dcreater3 days ago
I think the title buries the lede? Its specific to GPT-OSS and exposes the shady stuff Ollama is doing to acquiesce/curry favor/partner with/get paid by corporate interests
https://www.reddit.com/r/LocalLLaMA/comments/1jzocoo/finally_someone_noticed_this_unfair_situation/
https://github.com/ollama/ollama/issues/5245
Sadly it is not and the issue still remains open after over a year meaning ollama cannot run the latest SOTA open source models unless they covert them to their proprietary format which they do not consistently do.
No surprise I guess given they've taken VC money, refuse to properly attribute the use things like llama.cpp and ggml, have their own model format for.. reasons? and have over 1800 open issues...
Llama-server, ramallama or whatever model switcher ggerganov is working on (he showed previews recently) feel like the way forward.
What would be the best way to use llama.cpp and models that use GGUF these days? ramallama is a good alternative (I guess it is, but it's not completely clear from your message)? Or just use llama.cpp directly, in which case how to ensure I don't get rubbish (like the model asking and answering questions by itself without ever stopping)??
"GLM4-Air":
env:
- "CUDA_VISIBLE_DEVICES=1"
proxy: "http://127.0.0.1:9999"
cmd: >
/app/llama-server
--cache-type-k q8_0 --cache-type-v q8_0
--flash-attn
--ctx-size 32684
--jinja
-ngl 20
--model /modelfiles/GLM-4.5-Air-UD-Q4_K_XL-00001-of-00002.gguf
--port 9999
Ollama native models in their marketplace have these params supposedly set correctly to save you having to do this config but in practice this is hit or miss and often these change from day 0 of the release.
Even using llama.cpp as a library seems like an overkill for most use cases. Ollama could make its life much easier by spawning llama-server as a subprocess listening on a unix socket, and forward requests to it.
One thing I'm curious about: Does ollama support strict structured output or strict tool calls adhering to a json schema? Because it would be insane to rely on a server for agentic use unless your server can guarantee the model will only produce valid json. AFAIK this feature is implemented by llama.cpp, which they no longer use.
Here is some relevant drama on the subject:
https://github.com/ollama/ollama/issues/11714#issuecomment-3172893576
llama.cpp is designed to rapidly adopt research-level optimisations and features, but the downside is that reported speeds change all the time (sometimes faster, sometimes slower) and things break really often. You can't hope to establish contracts with simultaneous releases if there is no guarantee the model will even function.
By reimplementing this layer, Ollama gets to enjoy a kind of LTS status that their partners rely on. It won't be as feature-rich, and definitely won't be as fast, but that's not their goal.
(edit: I think -- after I read some of the links -- I understand why Ollama comes across as less of a hero. Still, I am giving them some benefit of the doubt since they made local models very accessible to plebs like me; and maybe I can graduate to no ollama )
I am biased since I effectively started with Ollama as my main local llm so take this response for what it is.
Still, you got me curious. Which defaults you consider hostile ( not disagreeing; this is pure curiosity )?
The infamous Ollama context limits, for one.
What's wrong with using an older well-tested build of llama.cpp, instead of reinventing the wheel? Like every linux distro ever who's ever ran into this issue?
Red Hat doesn't ship the latest build of the linux kernel to production. And Red Hat didn't reinvent the linux kernel for shits and giggles.
> What's wrong with using an older well-tested build of llama.cpp, instead of reinventing the wheel?
Yeah, they tried this, this was the old setup as I understand it. But every time they needed support for a new model and had to update llama.cpp, an old model would break and one of their partners would go ape on them. They said it happened more than once, but one particular case (wish I could remember what it was) was so bad they felt they had no choice but to reimplement. It's the lowest risk strategy.
A much lower risk strategy would be using multiple versions of llama-server to keep supporting old models that would break on newer llama.cpp versions.
llamacpp> ls -l \*llama\*
-rwxr-xr-x 1 root root 2505480 Aug 7 05:06 libllama.so
-rwxr-xr-x 1 root root 5092024 Aug 7 05:23 llama-server
I don't buy it. They're not willing to make an 700MB download a few megabytes bigger to ~730MB, but they are willing to support a fork/rewrite indefinitely (and the fork is outside of their core competency, as seen by the current issues)? What kind of decisionmaking is that?
Shouldn't any such regressions be regarded as bugs in llama.cpp and fixed there? Surely the Ollama folks can test and benchmark the main models that people care about before shipping the update in a stable release. That would be a lot easier than trying to reimplement major parts of llama.cpp from scratch.
There was drama that ollama doesn’t credit llama.cpp and most likely crediting it was „not aligning with their goals”.
A lot of things happened to get to the point they're getting called out aggressively in public on their own repo by nice people, and I hope people don't misread a weak excuse made in conversation as solid rationale, based on innuendo. llama.cpp has been just fine for me, running on CI on every platform you can think of, for 2 years.
EDIT: I can't reply, but, see anoncareer2012's reply.
> llama.cpp has been just fine for me.
Of course, so you really shouldn't use Ollama then.
Ollama isn't a hobby project anymore, they were the only ones at the table with OpenAI many months before the release of GPT-OSS. I honestly don't think they care one bit about the community drama at this point. We don't have to like it, but I guess now they get to shape the narrative. That's their stance, and likely the stance of their industry partners too. I'm just the messenger.
In the spirit of TFA:
This isn't true, at all. I don't know where the idea comes from.
You've been repeating this claim frequently. You were corrected on this 2 hours ago. llama.cpp had early access to it just as well.
It's bizarre for several reasons:
1. It is a fantasy that engineering involves seats at tables and bands of brothers growing from a hobby to a ???, one I find appealing and romantic. But, fantasy nonetheless. Additionally, no one mentioned or implied anything about it being a hobby or unserious.
2. Even if it wasn't a fantasy, it's definitely not what happened here. That's what TFA is about, ffs.
No heroics, they got the ultimate embarrassing thing that can happen to a project piggybacking on FOSS: ollama can't work with the materials OpenAI put out to help ollama users because llama.cpp and ollama had separate day 1 landings of code, and ollama has 0 path to forking literally the entire community to use their format. They were working so loosely with OpenAI that OpenAI assumed they were being sane and weren't attempting to use it as an excuse to force a community fork of GGUF and no one realized until after it shipped.
3. I've seen multiple comments from you this afternoon spiking out odd narratives about Ollama and llama.cpp, that don't make sense at their face from the perspective of someone who also deps on llama.cpp. AFAICT you understood the GGML fork as some halcyon moment of freedom / not-hobbiness for a project you root for. That's fine. Unfortunately, reality is intruding, hence TFA. Given you're aware, it makes your humbleness re: knowing whats going on here sound very fake, especially when it precedes another rush of false claims.
4. I think at some point you owe it to even yourself, if not the community, to take a step back and slow down on the misleading claims. I'm seeing more of a gish-gallop than an attempt to recalibrate your technical understanding.
It's been almost 2 hours since you claimed you were sure there were multiple huge breakages due to bad code quality in llama.cpp, and here, we see you reframe that claim as a much weaker one someone else made to you vaguely.
Maybe a good first step to avoiding information pollution here would be to invest time spent repeating other peoples technical claims you didn't understand, and find some of those breakages you know for sure happened, as promised previously.
In general, I sense a passionate but youthful spirit, not an astro-turfer, and this isn't a group of professionals being disrespected because people still think they're a hobby project. Again, that's what the article is about.
For transparency, I attended ICML2025 where Ollama had set up a booth and had a casual conversation with the representatives there (one of whom turned out to lead the Ollama project) before they went to their 2nd birthday celebration. I'm repeating what I can remember from the conversation, about ten minutes or so. I am a researcher not affiliated with the development of llama.cpp or Ollama.
> for a project you root for
I don't use Ollama, and I certainly don't root for it. I'm a little disappointed that people would assume this. I also don't use llama.cpp and it seems that is the problem. I'm not really interested in the drama, I just want to understand what these projects want to do. I work in theory and try to stay up to date on how the general public can run LLMs locally.
> no one realized until after it shipped.
I'm not sensing that the devs at Ollama are particularly competent, especially when compared to the behemoths at llama.cpp. To me, this helps explain why their actions differ from their claimed motivation, but this is probably because I prefer to assume incompetence over something sinister.
> as promised previously...
I don't think I made any such promises. I can't cite those claimed breakages, because I do not remember further details from the conversation needed to find them. The guy made a strong point to claim they had happened and there was enough frustrated rambling there to believe him. If I had more, I would have cited them. I remember seeing news regarding the deprecation of multimodal support hence the "I could swear that" comment (although I regret this language, and wish I could edit the comment to tone it down a bit), but I do not think this was what the rep cited. I had hoped that someone could fill in the blanks there, but if knowledgeable folks claim this is impossible (which is hard to believe for a project of this size, but I digress), I defer to their expert opinion here.
> llama.cpp had early access to it just as well.
I knew this from the conversation, but was told Ollama had even earlier discussions with OpenAI as the initial point of contact. Again, this is what I was told, so feel free to critique it. At that time, the rep could not explicitly disclose that it was OpenAI, but it was pretty obvious from the timing due to the delay.
> avoiding information pollution
I'm a big believer in free speech and that the truth will always come out eventually.
> I'm seeing more of a gish-gallop than an attempt to recalibrate your technical understanding...
> I sense a passionate but youthful spirit, not an astro-turfer,
This is pretty humbling, and frankly comes off a little patronizing, but I suppose this is what happens when I step out of my lane. My objective was to stimulate further conversation and share a perspective I thought was unique. I can see this was not welcome, my apologies.
Your rationale for your choices doesn't really matter, you made the choice to cast aspersions, repeatedly, on multiple stories in multiple comments.
Handwaving about how the truth will come out through discussion, even when you repeatedly cast aspersions you disclaim understanding of, while also expressing surprise you got a reply, while also always making up more stuff in service of justifying the initial aspersions is indicative of how coherent your approach seems to the rest of us.
In general, I understand why you feel patronized. Between the general incoherence, this thread, and the previous thread where you're applying not-even-wrong, in the Pauli sense, concepts like LTS and Linux kernel dev to this situation, the only real choice that lets anyone assume the best of you is you're 15-25 and don't have a great grasp of tech yet.
Otherwise, you're just some guy gossiping, getting smarmy about it, with enough of an IQ to explain back to yourself why you didn't mean to do it. 0 idea why someone older and technical would do all that on behalf of software they don't use and don't understand.
And also, Ollama is claiming not to use llama.cpp, despite continuing to use llama.cpp.
> Looking at ollama's modifications in ggml, they have too much branching in their MXFP4 kernels and the attention sinks implementation is really inefficient. Along with other inefficiencies, I expect the performance is going to be quite bad in ollama.
ollama responded to that
> Ollama has worked to correctly implement MXFP4, and for launch we've worked to validate correctness against the reference implementations against OpenAI's own. > Will share more later, but here is some testing from the public (@ivanfioravanti ) not done by us - and not paid or
leading to another response
> I am sure you worked hard and did your best. > But, this ollama TG graph makes no sense - speed cannot increase at larger context. Do you by any chance limit the context to 8k tokens? > Why is 16k total processing time less than 8k?
Whether or not Ollama's claim is right, I find this "we used your thing, but we know better, we'll share details later" behaviour a bit weird.
Do they? They probably care more about their "partners".
As GP said:
By reimplementing this layer, Ollama gets to enjoy a kind of LTS status that their partners rely onIronic that (according to the article) ollama rushed to implement GPT-OSS support, and thus broke the rest of the gguf quants (iiuc correctly).
Ollama is a scam trying to E-E-E the rising hype wave of local LLMs while the getting is still good.
Sorry, but somebody has to voice the elephant in the room here.
On llama.cpp breaking things, that's the pace of innovation. It feels like a new model with a new architecture is being released every week. Guess what? The same things we saw with drivers for Unix systems back in the day, no documentation. So implementation is based on whatever can be figured from the arxiv paper, other implementations transformers/vllm (python -> C), quite often these models released from labs are "broken", jinja.template ain't easy! Bad templates will break the model generation, tool calling, agentic flow, etc. Folks will sometimes blame llama.cpp, sometimes the implementation is correct but the issue is that since it's main format is guff and anyone can generate a gguf, quite often experimental gguf is generated and released by folks excited to be the first to try a new model. Then llama.cpp gets the blame.
Makes their VCs think they're doing more, and have more ownership, rather than being a do-nothing wrapper with some analytics and S3 buckets that rehost models from HF.
As far as I understand this is generally not possible at the model level. Best you can do is wrap the call in a (non-llm) json schema validator, and emit an error json in case the llm output does not match the schema, which is what some APIs do for you, but not very complicated to do yourself.
Someone correct me if I'm wrong
This is absolutely possible to do at the model level via logit shaping. Llama-cpp’s functionality for this is called GBNF. It’s tightly integrated into the token sampling infrastructure, and is what ollama builds upon for their json schema functionality.
Do you mean the functionality of generating ebnf grammar and from a json schema use it for sampling is part of ggml, and all they have to do is use it?
I assumed that this was part of llama.cpp, and another feature they have to re-implement and maintain.
Ollama has a hardcoded GBNF grammar to force generic json output for example, the code is here: https://github.com/ollama/ollama/blob/da09488fbfc437c55a94bc5374b0850d935ea09f/llm/server.go#L634-L654
Ollama can also turn user-passed json schema into a more tightly specified GBNF grammar, the code is here and is a bit harder to understand: https://github.com/ollama/ollama/blob/da09488fbfc437c55a94bc5374b0850d935ea09f/llama/sampling_ext.cpp#L62
This thread was about doing structured generation in a model-agnostic way without wrapping try/except around json.parse(), and GBNF is _the_ way to do that.
I'd recommend taking a look at https://github.com/containers/ramalama its more similar to what you're describing in the way it uses llama-server, also it is container native by default which is nice for portability.
> Ollama does not use llama.cpp anymore; we do still keep it and occasionally update it to remain compatible for older models for when we used it.
The linked PR is doing "occasionally update it" I guess? Note that "vendored" in the PR title often means to take a snapshot to pin a specific version.
Here’s the maintainer of ggml explaining the context behind this change: https://github.com/ollama/ollama/issues/11714#issuecomment-3172893576
It's "Ollama's forked ggml is incompatible with other gpt-oss GGUFs"
and it should link to GG's comment[0]
[0] https://github.com/ollama/ollama/issues/11714#issuecomment-3172893576
As far as the title, yeah I didn't work too hard on making it good, sorry
Docker Model Runner makes it easy to manage, run, and deploy AI models using Docker. Designed for developers, Docker Model Runner streamlines the process of pulling, running, and serving large language models (LLMs) and other AI models directly from Docker Hub or any OCI-compliant registry.
Whether you're building generative AI applications, experimenting with machine learning workflows, or integrating AI into your software development lifecycle, Docker Model Runner provides a consistent, secure, and efficient way to work with AI models locally.
For more details check this out : https://docs.docker.com/ai/model-runner/
The simple thing to do is to just use the custom quantization that OpenAI used for gpt-oss and use GGUF for other models.
Using Huggingface, LM Studio, etc. is the Linux metaphor of flexibility. Using Ollama is sort of like using macOS
classic Docker Hub playbook: spread the habit for free → capture workflows → charge for scale.
moat isn't in inference speed — it’s in controlling the distribution & default UX, once they own that, can start rent-gate it.
To me, one of the benefits of running a model locally is learning how all this stuff works, so Ollama never had any appeal. But most people just want stuff to work without putting in the effort to understand how it all fits together. Ollama meets that demand.
Figured it had to be Ollama doing Ollama things, seems that was indeed the case.
> llama.cpp failed to download the Ollama-downloaded gpt-oss 20b model
That should of course read
"llama.cpp failed to load the Ollama-downloaded gpt-oss 20b model"
Ollama forked to get it working for day 1 compatibility. They need to get their system back in line with mainline because of that choice. That's kinda how open source works.
The uproar over this (mostly on reddit and x) seems unwarranted. New models regularly have compatibility issues for much longer than this.