inference/research
verifiedhigh · 0.85updated 2026-10-11

mlx-lm issue #1047: Kimi K2.5 KV cache OOM crash on Mac Studio M3 Ultra 512GB under long prefill

repo

Fields

id
sources/src-github-mlx-lm-issue-1047
Publisher
ml-explore / community
Kind
repo
Published
2026-03-23
Accessed
2026-10-11
notes
Detailed trace of mlx_lm.server on Mac Studio M3 Ultra (512 GB unified memory) loading 418 GB model and running long prefill (116k tokens) until Metal command buffer crashes with Insufficient Memory (00000008:***).

Notes

Detailed trace of mlx_lm.server on Mac Studio M3 Ultra (512 GB unified memory) loading 418 GB model and running long prefill (116k tokens) until Metal command buffer crashes with Insufficient Memory (00000008:***).

Cross-references

This record references no other records.