fix(training): route ConvGrad outputs by shape, not by position

[runwangdl]

What

SplitConvGradPass read a fused ConvGrad's outputs positionally — outputs[0] as dX, outputs[1] as dW — and its docstring said so explicitly: "1 output (dX only): ConvGradX".

That assumption does not hold. ORT emits dX only when something downstream needs it. A trainable layer whose input needs no gradient — the first trainable layer of a network — emits dW alone. The positional read takes that lone dW to be dX and wraps it in a ConvGradX.

How it surfaces

Nothing downstream recovers. Conv2DGradXParser binds it as grad_in; the tiler slices it along its output channels; ConvGradXTileConstraintBase reads those cube offsets as batch offsets and derives a dY cube at (k, 0, 0, 0) on a tensor whose batch extent is 1:

AssertionError: Rectangle offset should be zero when the dimensions are the same.
Received rectangle HyperRectangle(offset=(1, 0, 0, 0), dims=(1, 16, 1, 1))
and reference shape (1, 16, 24, 24)

The assertion fires five layers away from the cause, which is why it reads as a tiler bug.

Fix

dX carries X's shape and dW carries W's, which tells them apart whatever order they arrive in — so classify by shape. Where shapes are unavailable the positional reading is kept, so graphs without shape annotations behave exactly as before. A node with no dX no longer produces a ConvGradX at all.

Silent mis-binding today

ResNet8 already contains such a node (layer1_0_conv1_b, dW only). It does not assert, because its shapes never reach that branch — so that graph has been binding a weight gradient as an input gradient all along, without any symptom.

Validation

	result
channel-split MobileNetV1	assertion above → clean deploy, planned L3 = 3246.8 KB
pytest -m "gap9_tiled and training and singlebuffer"	19 passed, 0 failed
pytest -m deeploy_internal	44 failures before and after (40 test_dmas.py, 4 test_deeploy_internal.py), all pre-existing

DeeployTest/test_split_convgrad.py builds the graphs by hand, runs in 0.16 s with no ONNX asset or backend, and covers dW-only, dX-only, both orders, the bias case, and which operands each split node reads. Two of its five tests fail without this change.

🤖 Generated with Claude Code