🧠 D: DOCUMENT — Problem & Opportunity

GroupNormalization (skainet-lang/skainet-lang-core/src/commonMain/kotlin/sk/ainet/lang/nn/normalization/GroupNormalization.kt) is reachable from the public groupNorm {} builder in NetworkBuilder, but its four forward helpers — reshapeForGroups, calculateGroupMean, calculateGroupVariance, reshapeFromGroups — are TODO(), so any network using it throws on the first forward pass. In the same file eps is never applied and gamma/beta are VoidOpsTensor placeholders whose set does nothing. The class is now marked @NotImplemented and shows as ❌ in the NN-layers table of the coverage matrix.

GroupNorm is the standard normalization in diffusion U-Nets, many detection backbones and small-batch training, where BatchNorm is unusable.

Summary: implement GroupNormalization.forward with torch.nn.GroupNorm semantics using ops that already exist.

🔍 A: ASSESS — Feasibility & Impact

• Feasibility: high — everything needed exists: reshape, mean(dim), variance(dim), sqrt, broadcasting arithmetic, and LayerNormalization.kt in the same package as a pattern. The StableHLO converter already lowers a groupNorm node.
• Semantics to match: torch.nn.GroupNorm(num_groups, num_channels, eps, affine): input [N, C, *] reshaped to [N, G, C/G * prod(*)], biased variance, (x - mean) / sqrt(var + eps), then per-channel gamma/beta.
• Risks: biased vs unbiased variance (PyTorch uses biased); check which one TensorOps.variance computes.

📚 R: RESEARCH — What Must Be Understood First?

• PyTorch reference cases (Lane 1)

🛠️ C: CODE — Implementation Plan

• Implement forward + fix eps and gamma/beta (Lane 2)
• Docs (Lane 5)
• Review (Lane 6)

Acceptance Criteria

• groupNorm {} network runs forward and matches PyTorch within 1e-5
• @NotImplemented removed from GroupNormalization; NN-layers table shows ✅

Lane breakdown

• [Lane 1 · numerics] GroupNormalization — PyTorch reference values for GroupNorm #1380 — [Lane 1 · numerics] PyTorch reference values for GroupNorm (size:xs, good first issue)
• [Lane 2 · kotlin-core] GroupNormalization — implement GroupNormalization.forward #1381 — [Lane 2 · kotlin-core] implement GroupNormalization.forward (size:s, good first issue)
• [Lane 5 · docs] GroupNormalization — document groupNorm in the NN DSL docs #1382 — [Lane 5 · docs] document groupNorm in the NN DSL docs (size:xs, good first issue)
• [Lane 6 · review] GroupNormalization — review GroupNormalization #1383 — [Lane 6 · review] review GroupNormalization (size:xs)

Lanes 0, 3 and 4 skipped: no design question, no platform code, ground truth is covered by the Lane-2 unit test against Lane-1 values.

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.