Fix/transformers runtime patch

[wildcraft958]

Summary

Fixes several issues that prevent loading pretrained LISAT models in different environments.

Problems Fixed

1. ValueError: Unknown vision tower: /home/patrickwu/...

The saved model config contains hardcoded absolute paths from the original training machine. When users download the model, this path doesn't exist.

Fix: builder.py now falls back to openai/clip-vit-large-patch14 when the config path doesn't exist locally.

2. KeyError: 'train_mask_decoder'

During inference loading via from_pretrained(), kwargs like train_mask_decoder and out_dim aren't passed, causing a KeyError.

Fix: LISAT.py and LISAT_eval.py now use kwargs.get() with sensible defaults.

3. Transformers version/registration errors

Strict version checks in transformers cause ImportError at import time. Additionally, duplicate AutoConfig.register calls raise ValueError: already used.

Fix: Runtime monkey-patch disables strict version checks before importing transformers, and wraps registration methods to ignore duplicates.

Files Changed

• model/LISAT.py — kwargs defaults + transformers patch
• model/LISAT_eval.py — kwargs defaults + transformers patch
• model/llava/model/multimodal_encoder/builder.py — vision tower fallback

Testing

• Syntax verified with py_compile
• Successfully tested in Modal deployment with Python 3.12 + torch 2.3 + transformers 4.31

---

Contribution by @wildcraft958