0.4.5.dev1 were better for GLM 5.2 MXFP4 speed

[maxkongerskov]

Why have GLM 5.2 MXFP4 on my M3 Ultra 512 been slower after updating?

As of writing this I am reverting back to 0.4.5.dev1 where I had 150-200 PP and 15-25 tk generation speed.

On the latest oMLX build I have poor performance, 87 pp tk and 9 generation tk

[jundot]

Thanks for the report. A few things would help me narrow this down:

1. How did you install the latest build, DMG, homebrew, or pip? Installs without the native GLM kernels silently fall back to a much slower path (GLM-5.2-mxfp4 PP drop from 160-180 tok/s to 40 tok/s with 0.5.0.dev2 #2137), and your numbers look similar to that case. You can check with python -c "from omlx.patches.glm_moe_dsa.kernels import fast; print(fast.native_available())", if it prints False you are on the fallback.
2. Roughly what context length were you testing at, and was MTP enabled?
3. Does the slowdown survive a full reboot? I recently confirmed a case where a memory leak left the GPU degraded until the machine was rebooted (v0.5.0: engine permanently deadlocks in mlx::core::eval_impl on first ~16k-token prompt (GLM-5.2 + MTP); rc1 and rc1+#2160+#2150 are fine #2197).

[maxkongerskov]

It was updated from DMG.

The prompt was a full project review inside Claude Code. Speed was gone.

Reverting back to dev1 brought pp speed to 150 but generation is still 9 tks, down from 25 tks.

Context length was 512000