feat(io): load a weight in kernel feed order, and have it say so

[michalharakal]

Closes #1120. Follows #1124 / #1126.

A weight can now be permuted into the order the packed matmul kernels read at load time, and the resulting TensorData declares BlockOrder.INPUT_BLOCK_MAJOR instead of pretending to be canonical.

That is the whole idea: every reader is right about the same bytes. The kernels address them in feed order deliberately; toFloatArray() and view decoding walk the logical grid and fetch each block from where that order put it. Before, feed-order bytes in a type claiming to be canonical decoded to plausible garbage — #1124, and #973/#968 before it.

The payoff

#1096's O(bytes) per-weight relayout stops running at all for such a weight. DefaultCpuOpsJvm relabels the shape over the same array instead.

Asserted structurally, not by timing:

assertSame(payload, (w.data as PackedBlockStorage).packedData, "the weight's own bytes must be untouched")

A relayout would necessarily have produced a different array, so this is the claim rather than a proxy for it.

Two things that had been conflated

	shape	declares an order?
relayoutPackedWeightForKernels	[in, out]	no — its shape and its blocks disagree
a loader-produced feed-order weight	[out, in]	yes, INPUT_BLOCK_MAJOR

The first is a private artifact for the JVM kernels, which address packedData in feed order themselves. #1126 established that it has no honest block order to declare — marking it was the first fix I tried there and it produced zeros. It is left exactly as it was. The second is the new, self-describing thing, and only it flows through views.

Reusing what already existed

The permutation is TensorView.prepack(INPUT_BLOCK_MAJOR), which already owned it — and already emits the conversion on the trace, so #1117 reports the prepack with no new code. A test counts exactly one per weight, and none for a weight kept as stored.

One constraint, stated rather than assumed

KERNEL_FEED requires WeightShapeOrientation.OUT_IN. Which block is "block b of output row o" has no answer while the tensor is still labelled in the file's ne order, so refusing beats permuting against a meaningless grid.

Tests

KernelFeedOrderTest (5): the order is declared and the shape kept; feed-order bytes differ from canonical yet decode to the same matrix — with the difference asserted, so a vacuous fixture cannot pass; the product agrees either way; the permutation happens once and is traced; the OUT_IN constraint refuses.

FeedOrderWeightNoCopyTest (2, commonTest so it runs on native too): the no-copy claim, and a hand-permuted feed-order weight agreeing with the canonical weight it came from.

All fixtures are three blocks per row — at one block per row the two orders coincide and every assertion here would hold vacuously (#968).

#1115's test asserting the old refusal is rewritten, not deleted: it now pins the constraint that remains.

Gate

scripts/pr-gate.sh — all legs passed. An earlier run failed SlicingTest.testPerformanceAccessPatterns on ChromeHeadless — an unrelated dense-slicing test that passes on its own, and the same class of load-sensitive assertion as the one #1107 fixed. Not touched here.

🤖 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-1127 artifact to view the complete documentation locally.

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