quant choices for 70B local with PyTorch

[benchplease]

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

I've gone back and forth between PyTorch and PyTorch for Qwen. The interesting difference for me has been speculative decoding.

Trying to avoid rebuilding the whole thing twice.

[bytehall]

Need a little more detail here. Is this local-only or are you routing anything externally?

[vrambandit]

Did a small comparison this afternoon. 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.

[fusedtinker]

Did a small comparison this afternoon. The issue went away once I changed one setting and restarted the stack cleanly. I expected the model choice to

That matches what I was seeing. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

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

[bytehall]

Not sure I agree there. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model. Worth testing the simpler explanation before adopting the more interesting one.

[cacheplane98]

following this because I'm in the same boat.

[turbosupply83]

following this because I'm in the same boat.

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at router policy. Not glamorous advice, but repeatable baselines save a lot of time.

[benchplease]

I don't think that's the whole story. I've seen people blame HBM constraints when the real issue was much more mundane.

[northquill]

Thanks, that helps. 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.

[phasestream53]

What worked for me was pretty simple. On B200, 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.