feat(LoRA): support AI Toolkit LoRA for FLUX

[keturn]

Summary

https://github.com/ostris/ai-toolkit is "an all in one training suite for diffusion models."

and I guess LoRA saved by it don't look exactly like the ones we support already.

Related Issues / Discussions

• https://discord.com/channels/1020123559063990373/1378150538927145071

QA Instructions

• test LoRA such as
  • https://civitai.green/models/1317862?modelVersionId=1848670
  • https://civitai.green/models/1633321/akira-toriyama-dragon-ball-style-chroma

Merge Plan

N/A

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[keturn]

Current state: LoRA install but don't run.

[keturn]

this might not be what we need, as I have another LoRA here that reports the same metadata 'software': '{"name": "ai-toolkit", "repo": "https://github.com/ostris/ai-toolkit", "version": "0.1.0"}', but it uses transformer.single_transformer_blocks keys instead of diffusion_model.single_blocks keys.

[keturn]

This could use more inputs to test with; I've only tested this on a couple files from a single author.

If anyone has a LoRA collection you can use something like this script to find which ones might be applicable:

#!/usr/bin/env python
# coding: utf-8

import os
from dataclasses import dataclass
from pathlib import Path

import safetensors

INVOKEAI_ROOT = Path(os.environ.get("INVOKEAI_ROOT", "/PATH/TO/InvokeAI"))

lora_dirname = INVOKEAI_ROOT / "models" / "flux" / "lora"
lora_filenames = list(lora_dirname.glob("*.safetensors"))


@dataclass
class LoraField:
    name: str
    ai_toolkit_metadata: bool = False
    diffusion_model_key: bool = False


def inspect_lora(lora_path):
    with safetensors.safe_open(lora_path, "torch") as f:
        return LoraField(
            name=lora_path.stem,
            ai_toolkit_metadata="ai-toolkit" in (f.metadata() or {}).get("software", ""),
            diffusion_model_key=any(k.startswith("diffusion_model.") for k in f.keys())
        )

records = [inspect_lora(lora_filename) for lora_filename in lora_filenames]
matching = [r for r in records if r.ai_toolkit_metadata or r.diffusion_model_key]

print("name\tmetadata\tdiffusion_model_key")
for record in matching:
    print(f"{record.name}\t{record.ai_toolkit_metadata}\t{record.diffusion_model_key}")

[hipsterusername]

@jazzhaiku please review.