Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepSeek open source DeepEP – library for MoE training and Inference

github.com

Post reply on HN
Print view View original

71–77 of 77 posts

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#71

Post by jimmydoe »

danans wrote: post #62
ur-whale wrote: post #50

The incentive behind the work of DeepSeek might very well be wrong (something along the lines of a state-sponsored attempt at shrinking the US first mover advantage in AI to nil) but the net result for everyone on the planet is simply fantastic. So even in the worst case (doing this for the wrong reasons): thank you DeepSeek, you are actually doing what OpenAI lied through their teeth to the whole world about doing f…

> The incentive behind the work of DeepSeek might very well be wrong (something along the lines of a state-sponsored attempt at shrinking the US first mover advantage in AI to nil) In the space of international relations, right and wrong don't apply nearly as much. Is open sourcing this any more "wrong" than the export ban on high end Nvidia GPUs? The open sourcing by DeepSeek (presumably with CCP consent) just happe…

It seems CCP is less hate worthy than they were two months ago. Comparing fake democracy with real authoritarian is kinda funny.

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#72

Post by chvid »

lithiumii wrote: post #52
breadwinner wrote: post #51

Zuckerberg should stop claiming Meta is open sourcing AI (they are even running TV ads) when they are only releasing the weights, and not the code. Only DeepSeek is real OSS AI.

Well technically even DeepSeek is not as OSS as OLMo or Open Euro, because they didn't open the data.

It is pirated material / material that breaks various terms of service but as I understand it is the stuff you can see in Anna's Archive and a bunch of "artificial" training data from queries to OpenAI ChatGPT and other LLMs.

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#73

Post by anticensor »

echelon wrote: post #54
anticensor wrote: post #42

Earlier quoted context omitted.

Their state-of-the-art speech to text model, Whisper, is available as open weights for free.

Strategically, they know that needs to run at the edge, and they want users to send them requests to their API without incurring latency or bad user experience. That is still a fair point, though, and it should be commended. And that hasn't been their only contribution, either.

They could've made it a trusted-computing-only model distributed with a proprietary encryption, unlocked with an expensive licence key if they wanted.

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#74

Post by kristianp »

antirez wrote: post #35
pama wrote: post #7

I feel like a kid in a candy shop. Some of these tricks would take way too long to reverse engineer correctly based on the papers. I hope that the releases this week start a renaissance of the use of MoE as baseline academic models.

From this point of view I don't understand what's happening between the actual SOTA models practice and the academic models. The former at this point are all MoEs, starting with GPT4. But then the open models, if not for DeepSeek V3 and Mixtral, are always dense models.

I assumed gpt 4o wasn't MOE, being a smaller version of gpt-4, but I've never heard either way.

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#75

Post by ein0p »

prjkt wrote: post #59
breadwinner wrote: post #51

Zuckerberg should stop claiming Meta is open sourcing AI (they are even running TV ads) when they are only releasing the weights, and not the code. Only DeepSeek is real OSS AI.

does pytorch count

PyTorch had the "first thing that didn't suck" advantage and now it has a completely dominant marketshare that prevents better alternatives from emerging. Where it sucks (e.g. on macOS) there are popular alternatives. But it's hard to be enthusiastic about a DL framework in 2025 which does not have native high performance quantization support, for example. Or one where FSDP is crudely bolted onto the side. They say "usability above all else", but I consider such things to be major usability deficiencies, which need to be addressed. But because PyTorch does not have to fight for marketshare, it'll be years before we see anything usable there.

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#76

Post by wbsun »

This feels like the 80s/90s when people hacking assembly or finding undocumented instructions to squeeze CPU for performance. Until one day either the compiler will be highly optimized enough or the GPU will be so powerful that such tricks won’t make much difference anymore, like CPUs nowadays :D

Re: DeepSeek open source DeepEP – library for MoE training and Inference

#77

Post by numba888 »

ein0p wrote: post #75
prjkt wrote: post #59

Earlier quoted context omitted.

does pytorch count

PyTorch had the "first thing that didn't suck" advantage and now it has a completely dominant marketshare that prevents better alternatives from emerging. Where it sucks (e.g. on macOS) there are popular alternatives. But it's hard to be enthusiastic about a DL framework in 2025 which does not have native high performance quantization support, for example. Or one where FSDP is crudely bolted onto the side. They say "…

I wonder how Meta trains its models. On vanilla Pytorch or they actually have some closed tools and frameworks?
Post reply on HN