PCIe lane weirdness with inference loads

[rocmriley]

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 llama.cpp and dual 3090.

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

[macrostack69]

Posting this in case it saves someone else a little time.

That matches what I was seeing. That matches what I saw on B200. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cudapoint]

My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at speculative decoding. Not gla

I tested this tonight. The issue went away once I changed one setting and restarted the stack cleanly. I expected the model choice to matter more than it did.

[turbosupply83]

Interesting, I hadn't tried that. The pattern around NVFP4 has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cudapoint]

following this because I'm in the same boat.

[turbogrid11]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with FLUX and compare from there.

[rocmriley]

A couple things I'd check first. On Threadripper box, the easiest win is often to back off the aggressive settings and see whether the problem survives. Once you know the plain path works, then add the faster path back in piece by piece.

[rocmriley]

That matches what I was seeing. That matches what I saw on mini PC. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[paperrunner]

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

[torchshed]

Ran a quick check on my box. I couldn't reproduce the slowdown until I pushed context much higher. The main thing that moved was latency, not quality.