speculative decoding in llama.cpp

[latentcraft]

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

I'm comparing two approaches here: keep it simple on dual 3090, or squeeze more out of the stack and deal with the complexity.

Interested in tradeoffs more than fan loyalty.

My use case is pretty ordinary: local chat, some document work, and occasional batch jobs.

[dockworks23]

Ran a quick check on my box. On A6000 with PyTorch, I saw basically the same behavior. I expected the model choice to matter more than it did.

[mangotinker73]

Ran a quick check on my box. On A6000 with PyTorch, I saw basically the same behavior. I expected the model choice to matter more than it did.

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.

[vrambandit]

Interesting, I hadn't tried that. That matches what I saw on B200. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[thetagrid15]

Maybe, but I had the opposite result. 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.

[macroforge]

Maybe, but I had the opposite result. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model. Wort

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

[northcraft90]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at structured output. Not glamorous advice, but repeatable baselines save a lot of time.

[plaingarden18]

Interesting, I hadn't tried that. I don't think your conclusion is far off.

[dockridge]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with FLUX and compare from there.