[Fix] Make vision encoder DropPath initialization safe on meta devices

[cyyecao-lappland]

Summary

Related to #1254.

The DropPath schedule inherits the default device from torch.linspace. Under a meta-device initialization context, extracting its Python scalar probabilities with .item() raises RuntimeError: Tensor.item() cannot be called on meta tensors.

• Explicitly construct the schedule on CPU in both internvl_chat and internvl_chat_gpt_oss.
• Preserve the existing schedule values and keep model parameters on the requested device.
• Add small-config regression tests for both real vision implementations, without downloading pretrained weights. The tests import the vision modules under temporary package namespaces to avoid unrelated chat-backend initialization; model and dependency implementations are not mocked.

Thanks to @Liu-524 for identifying the explicit-CPU workaround in this comment.

Validation

Tested on Linux with Python 3.12.8, PyTorch 2.10.0+cu128, Transformers 5.1.0, timm 1.0.30, Accelerate 1.15.0, pytest 9.1.1, and an NVIDIA RTX 4090.

CUDA_VISIBLE_DEVICES=0 python -m pytest tests/test_intern_vit_initialization.py -q

The same regression suite produced:

• Before the fix: 12 failed, 8 passed, with all failures caused by scalar extraction from meta tensors.
• After the fix: 20 passed (CUDA cases ran, rather than being skipped).

Coverage includes CPU/meta parameter placement, one/four layers, zero/nonzero DropPath rates, exact schedule preservation, and small-config state-dict checkpoint round trips from meta initialization to CPU/CUDA with matching finite forward outputs on the same device.

git diff --check passes. Repository pre-commit checks on the changed files pass or have no applicable files, except flake8: both existing model files report W604 for docstring backticks at line 191. Running the same checks on the unmodified baseline reproduces those errors.

Scope and known limitations

This is a focused initialization fix, not a claim of full Transformers 5.x compatibility.

A separate tiny-model save_pretrained() / from_pretrained() probe moves past the original .item() failure after this patch, but still fails with AttributeError: 'InternVisionModel' object has no attribute 'all_tied_weights_keys' under Transformers 5.1.0. That loading issue and the separate generate/GenerationMixin reports are outside this patch.

No full pretrained InternVL checkpoint or end-to-end image/chat inference was tested. The passing checkpoint tests use PyTorch state-dict loading (assign=True), not Hugging Face from_pretrained().