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

Just sharing a setup that ended up working reasonably well: Phi, LM Studio, 5090, with a lot of attention on long context.

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

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

I'm comparing two approaches here: keep it simple on Threadripper box, or squeeze more out of the stack and deal with the complexity.

Trying to avoid rebuilding the whole thing twice.

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

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 token cost.

Would rather learn from your scars than collect my own all at once.

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

Short version: TensorRT-LLM behaved well for me, but only after I stopped ignoring HBM constraints.

Curious if anyone else landed in the same place.

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 SGLang 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.

I figured I'd ask here before changing everything.

I'm mostly working with Phi in TensorRT-LLM on Threadripper box. The part I'm trying to understand is quant drift.

Would appreciate practical advice rather than theory.

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

I figured I'd ask here before changing everything.

I'm mostly working with FLUX in PyTorch on A6000. The part I'm trying to understand is HBM constraints.

Would appreciate practical advice rather than theory.

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

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

I tried this with DeepSeek using llama.cpp on A6000. The headline result was better than I expected, but long context still ended up being the limiting factor.

Happy to post more details if anyone wants them.

This is mostly for local work rather than public serving.

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

I tried this with Kimi using MLX on 4090. The headline result was better than I expected, but KV cache pressure still ended up being the limiting factor.

Curious if anyone else landed in the same place.

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

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

If you were starting from scratch today for this use case, what would you choose first and what would you ignore?

I'm especially interested in mistakes worth avoiding.

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