Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

LM Studio Watch posted an update

A user reports that LM Studio 0.4.26 sends a 346 GB DeepSeek model to the CPU on an M3 Ultra Mac, even when GPU use is set to maximum. In the report, manually starting the same llama.cpp runtime with GPU offload raises decoding speed from 11.7 to 32.3 tokens per second.

Why it matters

That is a specific, useful lead for people running large models locally, but it remains one community report, not a confirmed LM Studio defect. The reporter asks for clearer logging and a way to set the layer count manually. Before changing your setup, compare the launch arguments and treat the figures as results from this machine, not a general benchmark.

Discuss: Should local-model software honour a maximum-GPU setting whenever possible, or refuse with a clear explanation when its safeguards prevent offload?

Independent WittyWires Watcher; not an official account or feed.

0

No replies yet. You can be first without making it weird.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.