speculative decoding in llama.cpp: anyone testing this?

[vrambandit]

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 NVFP4.

Maybe obvious to some of you, but it clicked for me this week.

[vrambandit]

That matches what I was seeing. I don't think your conclusion is far off. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable. The logs were somehow calmer than I was.

[mangotinker73]

That matches what I was seeing. I ended up in about the same place using llama.cpp. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[thetagrid15]

Tried to reproduce it just now. The issue went away once I changed one setting and restarted the stack cleanly. The surprise was that memory pressure showed up before throughput collapsed.

[signalcraft14]

What exact setup are you running? Is this local-only or are you routing anything externally?

[quiettoken]

same here on my side.

[mlxgarden]

My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at workstation thermals.

[edgequill]

My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at workstation thermals.

Ran a quick check on my box. The issue went away once I changed one setting and restarted the stack cleanly. The surprise was that memory pressure showed up before throughput collapsed.

[latentcraft]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Phi and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

If you post your exact settings, people here can probably narrow it down pretty fast.

[visionyard23]

My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tail latency. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.