Mesh LLM: distributed AI computing on iroh

https://www.iroh.computer/blog/mesh-llm

by tionis • 11 days ago

347 94 11 days ago

iroh.computer

turtleyacht• 11 days ago

It sounds like iroh enables distributed compute without having to finangle custom hardware.

jmercouris2• 11 days ago

I thought about this too, but the throughput over a network is incredibly slow. It’s not usable for interactive use.

i38611• 11 days ago

That isn’t true. llama RPC is incredibly slow but staged splits in skippy are orders of magnitude faster.

stymaar7• 11 days ago

Throughput is not a problem as you just share relatively small vectors (a few kilobytes in size), the key issue is network latency.

jmercouris2• 9 days ago

I didn't mean the throughput of the network, I meant the throughput of the calculations. Yes, the latency is the issue.

darkpicnic4• 11 days ago

cocompute.ai is already doing this really well.

SwellJoe4• 11 days ago

Is it? I don't see anything on the website about splitting a model across multiple devices, only about putting local models on the internet, a wholly orthogonal problem (which is already easy with existing tools, since models use an http API).

darkpicnic4• 11 days ago

Good point. I know cocompute is working on splitting, but it’s not there yet; I was referring to the round-robin delegation within a trusted pool. Mesh LLM looks great too!

dnoberon• 11 days ago

Cool, always good to have more in the ecosystem. I love Iroh and hope this continues to succeed.

SwellJoe4• 11 days ago

I note the lack of performance information. I can only imagine it's much, much, slower than any other way to run a larger model (including, e.g. using system RAM and streaming some stuff from disk). Consumer networks, even 10gbit ethernet, are slow as hell compared to local RAM and even disks.

Are we talking 1 token per second for a split model? Less?

Edit: Found a number. On the models list, Qwen 235B A22B says "MoE 235B/22B, proven at 16 tok/s across 2 nodes". They don't say what the nodes are and what network connection they have, but that's a respectable speed. Not quite comfortable for interactive use, but pretty close.

woadwarrior01• 11 days ago

Perf should be fairly straightforward to ballpark. You'll need to transfer roughly 2 . hidden_size . num_shards bytes over the network per token during autoregressive decoding. And divide that number by chunk size during prefill.

i38611• 11 days ago

This was done on my home lab simulating 5ms latency and jitter between machines. Splits work quite well if you your nodes are over WAN at metro latency’s but not super fast on global WAN.

The idea is that you could take several machines without dedicated RDMA or NVLINK fabric and use them to serve a large model on hardware you own then share it with others.

I’m currently working on GLM 5.2 on my lab environment with around 10 tok/s on the same split.

zdw• 11 days ago

What hardware (CPU/GPU/memory) and network was used for this? What quantization for GLM 5.2? How much tuning of the split was needed?

SwellJoe4• 11 days ago

That sounds cool, but it's still pretty meaningless without information about what your home lab looks like. A few DGX Sparks wired up with their fancy super fast network is much different than a few laptops on wifi.

imrehg• 11 days ago

That's about the speed I get on a AMD Ryzen AI 9 HX 370 (inside a Framework 13), with Qwen3.6-35B-A3B, so doing the same on that much larger model...

stymaar7• 11 days ago

I note the lack of performance information. I can only imagine it's much, much, slower than any other way to run a larger model (including, e.g. using system RAM and streaming some stuff from disk)

Not necessarily, and I suspect there are plenty of configuration for which this isn't going to be the case. Let me explain why:

- when offloading the weights to RAM or NVMe, you need to transfer the massive weights from your slow storage to the GPU for each layer being processed for each token. And as such you are being bottlenecked by the transfer bandwidth (which is either the men bandwidth of your DRAM or the read speed of your disk)

- when using a distributed setup, the weights stay in the VRAM on each machine, the it's the GPU memory bandwidth that matters for the weights, and it's much higher than the two other bandwidth discussed above and as such the bottleneck isn't here. You need to tranfert data from a group of layers sitting on one device to the next one another device, but the amount of data is much smaller than the weights (we're talking about kilobytes of data, not gigabytes) so the network throughput isn't a limiting factor.

The limiting factor is the network latency: if you split your model between 4 devices, you'll have 3 times the network latency per token. If you're on a network with 1ms latency, that means 3ms of latency per token. Which means the theoretical upper bound for your inference speed without speculative decoding is 30tps (this theoretical limit assumes the computation itself is instantaneous).

So this is unlikely to be practical over the internet (too high of a latency) but on a local/enterprise network with speculative decoding it could totally work.

Edit: note that all of the above is about token generation, for prefill/prompt-processing the distributed setup will almost certainly win (because in this case, the network latency doesn't add up)

SwellJoe4• 11 days ago

Ah, that's interesting. I though there was more data crossing the network. So, why does a DGX Spark come with super fast network if 10Gbps ethernet would be sufficient for splitting a model? I never bought a second Strix Halo on the assumption that the pipe between them would be a limiting factor to using larger models, so obviously there's something I don't understand.

ul5255• 10 days ago

I’m staring at this comment for a while now: With 3ms latency combined per token, wouldn’t that mean (1 / latency) = 333 token/s for the theoretical upper bound? I’m not trying to nitpick, just curious if I misunderstand something.

chromatin• 10 days ago

A 1 gbe local network should have < 1 msec latency per hop so theoretical upper bound is substantially higher than 30 tps (again assuming instantaneous compute) => thus network latency should not be the limiting factor in reality, no?

darkpicnic4• 11 days ago

Does Mesh LLM encrypt the payload between nodes? Is it possible to read requests from other users?

tekacs• 11 days ago

I'm not affiliated, but yes – the main 'point' of iroh is that it's 'dial-a-key', QUIC with encryption based on the keys of the endpoints.

metadat• 11 days ago

Just wondering, why do you care about encryption in this context?

darkpicnic4• 11 days ago

If payloads to LLMs are being passed around to various nodes, even trusted ones (like friends and family), it gets awkward if you send something very personal. Think sending a medical question to medgemma:27b.

i38611• 11 days ago

I’m one of the contributors to Mesh LLM and happy to answer any questions. I authored the skippy engine that allows you to split large models across nodes.

iotapi322• 11 days ago

This is super impressive, We have a lab with lots of different epycs and different models - to bring them together this way is amazing. Well done!

i38611• 11 days ago

Thank you! AMD is a weak spot in our testing right now. If you’re willing to contribute or let us borrow some compute time, drop in on the Discord.

Lerc• 11 days ago

I have never really delved into kv cache implementation, do they run effectively separate caches per layer?

If so I can see it all dividing nicely, computation and data size wise and the only slowdown would be in search layer waiting for it's turn. If you pipelined it you could run multiple queries.

Is anyone doing best-of-n with a n stage pipeline running each query offset by one?

i38611• 11 days ago

Each stage has its own KV for the layers it hosts. You are on the money there, when one stage is waiting it's free for more parallelism. I am planning on exploiting this for more token verification through ngram spec decoding.

DerivativeBS• 11 days ago

Curious about: does it have fault tolerance if one of the machines goes down mid-inference? Can it dynamically reroute, or does it just retry?

i38611• 11 days ago

It can dynamically route. If a machine drops out of split, the topology is recalculated and the request is automatically retried.

Creamsicle472• 11 days ago

Hey, this is a super cool project. It's great to see a lot of the IPFS stuff resurfacing again.

A few questions:

1.) How does this handle privacy? If you're distributing compute this way then all actors in the compute graph will also know the sequence being computed.

2.) Any safeguards against malicious actors poisoning model activations?

i38611• 11 days ago

To be honest, both are very tough problems we don't have a good answer for yet. If that is something that concerns you, look into building a private mesh with trusted peers.

zmmmmm• 11 days ago

The obvious burning question is how performance looks over different network conditions on some standard models. Have you done much benchmarking? Is it mainly latency affected or is overall throughput less than the capacity of the GPUs due to being distributed?

sig_kill8• 10 days ago

Yeah, this is one area we’re struggling with due to the sheer volume of variations and conditions, but I’ve been thinking of collecting some real-time statistics around latency, prefill/decode, and model distribution… that way we can update some kind of live + aggregated performance numbers for interested parties.

whs2• 11 days ago

I wonder how security is done in this engine, since it's accepting input from anyone. llama.cpp's RPC layer seems to says that you shouldn't run it in public (I assume because it is lower level and may result in RCE on your GPU)

maxgashkov• 11 days ago

What is the incentive for me to join the public mesh? Do you have any fairness guarantees, e.g. if I contribute 1/8th of the VRAM required to run a particular model, do I get at least 1/16th of the inference share, or anything similar to this?

sig_kill8• 10 days ago

Great question! We’ve had a lot of discussions about the direction we want to take this, and how to best generate some kind of incentive / fairness reward.

And we’ve found ourselves hesitating on a direction because, at least for now, the primary use case in a useable env is private hosts you own in a mesh; which makes any sort of reward/ incentive structure somewhat unnecessary (for this setup).

When the public mesh becomes large enough and we get around to extending the existing “mesh governance” features is likely when we’ll add something like you’ve described.

For now, the public mesh is totally open with no restrictions or limitations.

stymaar7• 11 days ago

Is it a fully custom inference engine or are you reusing parts of an existing stack? (llama.CPP, vLLM, etc.)

i38611• 11 days ago

Our skippy library is a patch queue on top of llama that allows us to access internal information, such as activations, and filter tensors on model load.

cromka• 10 days ago

Would this benefit from integrating with the Colibri project announced here just days ago?

potluri• 10 days ago

How does this differ from exo?

sig_kill8• 10 days ago

There’s a breakdown on our docs page you can check out!

https://meshllm.cloud/docs/pages/exo-comparison/

throw12345678913• 10 days ago

A model gets partitioned by layer ranges into stages: layers 0 to 15 on one node, 16 to 31 on the next, and so on down the pipeline.

Numbers in this example are arbitrary. How does it actually work? What if the model’s number of layers is 33, or 34?

Is there a document explaining all constraints of this implementation?

sig_kill8• 10 days ago

It depends on the size / structure of the model… but generally you can check out the MeshLLM huggingface org for pre-computed model splits.

We have a job that will take popular models, and dice them up and post on HF.

When that specific model is served, the layers will be shovelled around the available nodes on the mesh.

dsegovia• 9 days ago

This is surprisingly similar to what I'm trying to build (at least at the network level), down to the use of iroh. I have not attempted model splitting though, and the angle is slightly different. Major props for this, it seems like a great project. The more we can push local inference, the better for everyone. Interested to see if there's collaboration opportunities. https://github.com/dsegovia90/cocompute

_superposition_• 11 days ago

I just wish I had the hardware to try it out!

dwoosley• 11 days ago

I’ve been curious what a polymorphic botnet that runs one (or multiple) distributed LLMs would be capable of doing. The idea would be to evolve the botnet delivery and payload using the clustered compute of all hosts in the botnet to run LLMs that guides the evolution of various botnet clusters. Bad cluster morphs get caught and cleaned off and bad delivery methods never spread, but the best versions survive to continue to grow.

What I envisioned for how it works is fairly similar to this, QUIC can actually be more difficult to detect than it seems since it’s very dynamic.

nullc• 11 days ago

Does this have intelligent expert handling for high parallelism MOE? You can get very high throughput for highly parallel MOE if you can mix different queries at each expert stage, but if the batch has to run together for the whole pipeline you get a parallelism loss instead of gain.

michaelneale• 9 days ago

@null not as yet, but did a tiny bit of research before. If i understand it this would help mainly with batching (ie concurrent sessions/users/turns etc) to keep the pipeline more busy (vs a batch of a fixed size going together), which is interesting, but most work so far has been on latency hiding for straight line performance to start.

We did have a version of things which used expert islands for MoE parallel alternative (I think that has been mostly scrubbed from the code). It showed early promise by having trunk+hot experts together, but as models got larger, it made both that very large on its own but also didn't seem to work as well (or we just weren't good at grouping experts), the idea being to really side step latency and route sessions to those islands. It felt like as models scaled it didn't stay smart (as diverse experts activated more than I thought they would). Could be our mistake (was an exciting possibility though - if you don't mind accepting some loss).

tldr; yes this seems a very nice enhancement for smarter batching/keeping things busy and seems like most larger models we look at are MoE!

whatjustin• 11 days ago

The real test is throughput. I'd like to see tokens/sec at higher concurrency and with uneven hardware.

Abishek_Muthian2• 11 days ago

I'm more interested in running distributed inference for purpose built small language models than these coding LLMs.

Say a distributed inference for image processing, SDR, local weather monitoring etc. These will run on mediocre specs and produce dependable output.

Nicely done OP.

unrvl22• 11 days ago

Something like this is nice, where instead of having 1 model with X active experts, you have 10 different models, all small and dense, trained on specific information. and loaded on 10 different servers, with one router.

Abishek_Muthian2• 10 days ago

True, I've been telling that coding LLMs have been detrimental to local AI scene. What happened to those nifty ML projects which can run on raspberry pi zero.

sig_kill8• 10 days ago

This is actually how I develop and use the mesh at home. Rather than splitting models, I aggregate disparate compute behind one endpoint, without having separate inference providers on each host and a gateway like LiteLLM

downrightmike• 11 days ago

difference between this and Exo?

sig_kill8• 10 days ago

Write up here: https://meshllm.cloud/docs/pages/exo-comparison/

MattPerry• 11 days ago

The first picture "gpu rig", "laptop", "server", "cloud node, etc made me realize how little compute I have. I don't have a laptop with 24GB VRAM or a workstation with 96GB. I think if I convinced all of my friends to run LLMs on their gaming PCs, I don't I would have the total VRAM in the picture.

As an aside, I saw this post mentions a public mesh, but I couldn't find any more information.

kennywinker2• 11 days ago

https://public.meshllm.cloud/

whs2• 11 days ago

I've been looking for similar distributed computing style LLM, and I found AI Horde and a few other smaller efforts like one from Aphrodite people and distributed training from Nous Research.

AI Horde seems to be the biggest of them all. Their API speaks KoboldCPP text completion (not even chat completion). It seems that the community (or at least the active people) strongly prefer it this way because the API exposes more tunables than chat completions, which for roleplay use seems to result in better result. I don't know what else you can use AI Horde for anyway since all other use cases likely will require tool use. Just this week I was set out to improve their OpenAI bridge to support chat templates and response parsing. We'll see if I could get it deployed officially then you might be able to use it to code, although you'll have to use RP models.

I think Horde do have a lot more abuse prevention. Workers needs to have 1 week of cumulative uptime to be considered trusted to prevent brigading - users can opt into trusted workers only. Running a worker give you kudos which is required for >512 max tokens generations and also free requests get bumped to last.

SubiculumCode• 11 days ago

All these ASICS being designed and specialized for AI but none seem to be being built for consumers. Reason?

kennywinker2• 11 days ago

I spent a while trying to get mesh-llm running, but none of the installable llama.cpp builds worked with my older gpu. It looks like it should be able to be used to proxy an external llama.cpp service, but I had no luck setting that up either. Seems very cool, but definitely some rough edges.

i38611• 11 days ago

I’d love a bug report - we can get it working for you!

josefrichter• 11 days ago

Is there a catch? If not, this would be super useful.

stymaar7• 11 days ago

The catch is that the token generation speed is going to be limited by network latency, making it unbearably slow to run over the internet.

It can be great on a local network though, especially if your workload is prefill-heavy (more text input to process than output tokens to emit).

luciana1u• 11 days ago

distributed AI computing so your hallucinations can be geographically diverse too

vigsterkr• 11 days ago

the https://query.mt/ project has been using iroh based mesh for a while. maybe give it a go, especially if you wanna use your mesh models on your mobile phone as well.

derdi• 10 days ago

They might want to make their docs less confusing. https://query.mt/showcase/gpu-sharing/ says "no API key needed", and talks about running llama.cpp on a friend's GPU. No API keys are needed for this. But then the page goes on about who "holds" API keys, and where API keys "stay", where API keys are "locked", etc.

dana321• 10 days ago

I knew this was possible, i asked chatgpt about a year ago and it said no the latency would be too big of a problem. I spent the best part of a year learning libp2p and was looking for a project to do with it at the time.

roger_• 10 days ago

Does this support Qwen 3.6 (e.g. 27B) and the myriad of llama.cpp options (batch sizes, quantization, etc.)?

I'd love to see some performance data.

sig_kill8• 10 days ago

Yes! I’ve worked on the settings interface between our runtime and llamacpp, these are documented and available via our config.toml file

Onavo• 10 days ago

Is this truly more secure though? The host can still see your data.

maccam912• 10 days ago

I have a macbook pro, figured I'd see how easy it was to contribute some vram...

And I can't overstate how easy it was. The swarm page thing had a little "join" button and said to run "mesh-llm --auto". And I did. And it worked first try. That is such an uncommon experience I had to report back. It handled picking a model to serve, downloading it from peers, and to test it I chatted with the model I was hosting, I could see the GPU doing work, etc.

It might be more of an endorsement for iroh than mesh-llm, although I'm sure getting it to all work seamlessly took work on both sides. But to whoever spent the time and energy trying to make it seamless, consider the effort recognized!

postpress• 10 days ago

This got me thinking about experiments with models talking to each other over WebRTC: https://xt-ml.github.io/shadow-claw/

Its sort of a "P2P mesh" :-) Watch four instances of the harness running together and collaborating on checking the weather: https://www.youtube.com/watch?v=h1les1A3gcg

lenerdenator• 10 days ago

Let's say I have a bunch of computers (Raspberry Pi 3+, MBP from 2017, Lenovo T420) sitting around that are on my local network. All are from the last 15 years and my main driver is a M2 MBP with 32 gb of RAM.

Would this let me utilize that spare hardware to do anything meaningful with a self-hosted coding-focused LLM at a somewhat-decent token generating speed, or is this still a pipe dream?

s4saif• 10 days ago

So, how is this different then exo ? it does the same job as well

i38611• 10 days ago

exo is Mac only.

throw12345678913• 10 days ago

no it isn't

brcmthrowaway• 10 days ago

How does this differ to darkbloom?

enbarca• 9 days ago

I'm looking at adopting a mesh network architecture for self-sovereign inference across longitudinal health records.

Mesh LLM: distributed AI computing on iroh

Related Stories