Jordan Conragan

Okay I was really annoyed by the claude-isms. I'm tired of seeing "nothing else. Simple now, load-bearing later.", "The honest edges", and even this seems like Claude wrote it: "The fix is one rule: ..." . It really detracts from the contents. Even the article itself seemed too long for the contents (encoding and decoding aren't necessarily inverse operations, renderers is a good fix for it). I kinda expect AI posts like this from community articles but not when those articles are from HF people, it was dissapointing

How did you distribute the weights in your Qwen 72B experiment? Did you just have it running w/ TP=8 on a single node, or did each node have it's own copy of Qwen 72B?

I noticed that having vLLM sleep didn't make it in to TRL, what's the reason for that?