Repository navigation
Replies: 2 comments
|
diffusers does not run a separate "LoRA server" - LoRA is baked into the model layers themselves, and the loader mixins wire weights into them. The whole mechanism lives in three files on current 1. The layers:
2. The adapter controls:
3. The loaders:
So the "dynamic" part you are asking about is literally: loading = writing Where to read, in order: |
|
Most of the plumbing is in The pipeline-level API: pipe.load_lora_weights("repo_or_path", adapter_name="a")
pipe.set_adapters(["a", "b"], adapter_weights=[1.0, 0.5])
pipe.disable_lora() # scale 0, weights stay in memory
pipe.unload_lora_weights()
pipe.fuse_lora() # bakes into base weights permanently
— |
Uh oh!
There was an error while loading. Please reload this page.
How does diffusers manage lora? Including dynamic loading, enabling, disabling, etc. Is there any detailed implementation reference?
All reactions