🧠 D: DOCUMENT — Problem & Opportunity

scaledDotProductAttentionBackward in skainet-compile/skainet-compile-dag/src/commonMain/kotlin/sk/ainet/lang/graph/DefaultExecutionTape.kt returns null for every input, so training any attention model silently produces no gradient for Q, K, V or anything upstream of them. It is now listed in stubbedBackwardOpNames and shows as ❌ stub in the Backward column of the coverage matrix.

Summary: implement the SDPA adjoint (dQ, dK, dV, honouring scale, causal and padding masks) so attention layers train.

🔍 A: ASSESS — Feasibility & Impact

• Feasibility: medium. The math is standard; the work is matching every forward option (scale, causal, mask / key-padding lengths from [Feature]: key-padding lengths for scaledDotProductAttention — mask as op metadata, lengths-based lowering #1343) and keeping memory reasonable (recompute the attention weights rather than storing them).
• Risks: numerical stability of the softmax backward; masked positions must get exactly zero gradient.

📚 R: RESEARCH — What Must Be Understood First?

• Derivation note covering every forward option (Lane 1)

🛠️ C: CODE — Implementation Plan

• Implementation (Lane 2)
• Gradient check (Lane 4)
• Review (Lane 6)

Acceptance Criteria

• Gradients match PyTorch autograd for torch.nn.functional.scaled_dot_product_attention on causal and non-causal cases
• scaledDotProductAttention removed from stubbedBackwardOpNames; matrix Backward column shows ✅

Lane breakdown

• [Lane 1 · numerics] SDPA backward — derivation note for the attention adjoint #1385 — [Lane 1 · numerics] derivation note for the attention adjoint (size:s, good first issue)
• [Lane 2 · kotlin-core] SDPA backward — implement scaledDotProductAttentionBackward #1386 — [Lane 2 · kotlin-core] implement scaledDotProductAttentionBackward (size:m)
• [Lane 4 · kotlin-core] SDPA backward — gradient check test for SDPA #1387 — [Lane 4 · kotlin-core] gradient check test for SDPA (size:s)
• [Lane 6 · review] SDPA backward — review SDPA backward #1388 — [Lane 6 · review] review SDPA backward (size:s)

Lanes 0, 3 and 5 skipped.

New here? Read Getting started as a contributor and the Issue taxonomy. Lanes marked good first issue assume no prior SKaiNET codebase knowledge. The gap markers and the coverage matrix were introduced in #1370.