on-device inference worth the effort

[cozysmith]

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 prefill/decode split decides everything downstream.

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

[sparrowshed]

Tried to reproduce it just now. My result was mixed: first run looked bad, second run looked normal.

[fastquill74]

Thanks, that helps. That matches what I saw on mini PC. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.

[amberrunner]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Mistral and compare from there. Not glamorous advice, but repeatable baselines save a lot of time.

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.

[macrotinker]

I think the bottleneck is somewhere else. I've seen people blame tokens/sec when the real issue was much more mundane. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.

If you post your exact settings, people here can probably narrow it down pretty fast.

[marbleyard32]

I think the bottleneck is somewhere else. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model. Could be a workload difference, but I wouldn't generalize too fast from one setup.

[turbosupply83]

My first guess would be this: If you're on TensorRT-LLM, I'd confirm the version first and then reduce variables until the behavior is obvious. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[cachelab]

Need a little more detail here. Are you testing this on 7900 XTX or something else?

[fastquill74]

A couple questions before I guess too hard. Is this local-only or are you routing anything externally?

[localcraft17]

Need a little more detail here. Are you testing this on A6000 or something else?