File: clear_undefinedness.h

package info (click to toggle)
pytorch-cuda 2.6.0%2Bdfsg-7
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid, trixie
  • size: 161,620 kB
  • sloc: python: 1,278,832; cpp: 900,322; ansic: 82,710; asm: 7,754; java: 3,363; sh: 2,811; javascript: 2,443; makefile: 597; ruby: 195; xml: 84; objc: 68
file content (22 lines) | stat: -rw-r--r-- 850 bytes parent folder | download | duplicates (3)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
#pragma once

#include <ATen/ATen.h>
#include <ATen/core/ivalue.h>
#include <ATen/core/jit_type.h>
#include <torch/csrc/Export.h>
#include <torch/csrc/jit/ir/ir.h>

namespace torch::jit {

// Undefinedness makes argument matching fail for regular tensor operations
// if 1+ arguments are undefined or possibly undefined tensors.
// Technically, undefined tensors are **not** tensors as the regular tensor
// operations do not know how to handle them.
// However, in practice, there are guards and conversion operators that
// **always** gate regular operations if undefined tensors may be present
// Eventually, we would love to move to the world where we use optionals
// in lieu of undefined tensors.
// When this happens, this pass will be removed
TORCH_API void ClearUndefinedness(const std::shared_ptr<Graph>& graph);

} // namespace torch::jit