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 NVFP4.

Simple and boring is a valid answer.

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

I've gone back and forth between Ollama and Ollama for DeepSeek. The interesting difference for me has been tokens/sec.

Trying to avoid rebuilding the whole thing twice.

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

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.

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

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.

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

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 consumer GPUs.

I'm especially interested in mistakes worth avoiding.

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

Short version: SGLang behaved well for me, but only after I stopped ignoring VRAM headroom.

Curious if anyone else landed in the same place.

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

I've gone back and forth between PyTorch and PyTorch for Qwen. The interesting difference for me has been speculative decoding.

Trying to avoid rebuilding the whole thing twice.

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 Ollama and 7900 XTX.

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

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 prefix caching.

Simple and boring is a valid answer.

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 Mixtral using Ollama on M3 Max. The headline result was better than I expected, but prefill/decode split still ended up being the limiting factor.

Happy to post more details if anyone wants them.

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