. As a result, I believe it’s important to explain ASGI and WSGI to a wider audience.
In this article, I explore how to improve the retrieval subsystem of RAG pipelines. I discuss common issues that can occur at the retrieval stage and provide practical solutions. Topics covered include tracking relevant metrics, addressing database coverage and embedding model suitability, query rewriting techniques, and strategies for enhancing context relevance through chunking.
I am one of the persons who worked on Aerotime Rewind. Though I already anticipated what would my year would look like, I was still a little surprised with my wordcloud. There were topics I had forgotten I worked on and them being there in the wordcloud gave me a little nostalgic feeling.
I also tried comparing my meeting DNA with Github contribution graph on Linkedin and counterintuitively they had positive correlation. One of the conclusions I drew out of that comparison was the fact that constant back and forths with my team kept my motivations up and since these meetings were always stacked together (which is what Aerotime is designed to do), they didnot hamper my deep work. Pretty amazing to see my deep work graph too.