[Flux2] Fix LoRA loading for Flux2 Klein by adaptively enumerating transformer blocks

[songkey]

What does this PR do?

The Flux2 model features the same underlying architecture as its successor, the Flux2 Klein series, but varies in the number of Transformer Blocks — these block differences correspond to the two variants with distinct computational footprints, namely the 4B and 9B versions.

Previously, this architectural mismatch was the root cause of errors when loading LoRAs fine-tuned on Flux2 Klein models (context: ostris/ai-toolkit#667).

The update now adaptively enumerates the number of Transformer Blocks, thereby enabling seamless LoRA loading across both the original Flux2 and the entire Flux2 Klein series.

Who is this for?

Users trying to load LoRAs trained on Flux2 Klein (4B/9B variants) into the diffusers Flux pipeline.

import torch
from diffusers import Flux2KleinPipeline

# Load base model
pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16)
pipe.to("cuda")

# Before: This would raise a shape mismatch error due to block count differences
# After: This now loads correctly by adaptively matching blocks
# Replace with actual path to a FLUX.2-klein-base-4B-trained LoRA
pipe.load_lora_weights("path/to/flux2-klein-base-4b-lora.safetensors")

Checklist:

• I have checked that the script runs locally.
• I have run make style and make quality.