> this won't speed up llama.cpp for everyone, just for users running it in this particular kind of Virtualization.framework VM.

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.

> The fix here works around a problem where the VM was causing llama.cpp to select the wrong kernels.

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

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

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: - https://github.com/openai/tart/issues/501 - https://github.com/openai/tart/issues/1032

UTM also has related cases where apps detect the Apple paravirtual Metal device but falls back to software rendering: https://github.com/utmapp/UTM/issues/7671

a win is still a win

agreed on the title. added more context below on the exact scope and why this is really a VM capability-reporting issue: https://news.ycombinator.com/item?id=49260087

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

Because nobody knows.

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.

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).

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.

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?

https://developer.apple.com/documentation/metal/mtlgpufamily

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

RunAnywhere or Conifer?

I'm hoping AMD wins when the RAM bubble bursts and their integrated AMD 395+ platform can keep getting faster and higher bandwidth.

We've also seen similar improvements on a M5 max. no M1 pro or M3 pro results yet though - would love to see someone try those

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.

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.

I think people generally throw Claude or Codex at the configuration challenge, so they don't know either.

No

[dead]