driver/framework news roundup on 4090

[southpatch]

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 prefill/decode split.

I'm especially interested in mistakes worth avoiding.

I care more about stability than squeezing the last possible tokens/sec out of it.

[deepseeked]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Phi and compare from there.

[fastlab]

That matches what I was seeing. The pattern around speculative decoding 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.

[dockriver]

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.

[deepseeked]

Tried to reproduce it just now. On mini PC with Ollama, I saw basically the same behavior.

[localpatch78]

A couple questions before I guess too hard. Do you see the same thing with a smaller batch or shorter prompt?

[deepseeked]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there.

[audiopoint18]

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

[thetashed24]

I hit almost the same issue last night.

[deepseeked]

What exact setup are you running? What model, what runtime, and roughly what context size?