[michalharakal]

Closes #1386, closes #1387. Part of #1384.

What

scaledDotProductAttentionBackward in DefaultExecutionTape returned null for every input, so training stopped at attention. This PR replaces the stub with the attention adjoint and removes the last entry from stubbedBackwardOpNames.

How

With S = s·Q·Kᵀ + masks, P = softmax(S), O = P·V and upstream dO:

dV = Pᵀ·dO
dP = dO·Vᵀ
dS = P ⊙ (dP − rowsum(dP ⊙ P))      (softmax backward, shared helper)
dQ = s·dS·K
dK = s·dSᵀ·Q
dM = dS reduced to the dense mask's shape

• P is recomputed in the backward, not stored, as the issue asked.
• The rule runs in a rank-5 layout [batch, kvHeads, nRep, seqQ, …]. Grouped-query K/V broadcast against the query heads of their group inside matmul, and dK/dV reduce over nRep with one sum. Gemma's layout works unchanged.
• Causal, window, rightContext and keyLengths are rebuilt from AttentionMask.additive, the same definition the forward kernel reads, so a masked score has P = 0 and exactly zero gradient. The dense mask operand is added trailing-aligned exactly as in the forward and receives a gradient. keyLengths is integer and gets none.

Tests

SdpaBackwardTest (12 cases, skainet-compile-dag commonTest):

• Central finite differences (eps 1e-2, tol 1e-2) for: plain, causal, grouped query (4 heads / 2 kv heads), key padding (padded keys get exactly 0f in dK and dV), window 2 with right context 1, dense mask [2,1,3,3], per-kv-head mask [1,2,1,3], a one-query decode step over 5 cached keys, and scale 0.5.
• PyTorch 2.5.1 parity within 2e-4 against torch.nn.functional.scaled_dot_product_attention for plain, causal and grouped+scaled cases. Fixtures were generated in float64 with seed 1384 and are embedded in the test.

Full :skainet-compile:skainet-compile-dag:jvmTest: 129/129. apiCheck clean (private and internal changes only).

Acceptance criteria of #1384

• Gradients match PyTorch autograd for causal and non-causal cases
• scaledDotProductAttention removed from stubbedBackwardOpNames; ops matrix Backward column reads ✅ (59 / 59)

Notes for review

• ops-status-matrix.adoc is hand-edited to what BackwardCoverageDumpTest emits for the now-empty stub set, since regenerating needs the full docs pipeline.
• [Lane 1 · numerics] SDPA backward — derivation note for the attention adjoint #1385 (derivation note) is not closed by this PR. The derivation above and the comments in the backward rule cover the same ground and can seed that note.

🤖 Generated with Claude Code

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1489 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.

[michalharakal]

Added TransformerBlockBackwardTest: a whole pre-norm Transformer block (gather embeddings, RMSNorm from primitives, 2-head causal SDPA, gated GELU FFN, class-token head) with all 204 parameter entries checked against central finite differences, worst diff 0.0017 at tol 0.02. End-to-end evidence that nothing in a block stops the gradient anymore.