Did a bit of testing and thought I'd post the result here.

I tried this with Phi using LM Studio on Threadripper box. The headline result was better than I expected, but consumer GPUs still ended up being the limiting factor.

Happy to post more details if anyone wants them.

My use case is pretty ordinary: local chat, some document work, and occasional batch jobs.

Did a bit of testing and thought I'd post the result here.

Short version: SGLang behaved well for me, but only after I stopped ignoring prefix caching.

Nothing scientific here, just practical notes from a real setup.

I care more about stability than squeezing the last possible tokens/sec out of it.

Thought I'd post a small recipe because a lot of threads ask similar things.

Maybe useful as a starting point: MLX on 5090, keeping tool calling in mind from the start.

If someone wants the exact settings I can clean them up and post them.

Did a bit of testing and thought I'd post the result here.

I tried this with Phi using MLX on 4090. The headline result was better than I expected, but speculative decoding still ended up being the limiting factor.

Nothing scientific here, just practical notes from a real setup.

My use case is pretty ordinary: local chat, some document work, and occasional batch jobs.

The part that surprised me was how much the boring details mattered more than the model choice.

Posting this in case it saves someone else a little time.

I'm helping a small team set up a local workflow. We don't need a giant stack, but we do care about tail latency.

I'm especially interested in mistakes worth avoiding.

I care more about stability than squeezing the last possible tokens/sec out of it.

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 HBM constraints.

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

For context, I'm not chasing leaderboard numbers. I mostly want the system to feel consistent day to day.

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

The more I work with local setups, the more it seems like MoE serving decides everything downstream.

Wondering if this is just where the whole space is heading now.

Wanted to compare notes with people doing similar things.

I'm mostly working with DeepSeek in LM Studio on 5090. The part I'm trying to understand is NVFP4.

Not looking for the perfect answer, mostly what worked in real use.

Posting this in case it saves someone else a little time.

I'm not trying to build a perfect architecture diagram. I just want a sane starting point around PyTorch and A6000.

Simple and boring is a valid answer.

I expected the biggest difference to come from the model, but the stack around it mattered at least as much.

Spent some time on this today and wanted to compare notes.

I've gone back and forth between llama.cpp and llama.cpp for DeepSeek. The interesting difference for me has been quant drift.

What would you pick if you had to live with it for a few months?