Why not just just have most of it in wasm? There are wasm runtimes that don’t require js, and I’m sure Google can adapt their engine. Then they could integrate the other side with rust

You can always compile to wasm, or ship a wasm decoder. You can even ship a JS decoder, but speed and resource consumption are important, ruling both out.

That would likely require copying the image data in and out of the wasm instance's memory, while a pure Rust solution could have safety with less overhead.

I was enticed by the promise of it being barebones but I still found it too bloated

So I cut out a lot of functionality in my own fork

https://github.com/Pyrolistical/mi

I took “expected to customize basically everything” to heart

But you kind of do.

For a untype language like js array, since it can be empty, you have to either always check the length, the item returned, or have a precondition to know the array is not empty.

All three of those cases is either code or context you’re holding in your head.

All that stuff is equivalent to a type system

Funny, I solved this problem by having Claude write a pi extension

https://github.com/Pyrolistical/pi-notify

Now my pi agent setups its own trigger to notify itself when a background process is done

Right! It’s not really rocket science. I’ve been thinking of building a general “wait” command that just blocks until some event occurs. Still more “active” waiting, but as far as I understand the efficiency problem comes from polling.

Until there is a machine that is natively functional, there is always going to an incentive to go lower level for more performance.

Even hardware (GPUs) that functional language could trivially exploit, it’s still higher performance to write low level code and manages all the memory imperatively

Ask for analogy in terms of a thing you are an expert in.

ie. i am an expert at zig, explain this c++ in terms of zig

Soooo anytime now to bring it back right??

I mean, they haven't announced the non-Pro iPhone 18 yet, so keep the hope alive!

No way you can get that much each without extremely quants

For a single r9700 you have 637 GB/s and for qwen 3.8 27b q4_k_xl the maximum tg/s is 33 before mtp

Now if you meant 4xr9700 tensor parallelism with mtp, 80 tg/s starts to make sense

You can get those numbers with https://codeberg.org/ggz14/radiance-vllm-mxfp4. I also can get it on a single R9700 but the 75 ~ 80 t/s is only peak acceptance of very predictable tokens like coding or json, and averages lower for prose. It's still much faster than regular llama.cpp.

Can confirm. I've a single R9700 and have maxed out at 45 tokens/second on llama.cpp with Q4 Qwen 3.6 27B (with MTP)

Q4 isn't an extreme quant, and I average 75 toks/s on code, 45 tok/s on prose with MTP.

I run it locally at q4_k_xl on a r9700 with kv cache bf16 and while it thinks a lot, it’s still fast enough to do the task.

This model had its knowledge replaced with reasoning ability. The chain of thought what makes this reasoning effective.

So this is why you need to let it think and don’t quantize the kv cache.

Or at least use a modern llama.cpp with KV activation rotation.

Nic needs pcie5 x16 slot, so on am5, take out your gpu