[Bug] AutoModel.from_pretrained() crashes with RuntimeError: Tensor.item() cannot be called on meta tensors

[aswin00000]

Bug Report

Model: OpenGVLab/VideoMAEv2-Base
File: modeling_videomaev2.py — VisionTransformer.__init__() line 389
Severity: Blocks AutoModel.from_pretrained() — the standard HuggingFace loading API

---

Minimal Reproducible Example

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "OpenGVLab/VideoMAEv2-Base",
    trust_remote_code=True
)

Full Traceback

RuntimeError: Tensor.item() cannot be called on meta tensors
  File "modeling_videomaev2.py", line 389, in __init__
    dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]

---

Root Cause

Hugging Face Transformers wraps every model constructor inside torch.device("meta") for memory-efficient loading:

with torch.device("meta"):
    model = cls(config)  # ← VisionTransformer.__init__() runs here

A meta tensor has shape and dtype but no actual data values. Line 389 calls .item() to extract a scalar — which is impossible on a meta tensor.

Call	Meta context?	Result
VideoMAEv2(config)	❌ No	✅ Works
AutoModel.from_pretrained(...)	✅ Yes	❌ Crashes

---

Proposed Fix

- dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
+ # Meta-safe: no PyTorch tensor needed here
+ if depth > 1:
+     dpr = [drop_path_rate * i / (depth - 1) for i in range(depth)]
+ else:
+     dpr = [0.0] * depth

Produces numerically identical values. No tensor operations during construction.

---

Environment

Item	Version
transformers	latest
torch	2.x
OS	Any

Happy to open a PR with this fix if maintainers confirm the approach.