[LoRA] accept fal Kontext LoRAs whose embedder keys lack the base_model.model. prefix

[christopher5106]

What does this PR do?

Fixes #14917

Some fal FLUX.1 Kontext LoRAs store the block keys under base_model.model. but the global embedder keys (time_in, vector_in, txt_in, img_in, guidance_in) without that prefix. _convert_fal_kontext_lora_to_diffusers does not map those keys, so they are left in original_state_dict and loading fails with "original_state_dict should be empty at this point".

This PR maps the unprefixed embedder keys to their diffusers names (time_text_embed.*_embedder.linear_{1,2}, context_embedder, x_embedder) before that check. Files without these keys convert exactly as before.

The reproduction in #14917 fails on main and converts all 1006 keys with this change.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? Yes, Claude Code.
  • Read the Coding with AI agents guide.
  • Self-review notes, final round:
    • The converter does not map the embedder keys when they do carry the base_model.model. prefix either. I left that case out because I have not seen such a file.
    • Only lora_A / lora_B weights are mapped for the embedders, like the rest of this converter. No bias or alpha keys appear on these modules in the files I have seen.
    • No test added: the change is in a private converter and the layout is third-party, so the reproduction lives in the issue.
• Did you read the contributor guideline?
• Was this discussed/approved via a GitHub issue? fal FLUX.1 Kontext LoRAs with unprefixed embedder keys fail in _convert_fal_kontext_lora_to_diffusers #14917, with a CPU-only reproduction.

Who can review?

@sayakpaul @BenjaminBossan

[sayakpaul]

Thanks! Do you have an example LoRA checkpoint?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[christopher5106]

I was not able to find out how users got this lora.

The unprefixed embedder names do exist in one place: Black Forest Labs' official Control LoRAs. Those use that naming on every key, with no prefix anywhere.

So our users' file mixes two formats. That points to a merged or hand-converted file rather than a trainer's direct output, but nothing confirms it.

│ Source │ Result │
│ fal's Kontext trainer, fal and comfy formats │ Ruled out by training │
│ fal's fast, portrait and turbo FLUX.1 trainers │ Ruled out by training │
│ Our Kohya Kontext trainer, minimal run on dev │ Ruled out: 1,128 tensors under lora_unet_ and lora_te1_, no prefix, no embedders │
│ musubi-tuner │ Ruled out by code: it saves FLUX keys under lora_unet as well │
│ Our own copies of the diffusers converter │ None writes this layout │
│ 2,492 FLUX.1 LoRA repos on the Hub, authenticated scan │ 36 files use the prefix, none has embedder keys

[sayakpaul]

Then how did you come up with this PR if you don't have a real checkpoint?