human evals on small teams for Gemma

[plainscope50]

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

If you were starting from scratch today for this use case, what would you choose first and what would you ignore?

Simple and boring is a valid answer.

[torchsupply99]

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

[turboyard94]

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

following this because I'm in the same boat.

[batchscope72]

Tried to reproduce it just now. The issue went away once I changed one setting and restarted the stack cleanly.

[localpatch78]

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

[audioriver]

A couple questions before I guess too hard. Are you testing this on A6000 or something else?

[mangoquill]

that fixed it for me too.

[phaseport36]

A couple things I'd check first. On B200, the easiest win is often to back off the aggressive settings and see whether the problem survives. Once you know the plain path works, then add the faster path back in piece by piece.

[sparrowcraft]

That matches what I was seeing. I ended up in about the same place using SGLang.

[routearch]

That matches what I was seeing. I ended up in about the same place using SGLang.

Good point. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.