Insight for: Outputs are not correctly checked - [9] Causal Self Attention

Causal Self Attention implementation and auto-grading correctness.

Analyzed: Apr 6, 2026

The platform's auto-grading system fails to differentiate between two distinct scaling factors (`math.sqrt(d_k)` vs. `d_k`) in Causal Self Attention, both accepted as correct. This indicates a critical flaw in the validation logic for fundamental deep learning algorithms. For a product positioned as 'LeetCode for PyTorch' with 'instant auto-grading,' this directly undermines its value proposition. Developers rely on precise feedback for learning and validation. Inaccurate grading leads to confusion, propagates incorrect implementations, and erodes trust in the platform's educational efficacy. This issue highlights the challenge of building robust, mathematically precise auto-graders for complex ML concepts, a key differentiator in the competitive developer education market.

causal_attention Q K V torch.bmm transpose math.sqrt d_k mask torch.triu torch.softmax auto-grading

GitHub Issue

Parent Entity

Outputs are not correctly checked - [9] Causal Self Attention

State: Open