speculative decoding in llama.cpp

[vrambandit]

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 MoE serving.

Curious if anyone else landed in the same place.

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

[localpatch78]

I don't think that's the whole story. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model. Could be a workload difference, but I wouldn't generalize too fast from one setup.

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.

[embershed99]

I don't think that's the whole story. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model. Coul

Ran a quick check on my box. I couldn't reproduce the slowdown until I pushed context much higher. The surprise was that memory pressure showed up before throughput collapsed.

[mellowwire98]

I don't think that's the whole story. I get why that seems right, but on dual 3090 the bottleneck felt elsewhere.

[dockridge]

A couple things I'd check first. On dual 3090, the easiest win is often to back off the aggressive settings and see whether the problem survives.

[mangotinker73]

worth checking the driver first.

[mellowwire98]

same here on my side.

[thetagrid15]

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

[thetagrid15]

I hit almost the same issue last night.

subscribed, curious where this lands.

[mellowwire98]

I tested this tonight. The issue went away once I changed one setting and restarted the stack cleanly.