Optimize llm-mlx startup performance

[simonw]

Refs:

• llm --help is slow (all commands are slow) with certain plugins installed llm#949

[simonw]

Based on what I learned from this issue:

• Optimize llm-sentence-transformers startup performance llm-sentence-transformers#20

And this bit of the profile trace: https://static.simonwillison.net/static/2025/llm-apr-23-profile.svg?x=36&y=1188

I suspect this is caused by these imports being slow:

llm-mlx/llm_mlx.py

Lines 4 to 6 in d545130

	import mlx.core as mx
	from mlx_lm import load, stream_generate
	from mlx_lm.sample_utils import make_sampler

(Because they import Torch.)

[simonw]

With those imports at the base of the module:

time llm --help

llm --help 0.55s user 0.15s system 99% cpu 0.709 total

Moving those imports into the execute() and ._load() methods:

llm --help 0.29s user 0.06s system 98% cpu 0.354 total

Saves nearly half a second.