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 quant drift.

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

This is mostly for local work rather than public serving.

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 KV cache pressure.

Maybe obvious to some of you, but it clicked for me this week.

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

I tried this with Phi using vLLM on 7900 XTX. The headline result was better than I expected, but tail latency 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.

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

This isn't revolutionary, but the combination of llama.cpp and 5090 felt surprisingly usable once I tuned around KV cache pressure.

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

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

The more I work with local setups, the more it seems like token cost decides everything downstream.

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

This is mostly for local work rather than public serving.

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

After a few runs on A6000, I ended up with decent performance, but the setup only really made sense once I adjusted around tool calling.

Happy to post more details if anyone wants them.

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.

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: Kimi, TensorRT-LLM, dual 3090, with a lot of attention on reranking.

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

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

The more I work with local setups, the more it seems like MoE serving decides everything downstream.

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

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

Two paths both seem viable, but they break down differently once token cost becomes important.

Interested in tradeoffs more than fan loyalty.

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: Qwen, vLLM, 5090, with a lot of attention on prefix caching.

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. I keep promising myself this is the last rebuild.