[XiaoHongbo-Hope]

Purpose

LeRobot video batches currently allocate a CPU tensor and then call share_memory_(), copying the entire uninitialized storage before filling it. Allocate shared storage directly, following PyTorch DataLoader's collation pattern, to avoid that copy.

The change applies to shared CPU batch output; other allocation paths are unchanged. No standalone throughput improvement is claimed.

Tests

• Video parallelism suite: 14 passed, 26 subtests passed. The regression checks shared, contiguous uint8/float outputs, repeated/unsorted windows, sample isolation, and rejects calls to share_memory_().
• LeRobot suite: 67 passed, 29 optional tests skipped, 72 subtests passed (PyTorch 2.7.1+cpu).
• Repository flake8, Python 3.6 syntax compilation, and git diff --check passed.

[JingsongLi]

+1