diffusers docker image: LoRA loading fixes (#380)

* diffusers docker image: load lora uncaught errors
broader exception type catch on fuse_lora
this fixes the follwing bug: when some error was encountered in _fuse_or_raise the lora adapter the exception was not always caught, as a consequence the lora adapter was not unloaded. Any subsequent call to the same adapter actually performed inference from the base model, as long as no new adapter was loaded

* diffusers image: upgrade peft dependency
Fixes some LoRA adapters loading issues with peft backend
See huggingface/diffusers#6809

Author: oOraph

docker_images/diffusers/app/lora.py

@@ -62,7 +62,7 @@ def _is_pivotal_tuning_lora(model_data):
     def _fuse_or_raise(self):
         try:
             self.ldm.fuse_lora(safe_fusing=True)
-        except ValueError as e:
+        except Exception as e:
             logger.exception(e)
             logger.warning("Unable to fuse LoRA adapter")
             self.ldm.unload_lora_weights()

docker_images/diffusers/requirements.txt

@@ -17,4 +17,8 @@ uvicorn>=0.23.2
 gunicorn>=21.2.0
 psutil>=5.9.5
 aiohttp>=3.8.5
-peft==0.7.1
+# Tmp: replace with version once there is a tag above 0.8.1
+# This commit includes a fix that concerns
+# https://github.com/huggingface/diffusers/issues/6809
+# peft==0.7.1
+git+https://github.com/huggingface/peft.git@a30e006bb2f82824ab24435b9f07932738b44966