speculative decoding in llama.cpp

[vrambandit]

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

The more I work with local setups, the more it seems like consumer GPUs decides everything downstream.

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

For context, I'm not chasing leaderboard numbers. I mostly want the system to feel consistent day to day. The logs were somehow calmer than I was.

[embershed99]

I'd start with the simpler stack.

[vrambandit]

Tried to reproduce it just now. On B200 with vLLM, I saw basically the same behavior. The surprise was that memory pressure showed up before throughput collapsed.

[vramridge]

Tried to reproduce it just now. On B200 with vLLM, I saw basically the same behavior. The surprise was that memory pressure showed up before throughpu

What worked for me was pretty simple. On M3 Max, the easiest win is often to back off the aggressive settings and see whether the problem survives.

[macroforge]

My first guess would be this: On GB200, the easiest win is often to back off the aggressive settings and see whether the problem survives. Not glamorous advice, but repeatable baselines save a lot of time.

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

[syshouse]

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

[macroforge]

Thanks, that helps. That matches what I saw on GB200. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[nanopatch]

Appreciate the detail. I don't think your conclusion is far off.

[vrampilot]

My first guess would be this: If you're on vLLM, I'd confirm the version first and then reduce variables until the behavior is obvious. Once you know the plain path works, then add the faster path back in piece by piece.

[mangotinker73]

that fixed it for me too.