---
library_name: transformers
tags:
- test-fixture
- paligemma
- peft
- lora
---

# hf-internal-testing/tiny-random-paligemma-lora-key-mapping

Tiny-random PaliGemma checkpoint bundling a LoRA adapter that **requires a `key_mapping` to load onto the
underlying `PaliGemmaModel`**.

It mirrors [`vidore/colpali`](https://huggingface.co/vidore/colpali) at tiny scale: the adapter's text weights
are stored under the old `language_model.model.layers.*` layout, so loading them onto today's
`PaliGemmaModel` (`language_model.layers.*`) needs:

```python
from transformers import PaliGemmaModel

model = PaliGemmaModel.from_pretrained(
    "hf-internal-testing/tiny-random-paligemma-lora-key-mapping",
    key_mapping={r"language_model\.model\.": "language_model."},
)
```

`PaliGemmaForConditionalGeneration` auto-bridges this (via the `llava` conversion) and does not need the
mapping; the bare `PaliGemmaModel` does. Every `lora_A` weight is filled with `0.0234` and every
`lora_B` weight with `0.0567`, so a test can assert the adapter was restored from the checkpoint.