speculative decoding in llama.cpp with SGLang

[embershed99]

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

I've gone back and forth between TensorRT-LLM and vLLM for FLUX. The interesting difference for me has been INT4.

Interested in tradeoffs more than fan loyalty.

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

[pixelcrate82]

What worked for me was pretty simple. If you're on PyTorch, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.

[batchcraft]

What exact setup are you running? Are you testing this on A6000 or something else?

[macroforge]

Interesting, I hadn't tried that. That matches what I saw on 4090. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

There are a few different conversations hiding inside this thread, so being explicit about the goal helps. I keep promising myself this is the last rebuild.

[dockridge]

Tried to reproduce it just now. My result was mixed: first run looked bad, second run looked normal.

[mangotinker73]

Ran a quick check on my box. The issue went away once I changed one setting and restarted the stack cleanly. I expected the model choice to matter more than it did.

[mellowwire98]

That matches what I was seeing. I don't think your conclusion is far off.

[quantsupply]

Maybe, but I had the opposite result. I've seen people blame prefix caching when the real issue was much more mundane.

[pixelcrate82]

What worked for me was pretty simple. If you're on PyTorch, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.

[northcraft90]

Ran a quick check on my box. On B200 with MLX, I saw basically the same behavior. The main thing that moved was latency, not quality.