RTX 5090 local inference on M3 Max

[ggufwire22]

What worked for me was pretty simple. If you're on MLX, 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.

[rasterhall]

I tested this tonight. On Threadripper box with llama.cpp, I saw basically the same behavior. The surprise was that memory pressure showed up before throughput collapsed.

[thetarunner]

same here on my side.

[fusedhall]

Tried to reproduce it just now. 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.

[torchshed]

Good point. I ended up in about the same place using LM Studio. The biggest change for me was simplifying the stack instead of adding one more optimization layer. I keep promising myself this is the last rebuild.

[dockkite69]

following this because I'm in the same boat.

[signaltrack44]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Kimi and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[thetarunner]

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

Two paths both seem viable, but they break down differently once MoE serving becomes important.

Interested in tradeoffs more than fan loyalty.

This is mostly for local work rather than public serving.

If I had to summarize the lesson so far, it would be that clean setup and observability beat clever guesses.

[batchwire]

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tail latency.