Having plugins installed slows even basic llm --help from 1.3s to 10.7s

[mcint]

I would like to make use of the extensive plugin offering to try various models available with ease and without unnecessary overhead. Latency is everything for UIs, and CLIs are generally great because they're fast and responsive, no memory and cpu hungry custom rendering to support text entry -- .5 s is noticable, 1.3 is checking my pulse, 10s is why am I using this tool. Could plugins be architecture to work in still more minimal ways until needed? Declare more of their offerings up front, defer initialization, or cache state, perhaps with modified time cache keys.

For subcommands that don't rely on loading accessory plugins, llm logs -t -n1, setting LLM_LOAD_PLUGINS='' speeds up access from 6.2s to 0.9s.

[mcint]

Plugins blocked with env var

With plugins -- seems domainted by sentence-transformers embedding plugin, especially its importing of torch modules.