61–70 of 346 posts
Earlier quoted context omitted.
> Is the claim that it only took $5M to train generally accepted? Based on Nvidia being down 18% yesterday I would say the claim is generally accepted.
> Nvidia being down 18% The only part of DeepSeek-R1 I do not like. I hope it's over, but I am not holding my breath.
Earlier quoted context omitted.
Can't this kind of repetition be dealt with at the ~~decoder~~ (edit: sampler) level, like for any models? (see DRY ~~decoder~~ sampler for instance: https://github.com/oobabooga/text-generation-webui/pull/5677 )
You likely mean sampler, not decoder. And no, the stronger the quantization, the more the output token probabilities diverge from the non-quantized model. With a sampler you can't recover any meaningful accuracy. If you force the sampler to select tokens that won't repeat, you're just trading repetitive gibberish for non-repetitive gibberish.
Indeed, that's posting before being fully awake.
> And no, the stronger the quantization, the more the output token probabilities diverge from the non-quantized model. With a sampler you can't recover any meaningful accuracy.
OF course you can't recover any accuracy, but LLM are in fact prone to this kind of repetition no matter what, this is a known failure mode that's why samplers aimed at avoiding this have been designed over the past few years.
> If you force the sampler to select tokens that won't repeat, you're just trading repetitive gibberish for non-repetitive gibberish.
But it won't necessary be gibberish! even a highly quantized R1 has still much more embedded information than a 14 or even 32B model, so I don't see why it should output more gibberish than smaller models.
An 80% size reduction is no joke, and the fact that the 1.58-bit version runs on dual H100s at 140 tokens/s is kind of mind-blowing. That said, I’m still skeptical about how practical this really is for most people. Like, yeah, you can run it on 24GB VRAM or even with just 20GB RAM, but "slow" is an understatement—those speeds would make even the most patient person throw their hands up. And then there’s the whole re…
Not everyone needs the largest model. There are variations or R1 with fewer parameters that can easily run on consumer hardware. With 80% size reduction you could run 70B on 8-bit on an RTX 3090. Other than that, if you really need the big one you can get six 3090s and you're good to go. It's not cheap, but you're running a ChatGPT equivalent model from your basement. A year ago this was a wetdream for most enthusias…
This line in the stuff inside the section suggests it's also been trained on YouTube clips:
>> "I'm not entirely sure if I got all the details right, but this is what I remember from watching clips and summaries online."
An excerpt from the generated summary:
>> "Set in the 23rd century during a Z-Corp invasion, the series features action sequences, strategic thinking, and humor. It explores themes of international espionage, space warfare, and humanity's role in the cosmos. The show incorporates musical numbers and catchy theme songs for an engaging viewing experience. The plot involves investigating alien warships and their secret base on Kessari planet while addressing personal conflicts and philosophical questions about space."
"It explores themes of international espionage, space warfare, and humanity's role in the cosmos" is the closest to correct line in the whole output.
Earlier quoted context omitted.
Not everyone needs the largest model. There are variations or R1 with fewer parameters that can easily run on consumer hardware. With 80% size reduction you could run 70B on 8-bit on an RTX 3090. Other than that, if you really need the big one you can get six 3090s and you're good to go. It's not cheap, but you're running a ChatGPT equivalent model from your basement. A year ago this was a wetdream for most enthusias…
I ran whatever version Ollama downloaded on a 3070ti (laptop version). It's reasonably fast. Generative stuff can get weird if you do prompts like "in the style of" or "a new episode of" because it doesn't seem to have much pop culture in its training data. It knows the Stargate movie, for example, and seems to have the IMDB info for the series, but goes absolutely ham trying to summarize the series. This line in the…
Probably was not r1, but one of the other models that got trained on r1, which apparently might still be quite good.
Random observation 1: I was running DeepSeek yesterday on my Linux with a RTX 4090 and I noticed that the models should fit into VRAM, which is 24GB. Or they are simply slow. So the Apple shared memory architecture has an advantage here. A 192GB Mx Ultra can load and process large models efficiently. Random observation 2: It's time to cancel the OpenAI subscription.
Earlier quoted context omitted.
My understanding is with MoE (Mixture of Experts), you can and should shard it horizontally. The whole model is 600GB, but only 37GB is active during the evaluation of any single output token. So you can load a different active subset of the MoE into each 89GB GPU, sharding it across something like 32 different GPUs (or can you get away with less? Wouldn't be surprised if they can infer on 8x H800 gpus). Some paramet…
Each MoE layer has its own router, and it activates 8 (out of 256) experts at a time. There's no reason to expect all of them to stay on the same GPU, so you're pretty much guaranteed to have to do all-to-all communication between the GPUs in your cluster after every layer for every token.
I had assumed the performance advantage for MoE came from minimising traffic between GPUs. But if it's per layer routing, then it's going to massively increase inter-gpu traffic compared to vertical slicing.
I guess that means the performance advantage actually comes when batching thousands of queries? The MoE routing would mean that on each MoE layer, each GPU shard gets a batch of queries that will all hit roughly the same subset of experts (and read the same weights from memory). The batches then shuffle between each MoE layer to re-optimise.
It's kind of like GPU raytracing where you get large performance gains by running coherency sorting on rays and batching similar rays together.
An 80% size reduction is no joke, and the fact that the 1.58-bit version runs on dual H100s at 140 tokens/s is kind of mind-blowing. That said, I’m still skeptical about how practical this really is for most people. Like, yeah, you can run it on 24GB VRAM or even with just 20GB RAM, but "slow" is an understatement—those speeds would make even the most patient person throw their hands up. And then there’s the whole re…
Earlier quoted context omitted.
While 192GB of ram is appealing, it's also quite expensive at $6000. For that price I rather buy a system with 5 used 3090s, which while being "only" 120GB of VRAM, you benefit from much faster tokens/s and prompt processing speed (the macs are notoriously slow at consuming large contexts).
Can I use that on the train though? I can with a 128GB MacBook, without it sounding like a helicopter taking off as well.
An 80% size reduction is no joke, and the fact that the 1.58-bit version runs on dual H100s at 140 tokens/s is kind of mind-blowing. That said, I’m still skeptical about how practical this really is for most people. Like, yeah, you can run it on 24GB VRAM or even with just 20GB RAM, but "slow" is an understatement—those speeds would make even the most patient person throw their hands up. And then there’s the whole re…
Oh the repetition issue is only on the non dynamic quants :) If you do dynamic quantization and use the 1.58bit dynamic quantized model the repetition issue fully disappears! Min_p = 0.05 was a way I found to counteract the 1.58bit model generating singular incorrect tokens which happen around 1 token per 8000!
Random observation 1: I was running DeepSeek yesterday on my Linux with a RTX 4090 and I noticed that the models should fit into VRAM, which is 24GB. Or they are simply slow. So the Apple shared memory architecture has an advantage here. A 192GB Mx Ultra can load and process large models efficiently. Random observation 2: It's time to cancel the OpenAI subscription.
So I'm thinking, inference seems mostly memory bound. With a fast CPU (for example 7950x with 16 cores), and 256GB of RAM (seems to be the max), shouldn't that give you plenty of ability to run the largest models (albeit a bit slowly). It seems that AMD Epyc CPUs support terabytes of ram, some are as cheap as 1000 EUR. why not just run the full R1 model on that - seems that it would be much cheaper than multiple of t…