# PSA for M5Ultra owners running LLMs: set your prefill step to 8192 — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/50098199f21c0ccc

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-25 11:10 ingested 2026-09-26 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wptimb/psa_for_m5ultra_owners_running_llms_set_your/)

## retained text



```
Prefill step size affects both PP performance and your drafter fetching logits for MTP (affects dflash as well, mlx-vlm needs to be patched slightly to support chunked prefill for dflash).  
It also needs to be reasonably large to be able to fill all your cores but not overfill - otherwise it will require more dispatches.  
In my tests I observe large gains up to 8k, e.g.:

GLM-flash-4bit with MTP --prefill-step-size 8192 on raw mlx-vlm:  
`Trial 1 (32768 prompt tokens): prompt_tps=1056.033, generation_tps=72.722, total_time=38.082`  
`Trial 2 (65536 prompt tokens): prompt_tps=919.958, generation_tps=73.671, total_time=78.203`  
`Trial 3 (131072 prompt tokens): prompt_tps=735.545, generation_tps=71.067, total_time=185.435`

GLM-flash-4bit with MTP --prefill-step-size 2048:  
`Trial 1 (32768 prompt tokens): prompt_tps=860.489, generation_tps=50.011, total_time=48.339`  
`Trial 2 (65536 prompt tokens): prompt_tps=785.604, generation_tps=51.245, total_time=93.425`  
`Trial 3 (131072 prompt tokens): prompt_tps=623.588, generation_tps=50.843, total_time=220.288`

omlx with MTP (total time is skewed as it's 128TG vs 512 above):  
`pp32768/tg128` `44136.5` `17.19` `742.4 tok/s` `58.6 tok/s` `46.353s` `709.7 tok/s` `176.66 GB`  
`pp65536/tg128` `86749.1` `21.21` `755.5 tok/s` `47.5 tok/s` `89.504s` `733.6 tok/s` `177.15 GB`  
`pp131072/tg128` `178622.6` `19.01` `733.8 tok/s` `53.0 tok/s` `181.156s` `724.2 tok/s` `178.45 GB`

note that some engines (like omlx) support adaptive step size, e.g. the prefill speeds I observed for qwen3.8-flash-next on omlx even though it's starting from 2048 matches 8k performance from raw mlx-vlm at 64k context and above and even works 10% better on smaller context, but as you can see it's not always the case.
```



document fetched 2026-09-26 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wptimb/psa_for_m5ultra_owners_running_llms_set_your
