DeepSeek V4.1 impressions for Llama

[rocmhouse90]

that fixed it for me too.

Thanks, that helps. I ended up in about the same place using PyTorch. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[routecraft]

I don't think that's the whole story. I get why that seems right, but on dual 3090 the bottleneck felt elsewhere. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.

[routecraft]

following this because I'm in the same boat.

[routecraft]

that fixed it for me too.

[mangofield33]

following this because I'm in the same boat.

Interesting, I hadn't tried that. That matches what I saw on A6000. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cozysmith]

I don't think that's the whole story. I get why that seems right, but on dual 3090 the bottleneck felt elsewhere. Not saying you're wrong, just that I

I don't think that's the whole story. I get why that seems right, but on 5090 the bottleneck felt elsewhere.

[bytesupply76]

Ran a quick check on my box. My result was mixed: first run looked bad, second run looked normal. I expected the model choice to matter more than it did.

[routecraft]

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at prefix caching. Once you know the plain path works, then add the faster path back in piece by piece.

[routecraft]

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 small local stacks.

Simple and boring is a valid answer.

My use case is pretty ordinary: local chat, some document work, and occasional batch jobs.

[routecraft]

that fixed it for me too.

Tried to reproduce it just now. The issue went away once I changed one setting and restarted the stack cleanly. The main thing that moved was latency, not quality.