Input grads fix for gemma3

[mmathew23]

This PR resolve #2131.

The hook to enable gradients for vision embeddings gets attached to a function that has no position arguments, only keywords. The torch pre_forward_hook doesn't pass kwargs so the hook throws an error as it expects one positional argument. The requires_grad_for_gradient_checkpointing has a workaround for CLIP embeddings and I add a check to to include siglip vision embeddings in this handler.

This would work as is, excecpt I noticed an issue in compiler.py. FULLGRAPH variable starts as "0" or "1" while the env variable is True or False. But the second time it's called UNSLOTH_FULLGRAPH = os.environ["UNSLOTH_FULLGRAPH"] == "1" followed by UNSLOTH_FULLGRAPH = UNSLOTH_FULLGRAPH == "1". This makes the first UNSLOTH_FULLGRAPH eval to True, but then the second eval to false. So fullgraph becomes false for everything compiled after siglip.

I fixed this so that it should eval to True. But then the issue still remains that we are trying to compile the embedding forward for Siglip and CLIP vision embeddings. I add a check right after the nn.Linear and ModuleDict check to specifically set fullgraph to false for these embeddings. Felt like a reasonable place to fix but I'm open to handling this in a more robust way.

I have a colab note book testing the changes here. Llava, llama, and gemma3 vision training all work, but gemma3 fails on a T4. There is a separate issue where the vision model bfloat16 weights do not get converted to float32 but the language component does. Will open a separate PR for that.

[johnray22]

Thank you. I used this code and can train now.

[liuluhaixiu]

Nice solution.A wonderful assist to my work in just correct moment.

[danielhanchen]

Fantastic work @mmathew23 ! :)