fix: replace numpy.where with paddle.where for tensor operations in P…

[danghoangnhan]

Replaced np.where(temp.cpu().numpy() == True) with paddle.where(temp == True)
to maintain computation on the PaddlePaddle device (GPU/CPU) without converting to NumPy.
This avoids unnecessary data transfer between devices and keeps the operation within Paddle’s tensor graph.

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

This PR attempts to optimize tensor operations by replacing np.where() with paddle.where() to avoid unnecessary device transfers between CPU and GPU. However, the implementation contains a critical bug due to API differences between NumPy and PaddlePaddle.

Key Changes:

• Modified forward_test method in rec_parseq_head.py to use paddle.where() instead of np.where() with CPU transfer

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The replacement of np.where with paddle.where is incorrect. When called with only a condition argument, np.where(condition) returns a tuple of arrays (one per dimension), while paddle.where(condition) returns a single tensor with shape [N, rank]. This breaks the indexing operation on line 369 query_mask[posi] = 0. Consider using paddle.nonzero(temp).T which returns indices in a format compatible with tuple-style indexing, or use paddle.nonzero(temp) and adjust the indexing accordingly.

Suggested change

	posi = paddle.where(temp == True)
	query_mask[posi] = 0
	posi = paddle.nonzero(temp).T
	query_mask[tuple(posi)] = 0