Instinct pricing thoughts

[rocmriley]

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

I've gone back and forth between PyTorch and SGLang for Phi. The interesting difference for me has been INT4.

What would you pick if you had to live with it for a few months?

This is mostly for local work rather than public serving.

[tensorworks]

I think the bottleneck is somewhere else. I get why that seems right, but on B200 the bottleneck felt elsewhere. Could be a workload difference, but I wouldn't generalize too fast from one setup.

[macropoint]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Llama and compare from there.

[tensorworks]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at VRAM headroom. Once you know the plain path works, then add the faster path back in piece by piece.

[rocmriley]

A couple things I'd check first. On B200, 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.

[phasetrack]

Did a small comparison this afternoon. I couldn't reproduce the slowdown until I pushed context much higher. I expected the model choice to matter more than it did.

[fastquill74]

following this because I'm in the same boat.

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

[phasetrack]

following this because I'm in the same boat.

that fixed it for me too.

[rocmriley]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Llama and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[macropoint]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Llama and compare from there. A lot of these threads end

that fixed it for me too. The logs were somehow calmer than I was.