domain tune without wrecking base behavior with PyTorch

[benchplease]

Spent some time on this today and wanted to compare notes.

Two paths both seem viable, but they break down differently once continuous batching becomes important.

Trying to avoid rebuilding the whole thing twice.

[fusedhall56]

Tried to reproduce it just now. My result was mixed: first run looked bad, second run looked normal. I expected the model choice to matter more than it did.

[kernelghost]

Need a little more detail here. What model, what runtime, and roughly what context size?

[deltastation]

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

[cacheline]

that fixed it for me too.

[mlxgarden]

Good point. The pattern around quant drift has been pretty similar for me.

[kernelghost]

Good point. I ended up in about the same place using llama.cpp. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.