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 VRAM headroom.

Simple and boring is a valid answer.

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.

Has anyone else run into this?

My current setup is TensorRT-LLM with Threadripper box, and I keep running into VRAM headroom.

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

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

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 vLLM and 4090.

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

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

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

I figured I'd ask here before changing everything.

I'm mostly working with Mistral in Ollama on 4090. The part I'm trying to understand is prefill/decode split.

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

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

I'm comparing two approaches here: keep it simple on dual 3090, or squeeze more out of the stack and deal with the complexity.

Trying to avoid rebuilding the whole thing twice.

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

If I had to summarize the lesson so far, it would be that clean setup and observability beat clever guesses.

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 token cost.

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

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.

I'm not trying to build a perfect architecture diagram. I just want a sane starting point around LM Studio and 7900 XTX.

I'm especially interested in mistakes worth avoiding.

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

I tried this with Mistral using Ollama on 7900 XTX. 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.

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

I'm trying to sanity-check my setup.

My current setup is PyTorch with dual 3090, and I keep running into HBM constraints.

If you've tested something similar, what did you keep and what did you stop bothering with? At this point the real bottleneck is my patience.

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

I've gone back and forth between LM Studio and PyTorch for Gemma. The interesting difference for me has been MoE serving.

Interested in tradeoffs more than fan loyalty.

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