CPU-only fallback performance for Gemma

[batchsize]

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 tokens/sec decides everything downstream.

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

This is mostly for local work rather than public serving.

[localpatch78]

That matches what I was seeing. The pattern around INT4 has been pretty similar for me. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

I'm happy to test one more variant if you want a second data point.

[fastlab]

That matches what I was seeing. I ended up in about the same place using TensorRT-LLM. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[routearch]

Interesting, I hadn't tried that. That matches what I saw on dual 3090. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[torchsmith25]

following this because I'm in the same boat. The logs were somehow calmer than I was.

[thetapath]

following this because I'm in the same boat. The logs were somehow calmer than I was.

worth checking the driver first.

[dockriver]

I'd approach it in a boring order. On mini PC, the easiest win is often to back off the aggressive settings and see whether the problem survives. Not glamorous advice, but repeatable baselines save a lot of time.

[vectorfield]

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at batch size. Once you know the plain path works, then add the faster path back in piece by piece.

[northgrid33]

That matches what I was seeing. The pattern around NVFP4 has been pretty similar for me.

[mangohouse]

Can you fill in a bit more context? Are you testing this on 7900 XTX or something else?