pytorch

1
import textwrap
2
from typing import Any
3

4
import gdb  # type: ignore[import]
5

6

7
class DisableBreakpoints:
8
    """
9
    Context-manager to temporarily disable all gdb breakpoints, useful if
10
    there is a risk to hit one during the evaluation of one of our custom
11
    commands
12
    """
13

14
    def __enter__(self) -> None:
15
        self.disabled_breakpoints = []
16
        for b in gdb.breakpoints():
17
            if b.enabled:
18
                b.enabled = False
19
                self.disabled_breakpoints.append(b)
20

21
    def __exit__(self, etype: Any, evalue: Any, tb: Any) -> None:
22
        for b in self.disabled_breakpoints:
23
            b.enabled = True
24

25

26
class TensorRepr(gdb.Command):  # type: ignore[misc, no-any-unimported]
27
    """
28
    Print a human readable representation of the given at::Tensor.
29
    Usage: torch-tensor-repr EXP
30

31
    at::Tensor instances do not have a C++ implementation of a repr method: in
32
    pytorch, this is done by pure-Python code. As such, torch-tensor-repr
33
    internally creates a Python wrapper for the given tensor and call repr()
34
    on it.
35
    """
36

37
    __doc__ = textwrap.dedent(__doc__).strip()
38

39
    def __init__(self) -> None:
40
        gdb.Command.__init__(
41
            self, "torch-tensor-repr", gdb.COMMAND_USER, gdb.COMPLETE_EXPRESSION
42
        )
43

44
    def invoke(self, args: str, from_tty: bool) -> None:
45
        args = gdb.string_to_argv(args)
46
        if len(args) != 1:
47
            print("Usage: torch-tensor-repr EXP")
48
            return
49
        name = args[0]
50
        with DisableBreakpoints():
51
            res = gdb.parse_and_eval(f"torch::gdb::tensor_repr({name})")
52
            print(f"Python-level repr of {name}:")
53
            print(res.string())
54
            # torch::gdb::tensor_repr returns a malloc()ed buffer, let's free it
55
            gdb.parse_and_eval(f"(void)free({int(res)})")
56

57

58
TensorRepr()
59