Commit 53c44a9a authored by Guangda Lai's avatar Guangda Lai Committed by TensorFlower Gardener
Browse files

Make a real repro for IConstantLayer: we need to call getType() before

getDimensions(), then if CreateConstantLayer() doesn't set the dtype, the test
will fail.

Before CreateConstantLayer() was added, graphs which trigger conversion of INT32
weights to IConstantLayer will crash the TRT conversion. This is a bug in TRT
5.0 and I has reported it to NVIDIA. CreateConstantLayer() fixed that by setting
the dtype of the output ITensor of the IConstantLayer, but the added test could
not reproduce the error when we remove the setting code. This change fix the
test, so when removing the dtype setting logic the test will fail.

Note that, after this fix, during conversion it'll report a lot of spammy logs
but won't crash. Once NVIDIA fixed the TRT bug we should remove the dtype
setting logic and the test should still pass.

PiperOrigin-RevId: 229758150
parent c9bd0744
Loading
Loading
Loading
Loading
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please to comment