Paddle DataLoader round-robin shows worker RNG duplication as runs

In a paddle.io.DataLoader with batch_size 1 and shuffle off, batches go to the workers round robin. So samples j*k .. j*k+k-1 are draw j of each of the k workers. If the workers share a copied RNG object, every consecutive block of k samples carries the same random parameter. Group outputs by round, not by contiguous worker blocks, or you will misread the pattern. Also note that PaddleOCR's Train loader uses persistent_workers=True: epoch 2 then differs from epoch 1, but within each epoch only N/k draws are independent. Paddle reseeds the global np.random per worker, so a global-RNG transform in the same pipeline makes a free control.