JAX for inference only

[kernelarch]

Posting this in case it saves someone else a little time.

I'm helping a small team set up a local workflow. We don't need a giant stack, but we do care about speculative decoding.

I'm especially interested in mistakes worth avoiding.

[mellowstream]

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

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.

[vramport]

Interesting, I hadn't tried that. I ended up in about the same place using SGLang. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[mangopath4]

Good point. That matches what I saw on dual 3090. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[vrambandit]

Not sure I agree there. I get why that seems right, but on dual 3090 the bottleneck felt elsewhere. Could be a workload difference, but I wouldn't generalize too fast from one setup.

[cachelab]

I tested this tonight. I couldn't reproduce the slowdown until I pushed context much higher. The logs were somehow calmer than I was.

[fusedhall56]

A couple things I'd check first. On 5090, 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.

[thetapath]

What exact setup are you running? Is this local-only or are you routing anything externally?

[rasterbench]

Not sure I agree there. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model.

[dockworks23]

Appreciate the detail. That matches what I saw on A6000. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.