dataset format recommendations: early impressions

[batchshed]

A couple things I'd check first. If you're on PyTorch, I'd confirm the version first and then reduce variables until the behavior is obvious.

[routecraft]

That matches what I was seeing. The pattern around speculative decoding has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cudapoint]

I don't think that's the whole story. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.

[quantghost]

Can you fill in a bit more context? What model, what runtime, and roughly what context size?

[thetaarch43]

A couple questions before I guess too hard. Is this local-only or are you routing anything externally?

[batchember]

Good point. I don't think your conclusion is far off.