distillation that keeps useful behavior with llama.cpp

[emberpeak]

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

I'm comparing two approaches here: keep it simple on B200, or squeeze more out of the stack and deal with the complexity.

Interested in tradeoffs more than fan loyalty.

[fusedriver]

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

[quietstack]

Interesting, I hadn't tried that. The pattern around long context has been pretty similar for me. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable. At this point the real bottleneck is my patience.

[quiettoken]

Interesting, I hadn't tried that. The pattern around long context has been pretty similar for me. I got better results once I stopped chasing one metr

I don't think that's the whole story. I've seen people blame embeddings when the real issue was much more mundane. Worth testing the simpler explanation before adopting the more interesting one.

[dockworks23]

My first guess would be this: On 4090, 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.

[latentscope35]

Thanks, that helps. 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.

[bytegarden73]

What worked for me was pretty simple. Before rebuilding anything, I'd test one clean baseline with FLUX and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[vramridge]

What worked for me was pretty simple. If you're on TensorRT-LLM, I'd confirm the version first and then reduce variables until the behavior is obvious.

[visionyard23]

What worked for me was pretty simple. If you're on TensorRT-LLM, I'd confirm the version first and then reduce variables until the behavior is obvious

Tried to reproduce it just now. On GB200 with TensorRT-LLM, I saw basically the same behavior. The surprise was that memory pressure showed up before throughput collapsed.

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

[quiettoken]

Interesting, I hadn't tried that. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.