Closed chimezie closed 3 months ago
The system has > 100GB free at the time the code is run
cc @SunMarc
Hi @chimezie, does this happen only with Mixtral-8x7B or with all the models ? From the traceback, the memory was completely used: MPS backend out of memory (MPS allocated: 163.01 GB, other allocations: 384.00 KB, max allowed: 163.20 GB)
This seems to happen only with Mixtral-8x7B. I was able to load Llama 3 8, Qwen1.5-14B, and internistai/base-7b-v0.2 for example, without any issue
Mixtral-8x7B is a very big model with around 100GB but you should be able to load the model since you have over 160GB. At which checkpoint does the loading fail ? Near the end ? You can track the memory consumption using the activity monitor on your mac.
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
System Info
Information
Tasks
no_trainer
script in theexamples
folder of thetransformers
repo (such asrun_no_trainer_glue.py
)Reproduction
Which results in:
Expected behavior
Should return a transformer model without error.