AdapTive-LeArning Speculator System (ATLAS): Faster LLM inference
together.ai182 points by alecco 13 hours ago
182 points by alecco 13 hours ago
Inference is impressively fast. But what about quality? In the Kimi vendor verifier (https://github.com/MoonshotAI/K2-Vendor-Verifier/), Together has one of the highest tool call failure rates (>300 failures over the benchmark, compared to 0-2 for the official API, groq, SiliconFlow, and Infinigence).
I don't know anything about Together quality in general, but the specific technique discussed here (speculative decoding) has no impact on the quality of generations. So you should be able to apply it to whichever model you want, and see the advertised speedup while retaining the quality of your base model.
> the specific technique discussed here (speculative decoding) has no impact on the quality of generations
I don't see why that would be true. As I understand, the verifier is checking if the tokens are good-enough, not if they're the exact same tokens it would have selected. The predicted tokens could be consistently slightly worse, which could have a cascading effect to make the overall output a lot worse.
It can be exact or not! Depends on the kind of sampling you are doing.
You can do exact verification, and as soon as a token mismatches you reject everything after that token from your draft. Relaxed acceptance techniques measure how wrong that mispredicted token is via some metric, and accept it if it’s close enough. So you get longer draft lengths with higher acceptance rates.
> the verifier is checking if the tokens are good-enough, not if they're the exact same tokens it would have selected
That's up to you, depends on how you implement it and how much you want to prioritize speed at the expense of quality, this is not an intrinsic attribute of speculative decoding. The verifier checks if the tokens predicted by the draft model are part of the top-k tokens predicted by the full size model at each steps. Set k to 1 and you will only accept perfect matches. Set k to > 1 and you will indeed start selecting "good enough" tokens, but will get faster inference.
But no matter what value you choose for k, the technique described in the article can apply and will result in faster inference at no loss when compared to a setup without this technique, with the same value of k.
Adding to the prior comments as my intuition matched yours, there’s a nice Reddit thread that gives some context into how it can be faster even if you require exact matches: https://www.reddit.com/r/LocalLLaMA/s/ARxHLqRjdM
The TLDR/key (from my understanding) is that verifying N tokens can be faster than generating N tokens.
> The TLDR/key (from my understanding) is that verifying N tokens can be faster than generating N tokens.
Yes. This is because to generate token n+1 you need token n etc. So generating from scratch is a sequential (thus slow) process. When we verify tokens, we can, for each token, use all preceding tokens as input and check that the output token matches the expectation. But since the full sequence we want to verify already exist, we can do it in parallel for each token we want to verify and not sequentially.
This is why training transformer models is much faster than RNN, we do the same thing during training, it's just that the sequence we compare to is the ground truth and not coming from another model.
I didn't know this! I've always thought speculative decoding was "if p(draft_token) > threshold, use it". You made me go read how it actually works and it's pretty neat!
That said, I still think some providers are cheating. Please correct me if the test below is flawed.
I generated texts at temperature = 0 vs temperature = 2. At high temperature, the distributions effectively become flatter, meaning the difference between real and draft effective distributions (the D_LK used in theorem 3.5 of 2211.17192) becomes smaller. When T=2, the model speaks complete gibberish, so the effective distribution must be pretty flat. This should mean fewer rejections --> a lot faster speculative decoding. Yet, I see no increase in throughput at all...
Not sure exactly what setup you are running, in theory yes, higher temperature for both model means higher chance of overlap and thus less rejections -> faster sampling (but worse quality overall).
However, if you have higher temperature but still are operating under a top-k sampling where k is small, not sure it's going to translate to any noticeable difference, since this will make your actual distributions very much non-uniform.
This is with Together's API via OpenRouter, running DeepSeek V3 0324 and Kimi K2 0905.
I didn't set a top-k. So it seems like Together must be doing something weird in their speculative decoding implementation.
Oh in that case there is definitely a top-k or top-p behind the scene, it might just not be exposed to the user as a param they can change through their API. I haven’t heard of anyone running a LLM in prod with actual pure sampling
If you compare "schema validation error count" plus "Count of Finish Reason others" then SiliconFlow and Infinigence is in the same bucket too. Maybe their API layer detected incorrect tool call and set finish reason to something else?
IMO this likely is what you get from running the model correctly as-is (i.e. using the same weight and activation dtype), so Together is not bad.
Moonshot AI themselves and Groq likely uses some sampler tricks to eliminate schema validation errors.
So really the only thing this shows is: Nebius, Chutes, AtlasCloud could be running something else (for example further quantized model). Or bugs.
Fair point. If Moonshot is holding back the true weights or inference techniques that affect correctness, then providers including Together should call them out on that. I for one would stop using Kimi if that is the case.
Anyway, Novita is doing significantly better on the vendor verifier chart than Together, so the low quality must be partially Together's fault at least.
I don't think it's weight being different or special inference techniques, more like they are not able to train the model to follow tool schema perfectly yet, and both Moonshot and Groq decided to use something like https://github.com/noamgat/lm-format-enforcer to make sure at least the output format is correct.
Do I understand this right?
A light-weight speculative model adapts to usage, keeping the acceptance rate for the static heavy-weight model within acceptable bounds.
Do they adapt with LoRAs?
>a faster speculator (also known as the draft model) proposes multiple tokens ahead, and the target model verifies them in parallel in a single forward pass
TIL. Bit of an aha moment - never understood till now how a big model can verify faster than it can generate
As with almost everything else in CS, it's a tradeoff. Pre-fill is compute bound, decoding is memory bandwidth bound. Speculative decoding works when the draft model is more often right that wrong, because most architectures have a lot more compute, compared to memory bandwidth.
So with a 4x speed-up, Together will give us at least 2x lower price for top-end models, right? :)
At first glance, this reminds me of how branch prediction is utilized in CPUs to speedup execution. As I understand it, this development is like a form of soft branch prediction over language trajectories: a small model predicts what the main model will do, takes few steps ahead and then verifies the results (and this can be done in parallel). If it checks out, you just jump forward, it not you take miss but its rare. I find it funny how small-big ideas like this come up in different context again and again in history of our technological development. Of course ideas as always are cheap. The hard part is how to actually use them and cash in on them.
A lot of optimizations in LLMs now are low hanging fruits inspired by techniques in classical computer science. Another one that comes to mind is paged KV caching which is based on memory paging.
No barrier to entry whatsoever? Backprop on the speculative decoding weights during inference to improve their accuracy on a per application basis?
Cool hack though, kudos. Wonder if they can make Groq or Cerebras do the same thing?
Will need some time to go through the details, but it’s increasingly rare to see teams consistently delivering meaningful improvements in the open. Impressive work!
I can't help but wonder how much longer we'll see this work shared openly.
> Built on top of Together Turbo Speculator, ATLAS reaches up to 500 TPS on DeepSeek-V3.1 and up to 460 TPS on Kimi-K2 in a fully adapted scenario — 2.65x faster than standard decoding, outperforming even specialized hardware like Groq
and yet, if you click on: https://openrouter.ai/moonshotai/kimi-k2-0905
You'll see Groq averaging 1,086tps vs Together doing 59tps. Groq and Cerebras often feel like the only games in town. I'd love that to be different (because I'd like more models!), but nobody else is coming close right now.
Comparing how quickly gpt-oss-120b runs gives a broader picture: https://openrouter.ai/openai/gpt-oss-120b -- Vertex (Google) and SambaNova do pretty good on it too, but still, the difference between a top provider and an also-ran is giant.
God I love OpenRouter.
> I'd love that to be different (because I'd like more models!), but nobody else is coming close right now.
I'm currently on the Cerebras Code subscription for like 50 USD a month because it more or less makes the rate limits I used to deal with other platforms disappear (without making me spend upwards of 100 USD paying per token): https://www.cerebras.ai/blog/introducing-cerebras-code
At the same time, their Qwen Coder 480B model is fine but I still find myself going for Claude or GPT-5 or Gemini 2.5 Pro for more complex issues (or ones where I need good usage of Latvian language), at least for programming tasks it'd eventually be super cool if they could offer more models.
Or have some sort of a partnership with Anthropic or whoever, because getting my questions answered at around 500-1500 TPS is really, really pleasant, especially for agentic use cases with code modifications, even if I still bump into the 128k context limits occasionally.
Interesting, if you take a look at the median throughput chart [0], groq goes insane after 7th Oct. Wonder what happened.
[0] https://openrouter.ai/moonshotai/kimi-k2-0905/performance
2x jump overnight. new LPU hardware? I checked the speed for groq's gpt-oss-120B, Llama4-maverick, and Llama4-scout; none of them had a noticeable change this month
Heavy quantization
They claim (or someone on Reddit who claims to be staff claims) that's not accurate: https://www.reddit.com/r/LocalLLaMA/comments/1mk4kt0/comment...
There's another angle to this comparison. Groq and Cerebras use custom chips, but I'm not sure about Together. In this case, Together is sharing results based on the B200 GPU. Another important point is the accuracy of these speed-ups compared to the baseline model. It's known that such tricks reduce accuracy, but by how much? Kimi has already benchmarked several providers. https://x.com/Kimi_Moonshot/status/1976926483319763130
> It's known that such tricks reduce accuracy
AFAIU, speculative decoding (and this fancier version of spec. decoding) does not reduce accuracy.
No it shouldn't do. "All" you're doing is having a small model run the prompt and then have the large model "verify" it. When the large model diverges from the small one, you restart the process again.
> Groq and Cerebras use custom chips
Not just custom chips, but custom chips which derive much of their performance from enormous amounts of SRAM. There's no denying that approach is fast, but it's also incredibly expensive, and SRAM scaling has slowed to a crawl so it won't get much cheaper any time soon.
This is an "expensive for whom" question. I'd be keen to know if they're burning investor money hosting these right now or if they're able to run these at cost.
But Groq/Cerebras are hardware accelerators. It's an unrelated optimization. I wouldn't be surprised if they could also use speculators (today or in the future).
> You'll see Groq averaging 1,086tps
What I don't understand is, Groq reporting 200tps for the same model: https://console.groq.com/docs/model/moonshotai/kimi-k2-instr...
OpenRouter numbers look fishy.
groq is quantizing, even though it's not labeled as such on openrouter (super frustrating)
Do you have a source for that? They are pretty close to the ref implementation on moonshot’s ranking
>Groq and Cerebras often feel like the only games in town.
SambaNova should be similar...they've got a similar specialized hardware approach
Do these numbers compare performance at the same cost?
You can see the cost in the links, and the answer is “pretty much” for the consumer. The backend maths, no idea.
Great, my slop memes can come out much faster now. This is the future of the world economy!