Brian Hirsh

hmm @JackCaoG I'm having a hard time understanding the XLA test failure from https://github.com/pytorch/pytorch/actions/runs/8991496548/job/24702482398. Do you know what's going wrong? I see that this error is a red-herring in the...

oh, it looks like it's just a segfault. the stacktrace from the XLA logs points to the FunctionalStorageImpl constructor (which I did change). I thought my check for `TensorImpl->has_storage()` would...

@pytorchbot merge

@pytorchbot merge

XLA no longer segfaults, but there is still an XLA test failing this assert: ``` assert n.args[0] in placeholders ``` I can't easily run the XLA test locally, so I'm...

ah @JackCaoG this is the printed graph from the failing test (at https://github.com/pytorch/pytorch/actions/runs/9003113930/job/24737735268, `TestDynamoBufferDonationAliasingWithCustomOp.test_manual_buffer_donation_for_inplce_op_repeat`): ``` %arg0_1 : [num_users=1] = placeholder[target=arg0_1] %dynamo_set_buffer_donor_ : [num_users=2] = call_function[target=torch.ops.xla.dynamo_set_buffer_donor_.default](args = (%arg0_1, True), kwargs =...

@pytorchbot merge

@pytorchbot merge

Hmm strangely `PYTORCH_TEST_WITH_DYNAMO=1 python test/functorch/test_aotdispatch.py -k test_set__not_allowed` does not repro the failure for me locally (I see the exception getting properly raised)

Yup that's right. It looks like `channel_shuffle` has a composite implementation, but mobile wants `XNNPACK` to handle it directly. From what I understand, the current state of `XNNPACK` is that...