Description

• Try to slove using wrong lora model cased by alias. The forbidden_network_aliases seems only used in generating the model card, but not used during inference.
• changes in code: when name.lower() in forbidden_network_aliases, get network_on_dick from available_networks rather than available_network_aliases.

I'm training lora models and I have two models with same 'ss_output_name' but different filename (since I changed the filename after training). When I run inference on webui, I find the result doesn't match the model name I write in prompt. After adding code for displaying the model path for name in prompt, I found it is using the wrong model.

This bug may caused by:
For example, there is a model A.safetensors and 'ss_output_name' in metadate is 'A'. And there is a model B.safetensors which 'ss_output_name' in metadata is also 'A'. During constructing the available_network_aliases dict, A.safetensors is processed before B.safetensors. After processing B.safetensors, key 'A' in available_network_aliases dict is mapped to B.safetensors. Even through this alias is stored in forbidden_network_aliases, network_on_disk obj is still queried in available_network_aliases. When I use '<lora:A:1>' in prompt, it will call B.safetensors instead of A.safetensors.

Screenshots/videos:

For reproducing the bug, I use two lora models from civitai.
l0ngma1d.safetensors
Yatogami Tohka(dal).safetensors
The original filename and 'ss_output_name' is same. For reproducing this bug, I changed the 'ss_output_name' in Yatogami Tohka(dal).safetensors to 'l0ngma1d'. And here are the models I used.
https://drive.google.com/drive/folders/1vRg9h2P3H28zTaz-9QZOS9Pw_3cI-bl9?usp=drive_link

Generation parameters

l0ngma1d

<lora:l0ngma1d:1>, 1girl
Negative prompt: blur
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 233, Size: 512x768, Model hash: fd0db59b48, Model: threeDelicacyWonton_v2, Clip skip: 2, Lora hashes: "l0ngma1d: 42ba22fee2d5", Version: v1.7.0-331-gcb5b335a

Yatogami Tohka(dal)

<lora:Yatogami Tohka(dal):1>, 1girl
Negative prompt: blur
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 233, Size: 512x768, Model hash: fd0db59b48, Model: threeDelicacyWonton_v2, Clip skip: 2, Lora hashes: "Yatogami Tohka(dal): 06acbeef6ed5", Version: v1.7.0-331-gcb5b335a

Inference with origin model

l0ngma1d	Yatogami Tohka(dal)

Inference with modified model

l0ngma1d	Yatogami Tohka(dal)

Inference with modified model (My code)

l0ngma1d	Yatogami Tohka(dal)

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests