model fallback rules that work

[cozyarch]

This keeps coming up in conversation so I figured I'd start a thread.

It feels like the conversation has shifted hard toward inference and serving details lately, and honestly that makes sense once you start dealing with KV cache pressure.

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

This is mostly for local work rather than public serving.

[routerunner]

I don't think that's the whole story. I get why that seems right, but on 5090 the bottleneck felt elsewhere. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.

[torchkite]

I don't think that's the whole story. I get why that seems right, but on 5090 the bottleneck felt elsewhere. Not saying you're wrong, just that I woul

Good point. The pattern around long context 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.

[rocmriley]

Thanks, that helps. The pattern around VRAM headroom has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[mlxgarden]

Ran a quick check on my box. On A6000 with Ollama, I saw basically the same behavior.

[llm_wren]

That matches what I was seeing. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[dockkite69]

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

[torchshed]

What worked for me was pretty simple. On dual 3090, 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.

[fastbench]

A couple questions before I guess too hard. What model, what runtime, and roughly what context size?