compat: torchvision MNIST loaders require torch.frombuffer

[lvyufeng]

Description

Several PyTorch beginner tutorials that use torchvision.datasets.FashionMNIST now fail after the torchvision import chain has been fixed, because Candle does not expose torch.frombuffer.

Repro

import candle as torch
import torchvision.datasets.mnist as mnist

# Simplest API-level repro
buf = bytearray(range(16))
t = torch.frombuffer(buf, dtype=torch.uint8, offset=4)

Current result:

AttributeError: module 'candle' has no attribute 'frombuffer'

Real-world failure site

torchvision.datasets.mnist.read_sn3_pascalvincent_tensor() uses:

parsed = torch.frombuffer(bytearray(data), dtype=torch_type, offset=(4 * (nd + 1)))

This currently blocks these tutorials under Candle's torch compat mode:

• quickstart_tutorial.py
• data_tutorial.py
• transforms_tutorial.py
• optimization_tutorial.py

Root cause

Candle uses np.frombuffer(...) internally in several places, but does not expose a public torch.frombuffer-compatible API at the top level. Once torchvision import succeeds, MNIST dataset loading reaches this missing API and fails immediately.

Notes

This appears to be a general PyTorch compatibility gap, not a tutorial-specific issue.