DisTrO trains multiple models in a bounded search space instead of full synchronization
Jeff Schmidt · The Quest for Community-Trained Open Source AI Models · Oct 1, 2024 · at 1:01:22
Jeff Schmidt (Jeffrey Quesnelle) is co-founder of Nous Research. He is explaining how DisTrO breaks traditional AllReduce AI training paradigms.
0:00 / 0:59exact quote · 59.7s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“So with distro, what we found is that rather than bringing everyone back home and averaging it back together, what you want to do is give each of those little nodes that are searching for the lowest point in the lost landscape, the freedom to move around. And they aren't actually coming home and all synchronizing. They each have the freedom to move around. But what you don't want is the freedom to move around and just go off on a tangent. So we sort of have like, he mentioned this rope. It's literally like this. They're all connected by like this bungee cord. And if one of them starts to fall down, like a really good lost landscape, they'll start pulling the other ones. But that diversity of search space, Where we actually aren't training one, we're not, we're breaking the paradigm of there being one model that's being trained. There's actually n number of models being trained, each of them getting to do their own little exploration, but within a bounded space, so that they're kind of like all looking around, and instead of like all coming home, they all phone home. And they just kind of say, here's what I've learned that's the best insights from what I've learned, versus being, here's, let's merge all together back into one.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Jeff Schmidt
Schmidt: AI industry relies on outdated 1990s architecture assumptions ripe for disruption
“It turns out that most assumptions in the AI space right now are a product of that's just how things had been done when there was not nearly as much energy and attention to it. So someone made an assumption maybe in the early nineties that everyone just kind o…”
Schmidt: Relying solely on fine-tuning is an existential threat to open-source AI
“Because that for us, if we're just fine tuning models. Right. That's like an existential threat, right? That's like an actual existential threat because the closed providers will continue to get better and we would be like dead in the water in a lot of sense.”
Schmidt: Sharing only key signals in distributed training yields equivalent model learning
“We know that like what we, what needs to be communicated between these things, the two, the different nodes are just these few key pieces of information. And that is necessary. That is a necessary condition or rather a sufficient condition To get the equivalen…”
Schmidt: Decentralized training will force NVIDIA to redesign chips around VRAM ratios
“What might happen sooner would be a redesign of the types of chips that NVIDIA or someone would make. Okay, under this model, we can dedicate more VRAM versus, there's like this question of how much VRAM versus how much processing power is on a die, and that, …”
Schmidt: Compute chips in NVIDIA's RTX 4090 and H100 are almost identical
“I think people don't actually realize that like a forty-ninety and like an H 100 are in a lot of ways the same card. For the non-gamers in the room, explain the forty-ninety. The chip that's inside of them is almost identical. The chip, the actual compute chip…”
Schmidt: Consumer gaming GPUs will become the sweet spot for distributed training
“Because you're able to distribute it so wide, I think the gaming GPU angle is really going to be like the sweet spot. If, as long as there's continued to be sort of like higher end gaming GPUs, and those are on comparison with the high end training GPUs, even …”