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 llama.cpp and GB200.

Simple and boring is a valid answer. At this point the real bottleneck is my patience.

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

Maybe useful as a starting point: Ollama on 4090, keeping tail latency in mind from the start.

Maybe useful as a starting point for newer folks.

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

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

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

This isn't revolutionary, but the combination of LM Studio and 5090 felt surprisingly usable once I tuned around tool calling.

Maybe useful as a starting point for newer folks.

Wanted to compare notes with people doing similar things.

I'm mostly working with Qwen in llama.cpp on 5090. The part I'm trying to understand is NVFP4.

Would appreciate practical advice rather than theory.

Has anyone else run into this?

I'm mostly working with Kimi in Ollama on 4090. The part I'm trying to understand is speculative decoding.

If you've tested something similar, what did you keep and what did you stop bothering with?

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

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

I tried this with Kimi using MLX on A6000. The headline result was better than I expected, but speculative decoding still ended up being the limiting factor.

Curious if anyone else landed in the same place.

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

I tried this with Phi using MLX on A6000. The headline result was better than I expected, but NVFP4 still ended up being the limiting factor.

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

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 MoE serving.

Simple and boring is a valid answer.

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 batch size.

I'm especially interested in mistakes worth avoiding.

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

One thing I keep noticing is that people underestimate tail latency until they actually try to run this stuff for more than one afternoon.

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

This is mostly for local work rather than public serving.

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