Hypura – A storage-tier-aware LLM inference scheduler for Apple Silicon (github.com)
177 points by tatef 10 days ago | 73 comments



simonw 10 days ago | flag as AI [–]

Suggestion for the maintainers: the comparison table currently lists some pretty old models, Qwen 2.5 14B and Mixtral 8x7B and Llama 3.3 70B.

A lot of people are reporting incredible results with the Qwen 3.5 MoE models on Apple hardware right now (streaming experts - see https://simonwillison.net/2026/Mar/24/streaming-experts/) - it would be great to get some of those models into that table.

Maybe the 1T parameter Kimi K2.5 too if you can get that to work, see https://twitter.com/seikixtc/status/2036246162936910322 and https://twitter.com/danpacary/status/2036480556045836603

tatef 10 days ago | flag as AI [–]

Thanks for sharing this! If you'd be interested in running the benchmark yourself with Hypura I'd happily merge into our stats. Otherwise will add to my todo list :)

Simon, A little offtopic but it seems that your website isn't working.

> An error occurred in the application and your page could not be served. If you are the application owner, check your logs for details. You can do this from the Heroku CLI with the command

I get this error when I go to simonwillison.net

Any random blog/link works for example though: https://simonwillison.net/2026/Mar/19/openai-acquiring-astra...

(I checked your website because I wanted to see if you had written something about trivy/litellm as well, I highly recommend checking out what has happened within litellm space if possible as I would love to read your thoughts on it)

Have a nice day simon!

Edit: now the website works but I am not sure what had gone wrong previously, (an issue from heroku maybe?) as its working now

Edit-2: after the website working, I am able to see that you have already made a post about it.

abtinf 10 days ago | flag as AI [–]

The lack of a token rate metric for the kimi example is disappointing.
zozbot234 10 days ago | flag as AI [–]

The latter link says they get ~1.7 tok/s which is quite impressive for a near-SOTA local model running on ordinary hardware.
vanyaland 10 days ago | flag as AI [–]

For a lot of local workloads, sub-1 tok/s is useless in foreground and perfectly acceptable in background. If the choice is “this crashes” vs “this finishes overnight,” that’s still a meaningful capability jump.
vicchenai 10 days ago | flag as AI [–]

the practical question is whether the read pattern is sequential enough to actually saturate nvme bandwidth or if the attention layer access pattern ends up being random enough to kill throughput. sequential reads on a decent nvme get you 5-7 GB/s, random reads drop to maybe 500 MB/s depending on queue depth.

for a 1T model youd need to stream something like 2TB of weights per forward pass at fp16. even at peak sequential thats 300+ seconds per token which is... not great for interactive use but maybe fine for batch inference where you dont care about latency.

still a cool proof of concept though. the gap between 'can run' and 'runs usefully' is where things get interesting.

tatef 10 days ago | flag as AI [–]

Yes, definitely agree. It's more of a POC than a functional use case. However, for many smaller MoE models this method can actually be useful and capable of achieving multiple tokens/sec.
zozbot234 10 days ago | flag as AI [–]

> for a 1T model youd need to stream something like 2TB of weights per forward pass

Isn't this missing the point of MoE models completely? MoE inference is sparse, you only read a small fraction of the weights per layer. You still have a problem of each individual expert-layer being quite small (a few MiBs each give or take) but those reads are large enough for the NVMe.

visarga 10 days ago | flag as AI [–]

But across a sequence you still have to load most of them.
nina57 10 days ago | flag as AI [–]

The sparse read pattern sounds clean on paper. Wait until your expert routing starts hitting the same few experts under load and you're back to sequential hotspots on one nvme while the others idle.
anvil54 10 days ago | flag as AI [–]

ran this on an M1 Pro last week and the access pattern is definitely more random than sequential in practice. queue depth stays low since there's no aggressive prefetch happening. curious what the 4K random numbers actually look like at QD1 on Apple silicon.
p_ing 10 days ago | flag as AI [–]

4K random read with a queue depth of 1 on an M1 Max is about 65MB/s.
marksully 10 days ago | flag as AI [–]

Where does "1T parameter model" come from? I can only see models with 70B params or less mentioned in the repo.
tatef 10 days ago | flag as AI [–]

I'm referencing it as being possible, however I didn't share benchmarks because candidly the performance would be so slow it would only be useful for very specific tasks over long time horizons. The more practical use cases are less flashy but capable of achieving multiple tokens/sec (ie smaller MoE models where not all experts need to be loaded in memory simultaneously)
causal 10 days ago | flag as AI [–]

Yeah title comes from nowhere in the link. No doubt it's possible but all that matters is speed and we learn nothing of that here...

The MoE point matters here ie sparse activation means you're not reading all 2TB per forward pass, but the access pattern flips from sequential to random which is exactly the worst case for NVMe. Been thinking about this a lot for agent inference workloads where you want consistent latency more than peak throughput.
baq 10 days ago | flag as AI [–]

Intel Optane rolling in its grave.
aitchnyu 10 days ago | flag as AI [–]

Memristors are also missing in this AI hype even when they were around the corner 10 years back.
liuliu 10 days ago | flag as AI [–]

Still have 4 brand new ones in my storage unit. Just in case these moments.

Joke aside (I do have them tho!), I don't think Optane is that much use (not to mention it is only 256GiB for my unit). It is useful legacy crutch if you have legacy software that is not designed to issue multiple reads / writes in parallel. If you do, it is really not faster than NVMe, especially these modern ones.

zozbot234 10 days ago | flag as AI [–]

It's not about being faster (except for small reads where latency dominates, which is actually relevant when reading a handful of expert-layers immediately after routing), it's the wearout resistance which opens up the possibility of storing KV-cache (including the "linear" KV-cache of recent Qwen, which is not append-only as it was with the pure attention model) and maybe even per-layer activations - though this has the least use given how ephemeral these are.

Is it too late for Intel to bring them back to life?
c0balt 10 days ago | flag as AI [–]

Yes, their NAND division has been sold, it is now mostly under solidigm. Maybe solidigm could bring it back, but it seems unlikely (given the previous commercial failure).

Nvidia and SK Hynix are bringing HBF to market for $$.

Wouldn't be Intel if they didn't quit halfway through on a good thing.

Still, couldn't one get a RAID 0 card with four drives to saturate a 16x lane? That's already the max one could push through PCIe anyhow.

0ptan3 10 days ago | flag as AI [–]

pmem
astrange 10 days ago | flag as AI [–]

> Consumer hardware (MacBook Pro, Mac Studio) ships with fast unified memory and NVMe storage, but limited capacity. A 32 GB M1 Max cannot naively load a 40 GB model — the OS will swap-thrash until the OOM killer intervenes.

macOS doesn't have an "OOM killer" in that sense. (It has an out of swap space killer but it's pretty weak.)

So what will happen is, either your memory wiring will fail, or else it will get really slow and panic.

Insanity 10 days ago | flag as AI [–]

This is a pretty cool project! Essentially this is like using Swap memory to extend your RAM, but in a 'smart' way so you don't overload the NVMe unnecessarily.

I do wonder in practice how the 'smarts' pan out, because putting a ton of stress on your NVMe during generation is probably not the best choice for it's longevity.

zozbot234 10 days ago | flag as AI [–]

This is not putting any stress or wear on the NVMe, it's a pure read workload.
hwolfe 10 days ago | flag as AI [–]

Right, read workloads are basically free from a wear standpoint. The concern is more about latency — NVMe reads add up when you're loading weights layer by layer, and that can noticeably slow generation versus keeping everything hot in RAM.
tatef 10 days ago | flag as AI [–]

Yes, exactly this.

> but in a 'smart' way so you don't overload the NVMe unnecessarily

"overloading NVMe"? What is that about? First time I've heard anything about it.

> because putting a ton of stress on your NVMe during generation

Really shouldn't "stress your NVMe", something is severely wrong if that's happening. I've been hammering my SSDs forever, and while write operations "hurt" the longevity of the flash cells themselves, the controller interface really shouldn't be affected by this at all, unless I'm missing something here.

tatef 10 days ago | flag as AI [–]

Hypura reads tensor weights from the GGUF file on NVMe into RAM/GPU memory pools, then compute happens entirely in RAM/GPU.

There is no writing to SSDs on inference with this architecture.


Even if there was a ton of writing, I'm not sure where NVMe even comes in the picture, write durability is about the flash cells on SSDs, nothing to do with the interface, someone correct me if I'm wrong.
max 10 days ago | flag as AI [–]

Same pattern as the old SGI NUMA systems circa 1995 — keep compute local, stream data in from slower tiers. Nothing new, but nice to see it done cleanly on consumer hardware.

People talk about "SSD endurance", but enough parallel I/O on M1/M2 can make the NVMe controller choke, with very weird latncy spikes.
Insanity 10 days ago | flag as AI [–]

I had assumed heat generation on the controller if it's continuously reading. But maybe it's not actually bad.
zozbot234 10 days ago | flag as AI [–]

It will be interesting to compare this to https://news.ycombinator.com/item?id=47476422 and https://news.ycombinator.com/item?id=47490070 . Very similar design except that this is apparently using mmap, which according to the earlier experiment incurs significant overhead.

It was written by an LLM, so... yeah.

Except this isnt using heavily quantised versions of the model thus reducing quality.
root_axis 10 days ago | flag as AI [–]

Are there any 1T parameter open source models?
zozbot234 10 days ago | flag as AI [–]

Kimi 2.5?

That model is "open weight", not open source. We have no idea what data Moonshot trained on.
root_axis 10 days ago | flag as AI [–]

Thanks, TIL.
nullbyte 10 days ago | flag as AI [–]

I am curious how the TPS compares vs default OS virtual memory paging

I wonder how many minutes per token on GLM 5.
amelius 10 days ago | flag as AI [–]

This is <1 tok/s for the 40GB model.

Come on, "Run" is not the right word. "Crawl" is.

Headlines like that are misleading.

feznyng 10 days ago | flag as AI [–]

Could still be useful; maybe for overnight async workloads? Tell your agent research xyz at night and wake up to a report.
maleldil 10 days ago | flag as AI [–]

Assuming 1 token per second and "overnight" being 12 hours, that's 43 200 tokens. I'm not sure what you can meaningfully achieve with that.
smlacy 10 days ago | flag as AI [–]

Yes, and with virtually zero context, which makes an enormous difference for TTFT on the MoE models.
monksy 10 days ago | flag as AI [–]

There needs to be something like this from Ollama. At the moment Ollama has a lot of flaws that prevent it from getting great performance. (My understanding is better GPU/CPU splits, etc). But Ollama is the only way to host an LLM and have it switch out on demand. Sigh.
zozbot234 10 days ago | flag as AI [–]

Ollama has very substandard support for mmap at present, which hurts inference with larger models. There are some recent pull requests in flight that should help address this to at least some extent https://github.com/ollama/ollama/pull/14525 https://github.com/ollama/ollama/pull/14134 https://github.com/ollama/ollama/pull/14864 but progress seems to be stalling. Their support for recent Qwen models seems to also have some bespoke incompatibilities with llama.cpp, which doesn't help matters; it's difficult to test the same model with both.
rubiquity 10 days ago | flag as AI [–]

llama.cpp and llama-swap do this better than Ollama and with far more control.

Don't even need to use llama-swap anymore now that llama-server supports the same functionality.
EnPissant 10 days ago | flag as AI [–]

You do not provide any comparison to llama.cpp with mmap.

You do not explain how any kind of predictor can work for MoE experts.

You do not explain how prediction can even be useful. I can predict the layers used in a dense model (all of them are used in order), but that doesn't help me much. It's still bottlenecked on bandwidth (hint: MoE doesn't change this).


OS paging would be significantly worse here. The kernel's page fault handler is reactive — it doesn't know you're about to read layer 47's FFN weights, so it can't prefetch. You stall on every fault, wait for the 4KB/16KB page to load, then resume. With 80 layers of dense FFN streaming, that's thousands of cold faults per token.

  What makes this approach faster is that the model's access pattern is completely deterministic during         
  inference. You know exactly which tensors are needed next because transformer layers execute sequentially. So
  you can issue large sequential reads and prefetch the next layer while the current one is computing on Metal. 
  The OS page cache can't do that — it has no concept of "layer N+1 comes after layer N."

  For MoE it's even more stark. The OS would page in all 8 experts on the first token that routes to each one,  
  then evict them under memory pressure with LRU, which has no idea that expert 3 fires 10x more often than
  expert 7. The neuron cache here is basically a domain-specific replacement policy.
EnPissant 10 days ago | flag as AI [–]

That assumes you have significant work to do between fetches (so you can prefetch while using the current data). With LLM decode you don't.
helix5 10 days ago | flag as AI [–]

Has anyone benchmarked madvise MADV_SEQUENTIAL against explicit async prefetch for this workload? My intuition is the granularity difference matters — madvise still operates on pages, not weight tensors, and the readahead heuristics might fight you on irregular access patterns.
zozbot234 10 days ago | flag as AI [–]

> The kernel's page fault handler is reactive — it doesn't know you're about to read layer 47's FFN weights, so it can't prefetch.

man 2 madvise

astrange 10 days ago | flag as AI [–]

That works for readahead but it's not good for random access. readv, aio, dispatch_io are better there.
erikcw 10 days ago | flag as AI [–]

Simon Willison wrote a good post about Dan Woods’ work on “Autoresearching Apple's "LLM in a Flash" to run Qwen 397B locally”.

[0] https://simonwillison.net/2026/Mar/18/llm-in-a-flash/

signal 10 days ago | flag as AI [–]

The framing assumes NVMe offloading is a capability gap to close, but it's really a workaround for running models too large for your hardware. Buying more RAM costs less than the engineering overhead here.

Don't post generated/AI-edited comments. HN is for conversation between humans

https://news.ycombinator.com/item?id=47340079

tatef 10 days ago | flag as AI [–]

Noted, thanks. I had LLM help positioning this message but I did the initial draft along with edits. Will keep in mind for the future.
DennisP 10 days ago | flag as AI [–]

That doesn't read like an AI-generated comment to me. He did mention he vibe-coded the project but that's not against the guidelines.
causal 10 days ago | flag as AI [–]

You need to change the title or actually include 1T parameter model content.
frikk 10 days ago | flag as AI [–]

This is interesting work, thank you for sharing. What hardware would you buy today for experimenting? Seems like the new gen of macbook pros are pretty powerful?
tatef 10 days ago | flag as AI [–]

Yes definitely. I use a M1 Max with 32gb of RAM daily and it's about on par from a performance standpoint with the new base M5 Pro 24gb. You can check the benchmarks in the repo if you're interested in seeing specific performance metrics, but investing in Apple hardware with as much memory as possible will generally get you furthest in this game.

Have you ever generated access frequency statistics for the experts in these models, something like a histogram?
Gracana 10 days ago | flag as AI [–]

ktransformers can do dynamic placement of experts and could presumably produce such a histogram, though currently its activation statistics are just a ".pt" file. https://github.com/kvcache-ai/ktransformers/blob/main/doc/en...

FWIW I never got it to work and did not dig into it much.

lostmsu 10 days ago | flag as AI [–]

Why would llama with --mmap crash?
zozbot234 10 days ago | flag as AI [–]

This doesn't surprise me all that much, mmap support gets little attention in general and interacts poorly with GPU-side inference. (And that's with it being default, you don't even really need to specify it as a CLI option.) OP has raised a discussion with the llama.cpp folks https://github.com/ggml-org/llama.cpp/discussions/20852 but little interest so far