upvote
From a debugging point of view, the author's conclusion was still completely reasonable given the evidence they had
reply
No it wasn't. A hardware defect so disastrous that it affects floating point computation on the neural engine, yet so minor that it does not affect any of the software on the device utilizing that hardware is exceedingly improbable.

The conclusion, that it was not the fault of the developer was correct, but assuming anything other than a problem at some point in the software stack is unreasonable.

reply
Nah.

All neural accelerator hardware models and all neural accelerator software stacks output slightly different results. That is a truth of the world.

The same is true for GPUs and 3d rendering stacks too.

We don't usually notice that, because the tasks themselves tolerate those minor errors. You can't easily tell the difference between an LLM that had 0.00001% of its least significant bits perturbed one way and one that had them perturbed the other.

But you could absolutely construct a degenerate edge case that causes those tiny perturbances to fuck with everything fiercely. And very rarely, this kind of thing might happen naturally.

reply
You are correct that implementations of numerical functions in hardware differ, but I do not think you correctly understand the implications of this.

>And very rarely, this kind of thing might happen naturally.

It is not a question of rarity, it is a question of the stability of the numerical problem. Luckily most of the computation in an LLM is matrix multiplication, which is s extremely well understood numerical problem and which can be checked for good condition.

Two different numerical implementations on a well conditioned problem and which requires much computation, differing significantly would indicate a disastrous fault in the design or condition of the hardware, which would be noticed by most computations done on that hardware.

If you weigh the likelihood of OP running into a hardware bug, causing significant numerical error on one specific computational model against the alternative explanation of a problem in the software stack it is clear that the later explanation is orders of magnitude more likely. Finding a single floating point arithmetic hardware bug is exceedingly rare (although Intel had one), but stacking them up in a way in which one particular neural network does not function, while other functions on the hardware run perfectly fine, is astronomically unlikely.

reply
> yet so minor that it does not affect any of the software on the device utilizing that hardware

You're being unfair here. The showpiece software that uses that hardware wouldn't install, and almost all software ignores it.

reply
The hardware itself is utilized by many pieces of software on any Apple device. Face ID uses it, Siri uses it, the camera uses it, there are also other Apple on device LLM features, where you could easily test whether the basic capabilities are there.

I highly doubt that you could have a usable iPhone with a broken neural engine, at the very least it would be obvious to the user that there is something very wrong going on.

reply
> The conclusion, that it was not the fault of the developer was correct, but assuming anything other than a problem at some point in the software stack is unreasonable.

Aah, the old "you're holding it wrong" defense.

reply
What do you mean? The developer is perfectly justified in being upset over a basic example not functioning correctly, due to bug on behalf of Apple's developers. It just wasn't reasonable to assume that the bug was due to malfunctioning hardware.
reply
Apple's documentation is utter garbage, but this code almost seems like a separate issue (and notably the MLX library uses loads of undocumented properties in metal which isn't cool). It looks like the change used to allow the NAX kernel to be used on the iPhone 17 or upcoming 18 if you're on 26.2 or later, to instead only allow it on the iPhone 17 Pro or upcoming 18. I'm fairly sure the GPU arch on the A19 is 17. They changed it so it will only use that kernel on the 17 Pro or upcoming 18, which is notable as the A19 Pro in the 17 Pro has a significantly changed GPU, including GPU tensor cores. The only real change here is that it would limit to the pro variants for the "17" model.
reply
> The neural accelerator exists in iPhones going back many years.

What has existed before is the Apple Neural Engine (ANE) which is very different from the newer Neural Accelerator support within the GPU blocks. In fact MLX does not even support ANE yet since at least in previous versions it was hardware-limited to computing FP16 and INT8 MADDs, and not even that fast.

reply
Sure, I directly and explicitly talked about Apple's version of tensor cores in the GPU. But the ANE is by every definition a neural accelerator. Yes, I'm aware of Apple's weird branding for their tensor cores.

"In fact MLX does not even support ANE yet"

I didn't say otherwise. The ANE is a fantastic unit for small, power-efficient models, like extracting text from images, doing depth modelling, etc. It's not made for LLMs, or the other sorts of experimental stuff MLX is intended for. Though note that MLX's author's reason for not supporting the ANE is that it has a "closed-source" API (https://github.com/ml-explore/mlx/issues/18#issuecomment-184...), making it unsuitable for an open-source project, and given that MLX didn't want to just lean on CoreML. But anyways, the ANE is fantastically fast at what it does, while sipping juice.

In any case, the code change shown should have zero impact on the running of MLX on an iPhone 16 Pro. MLX tries to really leverage platform optimizations so maybe another bifucation is making the wrong choice.

reply
The change's effects are dependent on what each SKU reports as its Metal architecture, both as identifying string (the equivalent to running 'metal-arch' in the Mac CLI) and as generation 'gen' number. Most likely you're misinterpreting the change as not affecting the iPhone 16 Pro, where in fact it does.

The MLX folks have various rationales for not supporting the ANE (at least as of yet), but one of them is that any real support requires implementing explicit splits in the graph of computations, where ANE-suitable portions are to be dispatched to the ANE and everything else goes back to the GPUs. That's not necessarily trivial.

reply
It used to be great, but those days are long gone, see the archived docs.
reply