pytorch

e9d0f5fb - Eliminate Named tensor warnings in XNNPACK and QNNPACK

Eliminate Named tensor warnings in XNNPACK and QNNPACK XNNPACK and QNNPACK propagate dimension names during allocation, but they never create named tensors from scratch. However, they were generating warnings ("Warning: Named tensors and all their associated APIs are an experimental feature...") from the call `namedinference::propagate_names_if_present_and_nonempty(t1, t2.names())` (because `t2.names()` returns a non-empty DimnameList with wildcards.) Introduce propagate_names_if_present_and_nonempty, which takes an `optional<DimnameList>` from `t2.opt_names()`, and use it from QNNPACK and XNNPACK to eliminate this warning. Another option would be to make `propagate_names_if_nonempty` take an optional, which would be more-or-less backward compatible since `propagate_names_if_nonempty(t1, t2.names())` would still be valid due to implicit conversion to optional. I chose not to do this just because it seemed riskier. Test Plan: For QNNPACK, just ran the NNAPI test, which was previously throwing a warning. For XNNPACK, its uses are gated on C10_MOBILE, so I needed to export a model. ```python import torch import torch.utils.bundled_inputs import torch.utils.mobile_optimizer import torch.nn.functional as F class MyModel(torch.nn.Module): def __init__(self): super().__init__() self.conv = torch.nn.Conv2d(1, 1, 1) self.convt = torch.nn.ConvTranspose2d(1, 1, 1) self.lin = torch.nn.Linear(1, 1) def forward(self, t): t = self.conv(t) t = self.convt(t) t = F.hardswish(t, False) t = F.hardswish(t[...,::2], True) t = F.adaptive_avg_pool2d(t, (1, 1)) t = torch.broadcast_to(t, (1, 4, 1, 1)) t = t.contiguous(memory_format=torch.channels_last) t = F.channel_shuffle(t, 2) t = F.max_pool2d(t.reshape(1, 1, 2, 2), (2, 2)) t = self.lin(t.reshape(1, 1).contiguous()) return t arg = torch.zeros(1,1,1,3) mtr = torch.jit.trace(MyModel().eval(), arg) mom = torch.utils.mobile_optimizer.optimize_for_mobile(mtr) mbi = torch.utils.bundled_inputs.bundle_inputs(mom, [((arg,))]) torch.jit.save(mbi, "/tmp/model.ptj") ``` ```bash nice env CMAKE_CXX_COMPILER_LAUNCHER=ccache CMAKE_C_COMPILER_LAUNCHER=ccache ./scripts/build_mobile.sh -DBUILD_BINARY=1 /build_mobile/bin/speed_benchmark_torch --model=/tmp/model.ptj --use_bundled_input=0 ``` Pull Request resolved: https://github.com/pytorch/pytorch/pull/77762 Approved by: https://github.com/zou3519