speculative decoding in llama.cpp

[mangoquill]

Did a bit of testing and thought I'd post the result here.

After a few runs on GB200, I ended up with decent performance, but the setup only really made sense once I adjusted around INT4.

Curious if anyone else landed in the same place.

For context, I'm not chasing leaderboard numbers. I mostly want the system to feel consistent day to day.

[embershed99]

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

[macroforge]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Mixtral and compare from there. Once you know the plain path works, then add the faster path back in piece by piece.

I'm happy to test one more variant if you want a second data point.

[kernelgrid]

I don't think that's the whole story. The reported numbers can look compelling, but in ordinary usage I didn't see the same benefit. Worth testing the simpler explanation before adopting the more interesting one.

[bytesupply76]

What worked for me was pretty simple. On 7900 XTX, the easiest win is often to back off the aggressive settings and see whether the problem survives. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[oxidestream]

following this because I'm in the same boat.

[dockridge]

My first guess would be this: On Threadripper box, the easiest win is often to back off the aggressive settings and see whether the problem survives.

[sparrowworks88]

Thanks, that helps. I don't think your conclusion is far off.

[mangotinker73]

Did a small comparison this afternoon. On B200 with vLLM, I saw basically the same behavior. The surprise was that memory pressure showed up before throughput collapsed.

[embershed99]

My first guess would be this: If you're on LM Studio, 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.