Paligemma 2 support

[baekbyte]

Quick question: is paligemma 2 supported?

I can't seem to find it listed under the model catalogue

[pauldev-hub]

Hi @baekbyte ,

Short answer: PaliGemma 2 is not in Unsloth's official model catalogue, but you can still fine-tune it using Unsloth's generic vision fine-tuning path.

Unsloth's vision/multimodal collection includes Llama 3.2, LLaVa, Qwen2 VL, Pixtral, and PaliGemma (original) — but PaliGemma 2 is not listed as an officially optimized model. (Hugging Face)

Why it's not in the catalogue:

PaliGemma 2 uses a unique architecture — it's a VLM designed primarily for fine-tuning on specific downstream tasks (captioning, VQA, OCR), not general chat. Unsloth focuses on chat/instruction models. PaliGemma 2 also uses its own special prompt format unlike standard chat templates.

Workaround — use FastVisionModel directly:

Since Unsloth supports "nearly all transformer-style vision models" via their auto-model system, you can try:
from unsloth import FastVisionModel

model, processor = FastVisionModel.from_pretrained(
"google/paligemma2-3b-pt-224", # or 10B/28B
load_in_4bit = True,
)

If the architecture isn't natively patched, it falls back to standard HuggingFace transformers with Unsloth's memory optimizations still partially applied.
For a dedicated optimized notebook, check if Unsloth has added one after this post was made — the team adds new models frequently.

If this helped you, please mark it as the answer — it helps others in the community who run into the same issue find the solution faster!