[chmjkb]

Description

Moves the EmbeddingGemma 2 state dict hooks onto the shared weight helpers and adds permuteRows. The last hook migration of phase 1 in #70.

• permuteRows(t, order): row i of the result is row order[i] of t, like torch t[order]. Lazy, keeps dtype and byteLength. Throws CHECKPOINT_MISMATCH when order is not a permutation.
• EmbeddingGemma 2 (+71 / −122):
  • Fused q/k/v (text, vision, audio) is concatRows. The rope reorder is permuteRows with a ropeOrder(rows, block) index list, replacing interleaveHeads.
  • qkv_norm tiles the permuted norms with concatRows, plus ones for v. byteLength still counts each norm once.
  • PLE projection and the audio post norm use scaleWeight. The vision position table is splitRows + reshapeWeight. The audio convs are reshapeWeight, permuteRows and transposeWeight.
  • The weightless-norm ones are f16 halfwords (exact).
  • Clip bounds, layer_scalar, rel and q_scale stay f32: CPU code and the attention kernel read them as f32.
• depthart: nothing fits the helpers (BatchNorm folds, a three-branch conv merge, per-channel tensor scales, -exp(x)), so it only gets one-line comments saying why each stays f32.

Names, shapes and byteLengths reaching the models are unchanged.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Performance (kernel tuning, graph building, anything related to performance)
• Other (chores, tests, code style improvements etc.)

Testing instructions

1. pnpm typecheck && pnpm lint && pnpm --filter runntime test:unit (941 pass)
2. New test/zoo/embeddinggemma2/stateDictHooks.test.ts: all three towers on F32 and F16 synthetic checkpoints, values against hand-worked math including the rope row order. permuteRows cases in test/core/weights/transforms.test.ts.
3. pnpm --filter runntime test:gpu: 1439 passed, 16 expected failures. The zoo leak tests bind random weights, so they don't run these hooks.

Related issues

Part of #70. Stacked on #78.

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

• scaleWeight keeps products in f32 (fixed in feat(core): add lazy weight helpers for state dict hooks and move MiniLM onto them #75), so the scaled PLE projection and audio post norms load the same values as on main.
• Not checked against the real hosted EmbeddingGemma 2 file headers (the Hub repo returned 401 from the agent's environment).