[Bug] tutorials do not build from a clean source tree

[areusch]

I think this might be a compiler caching bug cc @tqchen @jroesch @mbs-octoml . Not caught in the regression because the regression rebuilds only changed tutorials.

Steps to reproduce:

1. git checkout dc2f70e3c8a9b14b9e414ecf768ad32e6c3c3960
2. rm -rf build
3. docker/bash.sh ci_gpu tests/scripts/task_config_build_gpu.sh
4. docker/bash.sh ci_gpu tests/scripts/task_build.sh build -j16
5. docker/bash.sh ci_gpu bash -c 'cd docs && make clean'
6. docker/bash.sh ci_gpu tests/scripts/task_ci_setup.sh
7. docker/bash.sh ci_gpu tests/scripts/task_python_docs.sh

Will show this traceback somewhere along the way. micro_autotune was just trying to build a relay model. I think the shapes look correct to me. Rerunning task_python_docs.sh should cause them to build.

conv2d: requires that `0`, the input channels (0) divided by groups (1),                                                                                                          [1324/4555]
 must match the input channels of the weight `3`, where the weight shape is ([6, 3, 5, 5]).
The type inference pass was unable to infer a type for this expression.
This usually occurs when an operator call is under constrained in some way, check other reported errors for hints of what may of happened.
WARNING: /home/areusch/ws/tvm4/tutorials/micro/micro_autotune.py failed to execute correctly: Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/sphinx_gallery/gen_rst.py", line 480, in _memory_usage
    out = func()
  File "/usr/local/lib/python3.6/dist-packages/sphinx_gallery/gen_rst.py", line 465, in __call__
    exec(self.code, self.globals)
  File "/home/areusch/ws/tvm4/tutorials/micro/micro_autotune.py", line 179, in <module>
    lowered = tvm.relay.build(relay_mod, target=TARGET, params=params)
  File "../../python/tvm/relay/build_module.py", line 358, in build
    mod=ir_mod, target=target, params=params, executor=executor, mod_name=mod_name
  File "../../python/tvm/relay/build_module.py", line 172, in build
    self._build(mod, target, target_host, executor, mod_name)
  File "tvm/_ffi/_cython/./packed_func.pxi", line 323, in tvm._ffi._cy3.core.PackedFuncBase.__call__
  File "tvm/_ffi/_cython/./packed_func.pxi", line 267, in tvm._ffi._cy3.core.FuncCall
  File "tvm/_ffi/_cython/./base.pxi", line 163, in tvm._ffi._cy3.core.CALL
tvm.error.DiagnosticError: Traceback (most recent call last):
  27: TVMFuncCall
        at /home/areusch/ws/tvm4/src/runtime/c_runtime_api.cc:474
  26: tvm::runtime::PackedFunc::CallPacked(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1151
  25: std::function<void (tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)>::operator()(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
        at /usr/include/c++/7/bits/std_function.h:706
  24: std::_Function_handler<void (tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*), tvm::relay::backend::RelayBuildModule::GetFunction(std::__cxx11::basic_string<char, std::char_traits<c
har>, std::allocator<char> > const&, tvm::runtime::ObjectPtr<tvm::runtime::Object> const&)::{lambda(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)#3}>::_M_invoke(std::_Any_data const&,
tvm::runtime::TVMArgs&&, tvm::runtime::TVMRetValue*&&)
        at /usr/include/c++/7/bits/std_function.h:316
  23: tvm::relay::backend::RelayBuildModule::GetFunction(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, tvm::runtime::ObjectPtr<tvm::runtime::Object
> const&)::{lambda(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)#3}::operator()(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
        at /home/areusch/ws/tvm4/src/relay/backend/build_module.cc:181
  22: tvm::relay::backend::RelayBuildModule::Build(tvm::IRModule, tvm::runtime::Map<tvm::Integer, tvm::Target, void, void> const&, tvm::Target const&, tvm::runtime::String, tvm::runtime::St
ring)
        at /home/areusch/ws/tvm4/src/relay/backend/build_module.cc:288
  21: tvm::relay::backend::RelayBuildModule::BuildRelay(tvm::IRModule, std::unordered_map<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, tvm::runtime::NDAr
ray, std::hash<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::equal_to<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char>
 > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const, tvm::runtime::NDArray> > > const&, tvm::runtime::String)
        at /home/areusch/ws/tvm4/src/relay/backend/build_module.cc:479
  20: tvm::relay::backend::RelayBuildModule::Optimize(tvm::IRModule, tvm::runtime::Map<tvm::Integer, tvm::Target, void, void> const&, std::unordered_map<std::__cxx11::basic_string<char, std
::char_traits<char>, std::allocator<char> >, tvm::runtime::NDArray, std::hash<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::equal_to<std::__cxx11::
basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const, tvm::run
time::NDArray> > > const&)
        at /home/areusch/ws/tvm4/src/relay/backend/build_module.cc:329
  19: tvm::transform::Pass::operator()(tvm::IRModule) const
        at /home/areusch/ws/tvm4/src/ir/transform.cc:255
  18: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
        at /home/areusch/ws/tvm4/src/ir/transform.cc:267
  17: tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
        at /home/areusch/ws/tvm4/src/ir/transform.cc:481
  16: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
       at /home/areusch/ws/tvm4/src/ir/transform.cc:267
  15: tvm::relay::transform::FunctionPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
        at /home/areusch/ws/tvm4/src/relay/ir/transform.cc:160
  14: tvm::transform::Pass::operator()(tvm::IRModule) const
        at /home/areusch/ws/tvm4/src/ir/transform.cc:255
  13: tvm::transform::Pass::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
        at /home/areusch/ws/tvm4/src/ir/transform.cc:267
  12: tvm::transform::ModulePassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const
        at /home/areusch/ws/tvm4/src/ir/transform.cc:415
  11: tvm::runtime::TypedPackedFunc<tvm::IRModule (tvm::IRModule, tvm::transform::PassContext)>::operator()(tvm::IRModule, tvm::transform::PassContext) const
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1498
  10: tvm::IRModule tvm::runtime::detail::typed_packed_call_dispatcher<tvm::IRModule>::run<tvm::IRModule, tvm::transform::PassContext>(tvm::runtime::PackedFunc const&, tvm::IRModule&&, tvm:
:transform::PassContext&&)
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1444
  9: tvm::runtime::TVMRetValue tvm::runtime::PackedFunc::operator()<tvm::IRModule, tvm::transform::PassContext>(tvm::IRModule&&, tvm::transform::PassContext&&) const
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1369
  8: std::function<void (tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)>::operator()(tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*) const
        at /usr/include/c++/7/bits/std_function.h:706
  7: _M_invoke
        at /usr/include/c++/7/bits/std_function.h:316
  6: operator()
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1492
  5: unpack_call<tvm::IRModule, 2, tvm::relay::transform::InferType()::<lambda(tvm::IRModule, const PassContext&)> >
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1421
  4: run<>
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1382
  3: run<tvm::runtime::TVMMovableArgValueWithContext_>
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1382
  2: run<tvm::runtime::TVMMovableArgValueWithContext_, tvm::runtime::TVMMovableArgValueWithContext_>
        at /home/areusch/ws/tvm4/include/tvm/runtime/packed_func.h:1397
  1: operator()
        at /home/areusch/ws/tvm4/src/relay/transforms/type_infer.cc:857
  0: tvm::DiagnosticContext::Render()
        at /home/areusch/ws/tvm4/src/ir/diagnostic.cc:105
  File "/home/areusch/ws/tvm4/src/ir/diagnostic.cc", line 105

[mbs-octoml]

@electriclilies Lily, in your spelunking through build did you see any obvious global compile engine caching?

[areusch]

Failed again in mainline CI: https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-9050/1/pipeline

[mbs-octoml]

No luck on a local repro (using nvidia docker and the ci_gpu image) as nvcc.py is failing for some reason.

[areusch]

@mbs-octoml i think you can repro like this:

1. git checkout 44d3934be5d33590ba63139f9b756b05aec9d5c5
2. rm -rf build
3. docker/bash.sh ci_gpu tests/scripts/task_config_build_gpu.sh
4. docker/bash.sh ci_gpu tests/scripts/task_build.sh build -j16 # note: adjust -j16 on your box
5. docker/bash.sh -it ci_gpu bash -c 'cd docs && make clean && TVM_TUTORIAL_EXEC_PATTERN="(micro)|(dev/use_pass_infra)" make html'

it definitely will not repro if you've already built the docs once and don't run cd docs && make clean.

[mbs-octoml]

I can repro with my local config just with make clean & make html, no need for docker etc. Good, that's easier.

[zxybazh]

Same here in my PR's CI. #9053

[mikepapadim]

I had the issue also in the stagging Jenkins.

[mbs-octoml]

I'm looking again now.

[mbs-octoml]

On the BAD runs:

[14:56:21] /home/mbs/github/mbs-tvm/src/relay/ir/transform.cc:135: AlterOpLayout: Input module:
def @main(%data: Tensor[(1, 3, 10, 10), float32]) -> Tensor[(1, 6, 10, 10), float32] {
  nn.conv2d(%data, meta[relay.Constant][0] /* ty=Tensor[(6, 3, 5, 5), float32] */, padding=[2, 2, 2, 2], kernel_size=[5, 5], out_dtype="float32") /* ty=Tensor[(1, 6, 10, 10), float32] */
}


[14:56:21] /home/mbs/github/mbs-tvm/src/relay/ir/transform.cc:159: AlterOpLayout: Output module:
def @main(%data: Tensor[(1, 3, 10, 10), float32]) -> Tensor[(1, 6, 10, 10), float32] {
  %0 = layout_transform(%data, src_layout="NCHW", dst_layout="NCHW16c");
  %1 = nn.conv2d(%0, meta[relay.Constant][0] /* ty=Tensor[(6, 3, 5, 5), float32] */, padding=[2, 2, 2, 2], kernel_size=[5, 5], data_layout="NCHW16c", out_dtype="float32");
  layout_transform(%1, src_layout="NCHW16c", dst_layout="NCHW")
}

I can't find any matching rewrite on the GOOD runs -- they're all of the form:

[14:53:31] /home/mbs/github/mbs-tvm/src/relay/ir/transform.cc:135: AlterOpLayout: Input module:
def @main(%data: Tensor[(1, 3, 10, 10), float32]) -> Tensor[(1, 6, 10, 10), float32] {
  nn.conv2d(%data, meta[relay.Constant][0] /* ty=Tensor[(6, 3, 5, 5), float32] */, padding=[2, 2, 2, 2], kernel_size=[5, 5], out_dtype="float32") /* ty=Tensor[(1, 6, 10, 10), float32] */
}


[14:53:31] /home/mbs/github/mbs-tvm/src/relay/ir/transform.cc:159: AlterOpLayout: Output module:
def @main(%data: Tensor[(1, 3, 10, 10), float32]) -> Tensor[(1, 6, 10, 10), float32] {
  %0 = layout_transform(%data, src_layout="NCHW", dst_layout="NCHW3c");
  %1 = layout_transform(meta[relay.Constant][0] /* ty=Tensor[(6, 3, 5, 5), float32] */, src_layout="OIHW", dst_layout="OIHW3i3o");
  %2 = nn.contrib_conv2d_NCHWc(%0, %1, padding=[2, 2, 2, 2], channels=6, kernel_size=[5, 5], data_layout="NCHW3c", kernel_layout="OIHW3i3o", out_layout="NCHW3c", out_dtype="float32");
  layout_transform(%2, src_layout="NCHW3c", dst_layout="NCHW")
}

Eh?

[zxybazh]

So the only difference is dst_layout of NCHW3c v.s. NCHW16c?

[mbs-octoml]

As expected all is well if disable AlterOpLayout. I need to log whatever hidden state is driving that rewrite.