best local stack on mac right now

[labnest]

Thanks, that helps. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.

[vrambandit]

I think the bottleneck is somewhere else. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model.

Thanks, that helps. The pattern around KV cache pressure has been pretty similar for me. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[vrambandit]

same here on my side.

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Qwen and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[amberkite]

Thanks, that helps. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimizat

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Mistral and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[quietarch]

I tested this tonight. My result was mixed: first run looked bad, second run looked normal.

[mellowstream]

That matches what I was seeing. I ended up in about the same place using MLX. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[sanescope]

I tested this tonight. My result was mixed: first run looked bad, second run looked normal.

that fixed it for me too.

[circuitwire58]

What worked for me was pretty simple. If you're on SGLang, I'd confirm the version first and then reduce variables until the behavior is obvious. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[benchplease]

This keeps coming up in conversation so I figured I'd start a thread.

It feels like the conversation has shifted hard toward inference and serving details lately, and honestly that makes sense once you start dealing with prefill/decode split.

Curious whether that lines up with what other people are seeing.

[dockridge]

Tried to reproduce it just now. My result was mixed: first run looked bad, second run looked normal.