Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution

https://github.com/chiennv2000/orthrus

246 45 2 months ago

Paper: https://arxiv.org/abs/2605.12825 ; Code+models: https://github.com/chiennv2000/orthrus ; Disclosure: co-author.

Idea: Inject a trainable diffusion attention module into each layer of a frozen AR Transformer. Both heads share one KV cache. Diffusion head projects K=32 tokens in parallel; AR head verifies in a second pass and accepts the longest matching prefix. Output distribution is provably identical to the base model.

Results:

- Up to 7.8x TPF, ~6x wall-clock on MATH-500.

- 16% of params trained, <1B tokens, 24h on 8xH200.

- vs. diffusion LMs (Dream, Fast-dLLM-v2, SDAR, Mercury, Gemini Diffusion): they modify base weights and lose accuracy (Fast-dLLM-v2: -11 pts on MATH-500). Orthrus freezes the backbone; accuracy matches Qwen3-8B exactly.

- vs. Speculative Decoding (EAGLE-3, DFlash): no external drafter, no separate cache, zero TTFT penalty (no drafter to init/sync). KV overhead is O(1) (~4.5 MiB flat). Acceptance length on MATH-500: 11.7 vs. 7.9 (DFlash) vs. 3.5 (EAGLE-3).

- Single-step denoising beats multi-step (6.35 vs. 3.53 TPF). KL distillation beats CE on acceptance rate.

Limitations: strictly bounded by the frozen base model (inherits its biases, hallucinations, knowledge gaps); Qwen3-only evaluation; greedy + rejection sampling only.

ilaksh• 2 months ago

Amazing. Is it possible to do this with Qwen 3.6 27B? Will it work with quants (I assume so)?

sleepyeldrazi6• 2 months ago

From a quick and shallow view of the paper, it looks very feasible (with a little tinkering ) to be adapted to qwen3.6 27B. The process looks somewhat similar to training a LoRA, or in a way distilling your own model so that a mini model learns how to imitate it, and you glue them. I might bite the bullet and rent a gpu to do it for 3.6 27b, as this will solve a lot of my problems.

0-_-02• 2 months ago

3.6 already supports multi token generation AFAIK

littlestymaar2• 2 months ago

So, it's D-Flash but at each transformer layer and share the KV cache of the original model? Very smart!

foobar100002• 2 months ago

Kindof yeah - predictivity is a question though for larger layers - when trying to scale this up. But yeah, this is a "95% predictor in latent space is a 7x improvement in speed if done right" approach.

dot_treo7• 2 months ago

Do you plan on releasing the training code?

jbellis3• 2 months ago

Really cool work!

Does the training data budget scale with model size?

How would you compare the Gemma 4 draft model which is also integrated with the base kv cache?

jbellis3• 2 months ago

BTW the paper says

Since only (Qdiff,Kdiff,Vdiff) are updated during training, the total number of trainable parameters is approximately 16% of the full model.

But the code defines q_proj_diff, k_proj_diff, v_proj_diff, and o_proj_diff, and it only matches 16% when you include the O term.

gkapur• 2 months ago

On the limitation side:

Do you think this would scale to larger transformer models with more parameters per layer?

How would this work with MOE models or sparse models?

deflator• 2 months ago

I'm sure I don't understand all the technical aspects, but I do understand that this is frickin' cool. Nice work.

xiphias2• 2 months ago

The most interesting part of this idea for me is how it wasn't tried / implemented before, as it makes sense.

I haven't read the paper but of course DTree tricks work here as well

bertili• 2 months ago

Does this translate into a similar reduction in compute?

What's the catch?

littlestymaar2• 2 months ago

Does this translate into a similar reduction in compute?

No, quite the opposite actually. Like with speculative decoding this model will compute more tokens and discard the invalid ones.

What's the catch?

LLMs[1] are limited by memory latency and not by compute[2]: because they process tokens one at a time, you spend more time loading and unloading the weights on the GPU registers from VRAM than waiting for compute to happen. Techniques like these allow to process multiple tokens in parallel instead of one by one, and as such exploit better the compute of your graphic card. They do so by predicting which tokens are likely to occur and then verifying that the guess was correct.

For instance if the previous token is “hello”.

A regular autoregressive LLM will compute:

“hello” => “! ”,

then “hello! ” => “how ”,

“hello! how ” => “are ”,

“hello! how are ” => “you”.

and finally “hello! how are you” => “?<end>”

One at a time. Loading and unloading every weights 5 times from the GPU memory to its compute units.

With speculative decoding (I'd say this one isn't strictly speculative decoding, but it's a variant of the same principle), you have something that guesses that the whole sentence is going to be “how are you today?”, so the LLM can generate

“hello” => “! ”,

“hello! ” => “how ”,

“hello! how ” => “are ”,

“hello! how are ” => “you”.

“hello! how are you” => “?<end>”

“hello! how are you today” => “?<end>”

In parallel. So each weight would have been loaded only once from the VRAM instead of 5.

The last token will be discarded though, as the prefix “how are you today” doesn't match what has actually been generated. So in that particular example, you'd have gotten your 5 tokens 5 times faster than with pure autoregressive inference, but at the expense of a 6th token being generated and discarded immediately. So 5 times more token throughtput, but 20% compute cost increase per token.

[1]: autoregressive LLMs, that is. Which are the ones everybody uses because they are the most performant.

[2]: at least when run at low batch size, on your own computer for your personal use. On a datacenter, with many concurrent users, GPUs are actually compute-bound.

kreelman2• 2 months ago

Fantastic results. Well done. ...So this is built into the way the model works.. if I'm understanding it correctly.

I was wondering what would be involved in getting it to work with GGUF files, rather than safetensor files...

foobar100002• 2 months ago

Minor nit re[2]: for agentic workloads that are actually worth money - i.e., claude code and similar, things are either prefill-bound - which this does not help - or more importantly tps/user bound (at 150k+ context windows) - you want your big magic model to emit 200 tps/user. This is why Nvidia bought Groq (now LPU) and what Cerebras is trying to do, etc, etc. So for the stuff that makes money in the field - GPUs are not really compute bound once context lengths are large - but still memory transfer bound (may be KV-cache transfer, may be HBM->SRAM-on-chip, etc..)

dot_treo7• 2 months ago

It is all about moving the bottleneck. During prompt processing everything can be calculated in parallel, while during token generation you create a single token at a time. For example, using an RTX 4000 Ada, I'm getting 2700 t/s for prompt processing, and 48 t/s for token generation using an 8B class model.

Their approach is essentially a speculative decoding approach where multiple tokens are predicted at once and then verified. Therefore getting more tokens to be created at a speed that is closer to the prompt processing speed.

It seems to be special because their approach yields the exact same output distribution as the base model and it only takes a negligable amount of additional memory.

The main catch is that if your prompt processing speed is already bad, it will not help you all that much.

For example, the M-series Macs (up to M4) have a relative high generation speed compared to their prompt processing speed. That means they will not benefit as much (if at all). With the M5 the prompt processing speed has increased 4x, so those can expect to see a good uplift.

dehrmann• 2 months ago

multiple tokens are predicted at once and then verified

Reminds me a little of a carry lookahead adder.

DeathArrow• 2 months ago

If someone can make this work with GGUF and Quantized Qwen 3.6 or Deepseek 4 it would greatly help running local models.

hedgehog• 2 months ago

Multi-token prediction is available now, I'm still getting it set up but it sounds like it'll be 1.5x or 2x on the bigger models.

dot_treo7• 2 months ago

I've tried MTP, and that got me about 1.5x on average with a very spec friendly benchmark.

I didn't run the full benchmark with the demo code, just picked up a single prompt from it. The prompt is about 1300 token, the response is about 3200 token.

Baseline: 44.8 t/s With Orthrus: 164.6 t/s

Note: Don't use the `use_diffusion_mode=` config flag in their example to collect a baseline. Something about how the fallback to "normal" makes it grind to a crawl.

spwa4• 2 months ago

I don't understand. This distills a diffusion transformer out of Qwen3. And while the provably identical is nice, a full diffusion transformer would be a lot faster still.

yorwba• 2 months ago

A full diffusion transformer would need more forward passes (thus being slower) or produce worse output (because it can't properly account for dependencies between tokens when generating them independently in parallel), or both. Keeping the output identical to the autoregressive baseline ensures the speedup doesn't come at the cost of quality degradation.

dnlserrano• 2 months ago

I wonder what our man @antirez will make of this

boredatoms• 2 months ago

So will this help openai/anthropic have lower congestion in the afternoons if they implement something similar?

pornel• 2 months ago

No, it would make it worse.

This adds more computation and sacrifices throughput to improve latency of a serial single-user generation.

Large scale providers run inference in batches, sacrificing latency to gain throughput.

yencabulator• 2 months ago

Non-predicted token generation requires num_of_tokens_output passes over the weights.

Correctly-predicted token generation, requires num_of_tokens_output/prediction_size passes over the weights, plus a much smaller model to make those predictions.

Incorrectly-predicted token generation adds some overhead to the above, relative to the hit rate.

It sounds like good predictions would actually decrease the total overhead while improving latency. (Same FLOPs, but less memory bandwidth consumed -> probably run just as hot, but get more done.)

p0w3n3d• 2 months ago

I hope it helps finally running large models on normal hardware. Tying our work to two companies in the world is a bad bad thing. Quite risky. Against any threat modelling

Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution

Related Stories