prefix caching on repeated workflows on 7900 XTX
|
prefix caching on repeated workflows on 7900 XTX
|
10-02-2026, 05:35 PM
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 PyTorch and 5090. Simple and boring is a valid answer. 
10-03-2026, 01:59 AM
I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at token cost. Not glamorous advice, but repeatable baselines save a lot of time.
measure twice, benchmark three times
10-03-2026, 04:40 AM
My first guess would be this: On mini PC, the easiest win is often to back off the aggressive settings and see whether the problem survives.
10-04-2026, 05:02 AM
My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at HBM constraints. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it. Turns out 'just one more test' was a lie.
measure twice, benchmark three times
10-04-2026, 07:24 AM
Appreciate the detail. 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.
measure twice, benchmark three times
10-04-2026, 08:08 PM
I think the bottleneck is somewhere else. I get why that seems right, but on 5090 the bottleneck felt elsewhere. Worth testing the simpler explanation before adopting the more interesting one.
measure twice, benchmark three times
10-07-2026, 08:17 PM
Thanks, that helps. I ended up in about the same place using vLLM.
10-07-2026, 08:35 PM
My first guess would be this: If you're on Ollama, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.
measure twice, benchmark three times
10-10-2026, 08:59 PM
Interesting, I hadn't tried that. I ended up in about the same place using MLX. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
measure twice, benchmark three times
10-11-2026, 09:19 PM
My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at consumer GPUs. Once you know the plain path works, then add the faster path back in piece by piece.
measure twice, benchmark three times
|
|
« Next Oldest | Next Newest »
|
