odd little inference appliances: early impressions

[mangopath4]

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

One thing I keep noticing is that people underestimate KV cache pressure until they actually try to run this stuff for more than one afternoon.

Curious whether that lines up with what other people are seeing.

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

[turboscope]

Can you fill in a bit more context? Are you testing this on 7900 XTX or something else?

[vramarch]

Good point. I ended up in about the same place using Ollama. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[mintquill]

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

[bytepilot]

That matches what I was seeing. I ended up in about the same place using PyTorch. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[audiorunner74]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Kimi and compare from there. Once you know the plain path works, then add the faster path back in piece by piece.

[lablab]

That matches what I was seeing. 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.

[turbosupply83]

I'd approach it in a boring order. If you're on MLX, 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.

[mangopath4]

I tested this tonight. 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.

[latentscope35]

Tried to reproduce it just now. My result was mixed: first run looked bad, second run looked normal. The surprise was that memory pressure showed up before throughput collapsed.