DiffusionGemma Technical Report
arxiv.org
arxiv.org
The really interesting thing to me was that they didn’t need to train this model from scratch they just used their existing MOE checkpoint:
“To convert a decoder-only model (Gemma 4 26B A4B) into a denoiser, we can make use of something it is not directly using when generating tokens, namely the logits of all tokens!”
What makes me hopeful about this release is that possibly this same conversion can be applied to other open models and we might see a bunch of diffusion versions of existing local models. It’s exciting stuff!
I like the model a lot and it's fairly good at reasoning. You can also really bend it to your needs. It's designed for machines with more compute than memory bandwidth but IMO does really well on metal.
I've got it up to ~15tok/s on M3-class machines, but I wager there's a bunch of perf on M5 that I just don't have hardware access to unlock.
I tried to implement MTP using the other Gemma MTP heads but I failed to move that perf needle. There's some interesting research to be done about pre-seeding the diffusion canvas from draft models. DiffusionGemma with the right drafter can hit 20-30 tok/s on my machine, but I've been unable to combine the two together to make it faster than what it's been running at so far.
About the drafting/MTP, how would that work? As I understand, MTP/drafting is mostly useful for autoregressive models, not for a diffusion model, because each canvas (in the case of DiffusionGemma at least) is fast enough, what you care about is the "convergence" time before it figures out that the accuracy is good enough to move on to another canvas. This is my understanding at least. Would this drafting for a diffusion model like this mean you'd generate multiple canvas at the same time? That'd also seem strange, as then you'll cut performance for generating a complete canvas... I'm genuinely curious about how it could be applied to DiffusionGemma though, so please do expand :)
I'd be curious to see your CUDA approach - happy to combine efforts as well.
If your model can reason and write code at 1500 toks/sec, then you should end up totally bottlenecked on CPU time the entire time a prompt is active. If you aren't, then you're losing wall time versus competitors. But our whole development stack is based around the idea that programmers spend most of their time thinking, talking and coding, not waiting for the CPU (melting CI clusters being a painful exception to that).
What I'm imagining here is some sort of hybrid mode in which compiling code and running it through an interpreter can be overlapped, so an LLM can propose a change and immediately begin running unit tests while type errors that might affect some other module are found in parallel. And the tests would always run sharded, potentially on a remote cluster, even in local dev.
Obviously this approach is to some extent what made the JVM popular. Java compiles very fast because javac does little more than type checking, and the type system is simple. Then the JVM does the heavy lifting of compilation in parallel with it running. So although Java has a reputation for poor startup times, turnaround times for the JVM can be really excellent compared to something like C++, Swift or Rust. And a lot of startup time pain is just poor frameworks like old Springs that want to reflectively scan the app's files and do other inefficient stuff. More modern frameworks push more to incremental build tasks and can get startup down to very little, <0.5secs for a web server with DB connections to start for instance.
But it feels like this approach should be pushed much further. The model should spend all its time waiting on unit tests to run.
Performance-wise, I don't think diffusion decoding should be much worse than autoregressive decoding, see https://arxiv.org/html/2604.11035 .
As a local LLM user, I really want to see more small diffusion LLM/VLMs with decent performance coming.
I'm sure I have a fundamental misunderstanding of the technology, though.
Diffusion text models for me is the more interesting type of LLMs for local usage, as it really makes good use of single GPUs for single responses, rather than auto-regressive ones, and is a lot faster! Probably the fastest model I've been able to run so far, ending up doing ~670 tok/s (depending on the type of text) on a Pro 6000
My entire brain runs on sentences and words since I have no inner eye or whatever. So my thinking and writing both work kind of forward only.
I wouldn’t have thought that was too unique. But maybe it is?
If I'm asked a hard question, like what city I'd most like to visit, I kind of have to look up cities and cross reference them with ones I've noted that I want to visit, or have visited and liked enough to go back, then actually read the list, and figure out which one, but each time I read a city, I might have a, this would be great in summer or winter or they have a cool festival in June. Which leads to additional question I have to ask. Which I think about roughly the same speed I speak, maybe a word or two faster.
My wife is the complete opposite. She can put together a sentence in an instant in "parallel" so when she says it, she though it already, but not as a sentence.
Human brains all work so crazily weird, and we are only just now started to realize all the nuances.
If I'm just typing blind I'll somehow end up writing random homophones down with completely correct spelling, like hear instead of here, it's bizarre. So that part is sort of structured autogenerated noise that is then consciously either appended or inserted into random spots in text. Maybe it's more of an LLM first pass then diffusion refinement.
Simply predicting the surface of the data you are trying to generate is far more representationally efficient, and perhaps in the next few months we'll see a version of that for LLMs
Right now running it locally is about 7-11 t/s for me (16GB 4080 and plenty of RAM overflow). If it doubled or possibly tripled this, thats a game-changer
It's clearly worse than vanilla 26B-A4B, and lacks some things like structured outputs, and gets some tool calls wrong.
So you have to find a usecase or a hand rolled harness that leverages the cerebras-level TPS while not going off track during (even short) tasks.
RDNA4 is pretty similar to CDNA4, yet over a year after the release of "pro AI" cards like the r9700, they had basic kernels lacking in vllm (like w4a16 int4 kernels) while they were implemented in the datacenter CDNA4 cards.
AMD hardware runs well on llama.cpp because basically anything runs on llama.cpp, especially with vulkan. It's not high praise of AMD's software team to say llama.cpp runs well on their hardware
TIA
Would DiffusionGemma be suitable candidate for DFlash 2?