little utility you keep installed for Qwen

[batchshed]

Interesting, I hadn't tried that. I ended up in about the same place using MLX. The biggest change for me was simplifying the stack instead of adding

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

[rocmcrate26]

Need a little more detail here. Do you see the same thing with a smaller batch or shorter prompt?

[rocmriley]

Need a little more detail here. Are you testing this on B200 or something else?