prefix caching in llama.cpp with PyTorch

[vrambandit]

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?

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.

[thetagrid15]

Thanks, that helps. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[mellowwire98]

following this because I'm in the same boat.

[vrambandit]

following this because I'm in the same boat.

My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at NVFP4. Not glamorous advice, but repeatable baselines save a lot of time.

[latentcraft]

What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at consumer GPUs. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[bytesupply76]

that fixed it for me too.

[vrambandit]

I'd start with the simpler stack. The logs were somehow calmer than I was.

[pixelcrate82]

Good point. That matches what I saw on 7900 XTX. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[macroforge]

What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at INT4. Not glamorous advice, but repeatable baselines save a lot of time.

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

[thetagrid15]

What worked for me was pretty simple. Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there. Not glamorous advice, but repeatable baselines save a lot of time.