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 long context.

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

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

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 router policy.

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

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 TensorRT-LLM on 4090. The headline result was better than I expected, but NVFP4 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.

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

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

Maybe useful as a starting point: vLLM on 4090, keeping speculative decoding in mind from the start.

Feel free to tear it apart if you see something dumb.

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

I'm trying to sanity-check my setup.

I'm mostly working with DeepSeek in vLLM on B200. The part I'm trying to understand is MoE serving.

Would appreciate practical advice rather than theory.

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

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: Mistral, llama.cpp, M3 Max, with a lot of attention on tokens/sec.

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

After a few runs on GB200, I ended up with decent performance, but the setup only really made sense once I adjusted around structured output.

Happy to post more details if anyone wants them.

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

Short version: vLLM behaved well for me, but only after I stopped ignoring token cost.

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.

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

Two paths both seem viable, but they break down differently once workstation thermals becomes important.

Interested in tradeoffs more than fan loyalty.

This is mostly for local work rather than public serving.

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

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

I tried this with Qwen using Ollama on 7900 XTX. The headline result was better than I expected, but quant drift still ended up being the limiting factor.

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