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 INT4.

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.

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 continuous batching.

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

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

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.

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

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: DeepSeek, LM Studio, B200, with a lot of attention on continuous batching.

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

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

Maybe I'm missing something obvious.

I'm trying to decide whether to keep the current stack or change course. Main question is how people here are handling HBM constraints.

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

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.

Just sharing a setup that ended up working reasonably well: Phi, TensorRT-LLM, B200, with a lot of attention on small local stacks.

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.

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: DeepSeek, PyTorch, 4090, with a lot of attention on prefill/decode split.

Maybe useful as a starting point for newer folks.

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.

Short version: MLX behaved well for me, but only after I stopped ignoring prefill/decode split.

Happy to post more details if anyone wants them.

This is mostly for local work rather than public serving.

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?

Simple and boring is a valid answer.

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

I tried this with DeepSeek using SGLang on 7900 XTX. The headline result was better than I expected, but VRAM headroom still ended up being the limiting factor.

Curious if anyone else landed in the same place.

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