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

My first guess would be this: On M3 Max, the easiest win is often to back off the aggressive settings and see whether the problem survives.

I'd approach it in a boring order. If you're on TensorRT-LLM, I'd confirm the version first and then reduce variables until the behavior is obvious.

Did a small comparison this afternoon. I couldn't reproduce the slowdown until I pushed context much higher. The surprise was that memory pressure showed up before throughput collapsed.

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

I've gone back and forth between SGLang and llama.cpp for Llama. The interesting difference for me has been speculative decoding.

What would you pick if you had to live with it for a few months?

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

I hit almost the same issue last night.

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. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

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

The more I work with local setups, the more it seems like structured output decides everything downstream.

Wondering if this is just where the whole space is heading now.

I'd start with the simpler stack.