Yes I know. We helped show the PyTorch team how to train Llama efficiently – they have a whole GPU Mode talk about it. These issues do not at all line with your claims however.
@jeremyphoward
-
Understanding torch.compile: Common misconceptions explained
By
–
Whoever wrote that seems to not understand fundamentally what torch.compile does, or why it does it.
-
DDP Gradient Bucketing and Activation Checkpointing Optimization
By
–
I'm sorry, this makes no sense.
– Gradient bucketing is literally part of DDP
– DDP is a networking thing, obs you don't compile that
– Activation checkpointing works fine once hooks are disabled during tracing (ofc) -
Understanding the Inevitable Problem with AI and Technology
By
–
Everyone sees the problem eventually. They often slam into it at great speed, in fact…
-
Claude Uses ULMFiT Tokenization Technique from 2017
By
–
According to this research it sounds like Claude uses the same tokenization trick I invented for ULMFiT in 2017! 😀
-
Paywall llms.txt feature enhances premium customer experience
By
–
This is kinda cool actually — an llms.txt behind a paywall; very nice extra touch to provide paying customers! 😀
-
Managing Multiple CUDA Toolchains: uv Limitations vs Conda
By
–
E.g having multiple CUDA toolchains and switching between them to test whether your lib works right on each, isn't possible with uv. But with conda it's easy.
-
UV Package Manager: Project-Based Environments and Development Tools
By
–
I use uv. I don't quite love it yet, because if you're not doing project-based envs (which I hate) then it's hard to get it to work right. Comparing it to conda is a bit silly though since they do different things.
-
Bootstrap Framework Remains Underrated and Highly Effective
By
–
Counterpoint – bootstrap nowadays is actually pretty great and criminally under-rated 😀
-
UIKit Benefits and Trade-offs in Development
By
–
UIKit has some benefits in terms of being able to leverage it too, but it's another thing to deal with…