It looks to me like this won't speed up llama.cpp for everyone, just for users running it in this particular kind of Virtualization.framework VM.<p>The fix here works around a problem where the VM was causing llama.cpp to select the wrong kernels.
> this won't speed up llama.cpp for everyone, just for users running it in this particular kind of Virtualization.framework VM.<p>correct. these figures apply to llama.cpp inside the macOS guest configuration we tested. Lume is the VM frontend we used, while Apple's Virtualization.framework provides the virtual GPU. bare-metal llama.cpp is unaffected.<p>> The fix here works around a problem where the VM was causing llama.cpp to select the wrong kernels.<p>mostly, with one nuance: llama.cpp is selecting the correct kernels for the capability answers it receives. the stock guest reports an older Apple GPU family and a 32 KB threadgroup memory limit, so llama.cpp chooses slower kernels. Our process-scoped layer reports the tested Apple 9 and 64 KB values while allowing llama.cpp to select newer paths that the paravirtual GPU successfully execute<p>the layer itself though works at the Metal API boundary, independently of llama.cpp. other Metal compute and graphics apps now may select newer paths from the same capability answers, although this is still preliminary and each app needs separate testing. for example, MLX-LM stayed flat in our tests<p>historically related limitations have been coming up across Apple Silicon VM frontends for a while e.g. Tart tracked MPS/GPU support back in 2023:
- <a href="https://github.com/openai/tart/issues/501" rel="nofollow">https://github.com/openai/tart/issues/501</a>
- <a href="https://github.com/openai/tart/issues/1032" rel="nofollow">https://github.com/openai/tart/issues/1032</a><p>UTM also has related cases where apps detect the Apple paravirtual Metal device but falls back to software rendering: <a href="https://github.com/utmapp/UTM/issues/7671" rel="nofollow">https://github.com/utmapp/UTM/issues/7671</a>
a win is still a win
That makes sense. The title initially sounded like a general llama.cpp speedup on Apple Silicon, but if the improvement comes from fixing kernel selection inside Virtualization.framework VMs, that distinction is pretty important.
What I don't get, which this article doesn't talk about, why would Apple’s Virtualization.framework expose a lesser Metal profile instead of reporting all capabilities supported by the host GPU?
Because nobody knows.<p>Apple doesn't let you "pass" the GPU through to a VM like most other ARM/x86_64 processors (forwarding interrupts and PCIe memory regions). There are symbols defined to do this within the kernel (if you dump the binary) but they aren't used in retail macos.<p>Instead you end up creating a paravirtual device that emulates the GPU acting like a 'normal PCI device' which you give to clients. This is usually reserved (by other hardware vendors) for when you're doing multi-tenat time sharing of higher end GPUs (like Nvidia enterprise cards can do).<p>These paravirtualized GPUs then just have 'less features' and Apple (being Apple) states no reason why.
All M-series chips support Metal 4. Wonder if we can fix this with a simple override somewhere.
QEMU/kvm does this as a default, because keeping a more generic CPU / etc makes moving VMs between machines with different hardware possible. If you try to move a VM it won't work, of course, if the new machine doesn't support what the old did.<p>Not sure of this is why Apple does it. With KVM, you tend to pick a baseline that all your machines support.
Because it cannot be safely virtualised?
> 11.08× faster and generated tokens 16.36× faster than the same workload in the same stock VM.<p>So this was the comparison, for me the title was a bit confusing
yeah fair point. it's always tricky to get the whole idea across within HN's title limit. tldr: we ran the same workload in the same Lume macOS VM on the same Apple Silicon host, first with stock Metal capability reporting and then with our process-scoped dynamic library. The 11.08x figure is prompt processing, while 16.36x is token generation. the mechanism technically extends to graphics workloads too but these figures are specifically from llama.cpp
I don’t understand what Apple 1-9 are. At first I thought it was M series chips but there is no M9 (yet)
my whole setup is buy more RAM, run it on CPU, and tell myself the GPU is just a personality trait I'm working on.
I recall there was another YC startup that was working on Mac-specific ML optimizations for local inference (and perhaps fine-tuning).<p>I wonder if their work is related?
All this work to get the Mac to be a platform useful for AI is being done despite Apple's efforts. They're famously pissed at Nvidia since the Nvidia + Intel Macs due to heat and other issues. But then the OS is a bit closed off and they move slow and are more focused on milking iOS and the App Store and services for money BUT the PA-Semi purchase and Apple Silicon and everything following it has made the hardware just so amazing and useful that despite all that people build for it.<p>I love the platform. I'm happy to see people building on it.<p>AND! if we ever get an M7 chip with the rumored 1.5TB of available ram all this work will not have been in vain. You think the ai acceleration is nice in the M5 wait till M7 and M8.
apple silicon is what made us start Lume in the first place last year. the hardware is so good (M1 is now 6 years old!) that people keep pushing through the gaps in the platform. just yesterday we ran a fully offline computer-use agent with Cua Driver and Muse Glimmer, all locally on Apple Silicon, an now the same kind of agent can run isolated inside a macOS VM and use Apple’s GPU path too
[dead]
[dead]
[flagged]
What hardware do you run? I have a first-gen mac studio, and I just run cmake and build with no special options. Same thing with llama-server, I just specify the model and use the built-in web UI.<p>For reference, I get ~26 tok/sec with the new Muse 30B model.
There are certainly challenges. When setting up a new model, I get AI to walk me through the commands using llama-benchmark that determine the best parameters for my particular configuration and needs. Once you've got that it's pretty easy to port those parameters to llama-server. It takes me about an hour to run through this process. It would be great if there was a registry of hardware, models, configuration parameters, and resulting tokens per second. Maybe one day we'll get there.
What kind of parameters do you end up changing, and how much difference does it make. Perhaps I am missing something and get more tok/sec, but I usually just do a git pull, then rebuild the latest whenever I get a new model.<p>In the past, I had to play with chat templates for some models to work with agents for tool calling. But I've never had to do anything other than specify the model, and tweaking the context size in some cases.
> It takes me about an hour to run through this process<p>Yeah not doing that
I think people generally throw Claude or Codex at the configuration challenge, so they don't know either.
No
[flagged]
The Claudish in the blogpost makes it really hard to ready. Also, TinyLlama 1.1B lol.