> For extreme performance, we discover and use an out-of-doc PTX instruction: ld.global.nc.L1::no_allocate.L2::256B. This instruction will lead to an undefined behavior: accessing volatile GPU memory with non-coherent read-only PTX modifiers .nc. But the correctness is tested to be guaranteed with .L1::no_allocate on Hopper architectures, and performance will be much better.

Practically speaking, is it possible for NVIDIA to "pull the rug" later, intentionally or otherwise, by subtly changing the behaviour of this out-of-doc instruction on new architectures?

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.

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.

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.

I understand the reasoning and I hope there is legislation in the future that basically goes "If you can't produce the data, you can't charge more than this for it". Basically, LLM producers will have to treat their product as a commodity product that can only be priced based on the compute resources plus some overhead.

Come on... Meta has been refining pytorch for more than a decade. It basically contains all that you need to train LLMs, including the latest technologies. What more do you need? The part of the code that is specific to Meta infrastructure?

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.

There's hilariously nothing open about OpenAI, and that was the plan from the start. From the email by Ilya Sutsekver, OpenAI was always going to keep all it's research and code as proprietary information. Open supposedly meant the benefits would be shared. So they basically just became a SaaS with a free tier, like most of them. Musk was right when he called them out for fishing for money as if they were a non profit, but always had plans to become a company

Now. It’s amazing to me that everyone is like fuck OpenAI deepseek is the savior, when OpenAI’s papers and code jump started an AI revolution just a few years ago. Let’s wait the same number of years and see what deepseek does.

Too late, there is no more innovation from openai all the people that were the drivers left for Anthropic and the others. They had some of the biggest funding, had the advance... And yet they lost it.

That’s an impossible ask. Sam is the pinnacle of capitalist ruling class, he’s a pure businessman. He has no interest in giving anything for free unless there’s a business plan. He doesn’t care about humanity. He’ll pretend to change the world and tell you that they’re inventing AGI, Q*, strawberry or whatever they’re branding it, but the reality is he knows it’s all over and unless there’s a major breakthrough this company will be in major financial trouble. Sorry for the rant but he doesn’t deserve much respect for turning all this science to grift. He’s actually the person the old openai board warned everyone about.

Much of the hype around DeepSeek is due to their extraordinarily low training and inference costs. They achieved this by optimizing their training code, apparently using PTX in addition to CUDA. PTX is kind of an intermediate assembly language for NVIDIA GPUs and people are eager to see how it was used.

"For extreme performance, we discover and use a behavior-out-of-doc PTX instruction: ld.global.nc.L1::no_allocate.L2::256B. This instruction will lead to an undefined behavior: accessing volatile GPU memory with non-coherent read-only PTX modifiers .nc. But the correctness is tested to be guaranteed with .L1::no_allocate on Hopper architectures, and performance will be much better. If you find kernels not working on some other platforms, you may add DISABLE_AGGRESSIVE_PTX_INSTRS=1 to setup.py and disable this, or file an issue."

My understanding of the L2 part is that it asks for a 256b pre-fetch (only available on some platforms it seems) but they use vectors of 4 32bits signed ints max so not sure why only the 256 would work or if the fact that it did fetch the next 128 helps.

Now it includes the highly anticipated PTX! Of course, I don’t understand it, but I’ve already click the star and even the fork button, which basically means I’ve mastered it, right? I feel incredibly powerful right now...