Hypura – A storage-tier-aware LLM inference scheduler for Apple Silicon

213 points by tatef a day ago on hackernews | 69 comments

Where does "1T parameter model" come from? I can only see models with 70B params or less mentioned in the repo.

: causal | a day ago
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...
: [OP] tatef | a day ago
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)

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.

: salynchnew | a day ago
It was written by an LLM, so... yeah.
: jeffybefffy519 | a day ago
Except this isnt using heavily quantised versions of the model thus reducing quality.

Insanity | a day ago

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 | a day ago

This is not putting any stress or wear on the NVMe, it's a pure read workload.

: [OP] tatef | a day ago
Yes, exactly this.

embedding-shape | a day ago

> 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.

Insanity | a day ago

I had assumed heat generation on the controller if it's continuously reading. But maybe it's not actually bad.

: throwway120385 | a day ago
Just pop a heatsink on it and call it good.

[OP] tatef | a day ago

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.

: embedding-shape | 23 hours ago
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.

hrmtst93837 | 23 hours ago

People talk about "SSD endurance", but enough parallel I/O on M1/M2 can make the NVMe controller choke, with very weird latncy spikes.

monksy | a day ago

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.

rubiquity | a day ago

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

circularfoyers | a day ago

Don't even need to use llama-swap anymore now that llama-server supports the same functionality.

: rubiquity | 18 hours ago
I did not know that. Thanks for sharing!

zozbot234 | a day ago

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.

baq | a day ago

Intel Optane rolling in its grave.

liuliu | a day ago

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 | a day ago
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.

speedgoose | a day ago

Is it too late for Intel to bring them back to life?

: c0balt | a day ago
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).
: walterbell | a day ago
Nvidia and SK Hynix are bringing HBF to market for $$.

0ptan3 | a day ago

pmem

moffkalast | a day ago

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.

aitchnyu | a day ago

Memristors are also missing in this AI hype even when they were around the corner 10 years back.

nullbyte | a day ago

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

anshulbasia27 | a day ago

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.

zozbot234 | a day ago

> 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 | 21 hours ago

That works for readahead but it's not good for random access. readv, aio, dispatch_io are better there.

zozbot234 | 21 hours ago

This claim is a bit apples and oranges (no pun intended!). madvise is all about providing hints to the kernel to tune the page cache and readahead (including possibly disabling readahead altogether). it's not about performing reads into private memory buffers, which is actually where the options you mentioned fit in.

: astrange | 3 hours ago
Triggering reads is also how you get pages into the page cache, so it helps to know how to do it.

EnPissant | a day ago

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.

EnPissant | a day ago

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).

amelius | a day ago

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

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

Headlines like that are misleading.

smlacy | a day ago

Yes, and with virtually zero context, which makes an enormous difference for TTFT on the MoE models.

feznyng | a day ago

Could still be useful; maybe for overnight async workloads? Tell your agent research xyz at night and wake up to a report.

maleldil | a day ago

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.

: zozbot234 | 21 hours ago
Sure, but if long-term throughput is a real limitation there's plenty of ways to address that while still not needing to keep anywhere close to all model weights in RAM (which is still the conventional approach with MoE). So the gain of a smaller memory footprint is quite real.

speedgoose | a day ago

I wonder how many minutes per token on GLM 5.

vicchenai | a day ago

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.

zozbot234 | a day ago

> 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 | a day ago
But across a sequence you still have to load most of them.

[OP] tatef | a day ago

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.

p_ing | a day ago

4K random read with a queue depth of 1 on an M1 Max is about 65MB/s.

erikcw | a day ago

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/

root_axis | a day ago

Are there any 1T parameter open source models?

zozbot234 | a day ago

Kimi 2.5?

root_axis | a day ago

Thanks, TIL.

ai-inquisitor | a day ago

That model is "open weight", not open source. We have no idea what data Moonshot trained on.

: airspresso | 10 hours ago
I think we lost that terminology war. Open source models mean open weight. There are only a couple examples of fully open source models with open data and code, and the labs are not incentivized to go that far.

vanyaland | a day ago

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.

joelthelion | 7 hours ago

How much are you going to spend on electricity though? Is this really going to be more cost-effective than just using openrouter?

: austinthetaco | 6 hours ago
There are many other reasons someone might want to run a model locally outside of cost savings, ownership of data flow and use in locations without internet to name a couple.

shubhamintech | 23 hours ago

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.

simonw | 22 hours ago

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

Imustaskforhelp | 22 hours ago

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 | 21 hours ago

The lack of a token rate metric for the kimi example is disappointing.

: zozbot234 | 20 hours ago
The latter link says they get ~1.7 tok/s which is quite impressive for a near-SOTA local model running on ordinary hardware.

[OP] tatef | 21 hours ago

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 :)

astrange | 21 hours ago

> 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.

msbhogavi | 19 hours ago

"As much memory as possible" is right for model capacity but misses bandwidth. Apple Silicon has distinct tiers: M4 Pro at 273 GB/s, M4 Max at 546 GB/s, M4 Ultra at 819 GB/s. Bandwidth determines tok/s once the model fits in memory. An M4 Max gives you 2x the decode speed of an M4 Pro on the same model.

For what Hypura does, the Max is the sweet spot. 64GB loads a 70B at Q4 with room to spare, and double the bandwidth of the Pro means generation is actually usable instead of just technically possible.

dev_tools_lab | 8 hours ago

Nice work on the scheduler. Have you benchmarked parallel inference across multiple models? Running GPT, Claude and Gemini simultaneously on the same input is where latency becomes a real constraint.

zozbot234 | 7 hours ago

GPT-OSS exists but Claude and Gemini aren't available locally, lol.

: dev_tools_lab | 5 hours ago
True, Claude and Gemini aren’t local yet — I mostly meant running all available local models in parallel.
Even with just open-source LLMs, you can see interesting differences in flagged issues when cross-validating outputs.

dangoodmanUT | 3 hours ago

With unified memory and such a strong os-hardware integration, one would hope that swap could handle this task