Fusion print is broken with the presence of castOp.

[jjsjann123]

🐛 Bug

A type assert is mysteriously failing when I have a fusion with a castOp. I briefly scanned through the code in arith.cpp looks like the type is properly set.

To Reproduce

I have the repro here: https://github.com/csarofeen/pytorch/tree/castOp_repro
(basically just trying to print fusion with a castOp).

build and run with ./test_jit --gtest_filter="*GPU_FusionCast*

C++ exception with description "dtype_ != DataType::Null INTERNAL ASSERT FAILED at "../torch/csrc/jit/codegen/cuda/ir_base_nodes.cpp":134, please report a bug to PyTorch. Value does not have a data type.
Exception raised from getDataType at ../torch/csrc/jit/codegen/cuda/ir_base_nodes.cpp:134 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0x6b (0x7fa9c29ef99b in /volume/codegen_project/pytorch_codegen/build/lib/libc10.so)
frame #1: torch::jit::fuser::Val::getDataType() const + 0x3d4 (0x7fa9c5b1de44 in /volume/codegen_project/pytorch_codegen/build/lib/libtorch_cuda.so)
frame #2: torch::jit::fuser::IRPrinter::handle(torch::jit::fuser::UnaryOp const*) + 0x1ad (0x7fa9c5b3d4dd in /volume/codegen_project/pytorch_codegen/build/lib/libtorch_cuda.so)
frame #3: void torch::jit::fuser::Expr::constDispatch<torch::jit::fuser::OptInConstDispatch*>(torch::jit::fuser::OptInConstDispatch*, torch::jit::fuser::Expr const*) + 0xb0 (0x7fa9c5ae76c0 in /volume/codegen_project/pytorch_codegen/build/lib/libtorch_cuda.so)
frame #4: torch::jit::fuser::IRPrinter::handle(torch::jit::fuser::Fusion*) + 0x55 (0x7fa9c5b3a745 in /volume/codegen_project/pytorch_codegen/build/lib/libtorch_cuda.so)
frame #5: torch::jit::fuser::operator<<(std::ostream&, torch::jit::fuser::Fusion*) + 0x6a (0x7fa9c5b3caaa in /volume/codegen_project/pytorch_codegen/build/lib/libtorch_cuda.so)
frame #6: torch::jit::testGPU_FusionCastOps() + 0x8f (0x5605261e181f in ./test_jit)
frame #7: void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) + 0x4a (0x56052623d05a in ./test_jit)
frame #8: <unknown function> + 0x224fc5 (0x560526232fc5 in ./test_jit)
frame #9: <unknown function> + 0x2255d5 (0x5605262335d5 in ./test_jit)
frame #10: <unknown function> + 0x225885 (0x560526233885 in ./test_jit)
frame #11: testing::internal::UnitTestImpl::RunAllTests() + 0xc1c (0x5605262348dc in ./test_jit)
frame #12: testing::UnitTest::Run() + 0x98 (0x560526234b98 in ./test_jit)
frame #13: main + 0xc8 (0x560526081528 in ./test_jit)
frame #14: __libc_start_main + 0xe7 (0x7fa9c1c91b97 in /lib/x86_64-linux-gnu/libc.so.6)
frame #15: _start + 0x2a (0x56052608d30a in ./test_jit)
" thrown in the test body.

[kevinstephano]

I ran into this as well. It seems like the Val for the input or output loses its DataType by the time it got printed. This is prior to lowering.

[csarofeen]

@kevinstephano is this still broken? I thought you fixed it.

[kevinstephano]

This was fixed with #168. Closing...