I took a random book excerpt with 23,000 words (±30k input tokens) and used it as context. Jev still responds in 800ms, sometimes 500ms. That's in the neighbourhood of 20-50,000 tok/s prefill, which is obviously not possible with normal LLMs, not even Cerebras is this fast.
I imagine it’s not so hard to optimize a model for this use case.
Off the top of my head, I would skip all the modern linear attention / state space stuff and use classical attention. But run prefill in a fully sliding-window mode so that “state” tokens simply don’t attend to far away tokens, or maybe also allow everything to attend to the first few tokens (and train like this). Now prefill is almost embarrassingly parallel, and you can make it fully parallel by duplicating work at block boundaries. (I’m not saying this is an awesome architecture if you want excellent results, but I’m also not convinced that Jev gives excellent results…)
The let queries attend to everything.
And architect the stack around this. Don’t try to cache the KV data — process the queries as you go so that the each input block and layer’s K and V data is computed, attended to, and discarded.
I’m curious whether Cerebras actually is a good device for this. Cerebras is kind of low on RAM, but if you don’t need to store KV data, maybe the entire computation fits on the die.
i have tested jev for my use cases and its horrendously wrong, but then the follow up from jev's team is "oh, you need to boil the question down further". it's a spiral of how much do you wanna dumb down the ask so that it answers it correctly. i'll pass for now.
The prefill is the only blocking part, and you can prefill the whole context up to the point where they diverge, then prefill each question and decode the one token in parallel for each question.
If you batch vLLM calls with the same prompt prefix to the same process, it'll deduplicate the prompt prefix across batched requests (+/- the block size) and decode in parallel for each.
That's with a vanilla LLM. If you modify the LLM you can pull that in-graph, but it isn't really necessary.
There's some speculation that Jev is an open weight model with novel post-training (RLCD). So, if these folks have competitive accuracy with just the base model, it may raise some questions about the necessity of Jev's architecture. You generally don't want to find yourself competing only on price.
> it may raise some questions about the necessity of Jev's architecture
When I hear “architecture” I am thinking number of parameters and latency.
When I hear “accuracy” I think training recipe, data, and (later) number of parameters.
So when you say that Jev’s architecture may not be necessary, the evidence I expect to see is comparable quality at comparable latency. Not equal quality at 2x latency and 4x the cost.
Also, while it's clearly got a lot of training on some use cases, others that probably weren't in the training set have worse good decision rates than a random number generator. If you can rebuild the architecture, you can train it on your use case.
I mean just from what's known of the funding and timeline it pretty much has to be based on open weights.
But it is likely more than just a fine tune + novel training. At the very least the LM head is swapped out for a classifier one and then or also idk, bidirectional attention for the encoding pass I'm out of my depth at this point and will stop guessing. The training is probably where they have the biggest moat though, not that it's necessarily huge.
I have a project that fits jev as advertised almost comically well and I've been playing with it, and the various hacks and open versions. Jev doesn't necessarily perform better overall but it is quite different. It's sensitive to prompt phrasing in ways the others aren't, it's easy to generate questions where all the other models cluster in confidence but jev is an outlier. Not necessarily more correct, but it does feel like it's getting its answers in a different way.
I'm guessing just as much as anyone else but I've been spending a ton of time on this the last couple weeks, it landed right when I was most ready to dig into it.
> We measured latency in separate runs with one request at a time, because timings taken under load measure the queue rather than the model.
> As Privatemode is hosted in the EU and Jev is hosted in the US, we ran four of the datasets from Germany and from the US at the same time. From Germany, Privatemode answered in 180 ms and Jev in 264 ms. From the US, the order reverses: 164 ms for Jev against 299 ms for Privatemode.
Not reading TFA before commenting is okayish I guess. Confidently doubling down with a direct contraction to a short and clear quote in a reply is just polluting the discussion with noise.
Is this a joke? “Jev-like” properties? People have been using LLMs as classifiers or rankers in a similar way for ages. I feel like we’re losing our minds
I took a random book excerpt with 23,000 words (±30k input tokens) and used it as context. Jev still responds in 800ms, sometimes 500ms. That's in the neighbourhood of 20-50,000 tok/s prefill, which is obviously not possible with normal LLMs, not even Cerebras is this fast.
Off the top of my head, I would skip all the modern linear attention / state space stuff and use classical attention. But run prefill in a fully sliding-window mode so that “state” tokens simply don’t attend to far away tokens, or maybe also allow everything to attend to the first few tokens (and train like this). Now prefill is almost embarrassingly parallel, and you can make it fully parallel by duplicating work at block boundaries. (I’m not saying this is an awesome architecture if you want excellent results, but I’m also not convinced that Jev gives excellent results…)
The let queries attend to everything.
And architect the stack around this. Don’t try to cache the KV data — process the queries as you go so that the each input block and layer’s K and V data is computed, attended to, and discarded.
I’m curious whether Cerebras actually is a good device for this. Cerebras is kind of low on RAM, but if you don’t need to store KV data, maybe the entire computation fits on the die.
i have tested jev for my use cases and its horrendously wrong, but then the follow up from jev's team is "oh, you need to boil the question down further". it's a spiral of how much do you wanna dumb down the ask so that it answers it correctly. i'll pass for now.
also, 30k input tokens is a lot.
The prefill is the only blocking part, and you can prefill the whole context up to the point where they diverge, then prefill each question and decode the one token in parallel for each question.
If you batch vLLM calls with the same prompt prefix to the same process, it'll deduplicate the prompt prefix across batched requests (+/- the block size) and decode in parallel for each.
That's with a vanilla LLM. If you modify the LLM you can pull that in-graph, but it isn't really necessary.
Fyi, I haven't tested this yet.
When I hear “architecture” I am thinking number of parameters and latency.
When I hear “accuracy” I think training recipe, data, and (later) number of parameters.
So when you say that Jev’s architecture may not be necessary, the evidence I expect to see is comparable quality at comparable latency. Not equal quality at 2x latency and 4x the cost.
But it is likely more than just a fine tune + novel training. At the very least the LM head is swapped out for a classifier one and then or also idk, bidirectional attention for the encoding pass I'm out of my depth at this point and will stop guessing. The training is probably where they have the biggest moat though, not that it's necessarily huge.
I have a project that fits jev as advertised almost comically well and I've been playing with it, and the various hacks and open versions. Jev doesn't necessarily perform better overall but it is quite different. It's sensitive to prompt phrasing in ways the others aren't, it's easy to generate questions where all the other models cluster in confidence but jev is an outlier. Not necessarily more correct, but it does feel like it's getting its answers in a different way.
I'm guessing just as much as anyone else but I've been spending a ton of time on this the last couple weeks, it landed right when I was most ready to dig into it.
> As Privatemode is hosted in the EU and Jev is hosted in the US, we ran four of the datasets from Germany and from the US at the same time. From Germany, Privatemode answered in 180 ms and Jev in 264 ms. From the US, the order reverses: 164 ms for Jev against 299 ms for Privatemode.
turns out there is a trick to keeping the context filled and only evaluating a handful of choice tokens https://www.youtube.com/watch?v=bcGO7xre46o
What isn't obvious is why people keep shouting "Jev Jev Jev" all the time.
Astroturf.
I cannot fathom how people are not seeing the multiple daily posts as anything but the spam they are.