Add helpful pretty pretting summaries to torch for lldb debugging (#97101)

# Summary
Add support for pretty printing of tensors when using lldb similiar to what is currently available for gdb

<img width="772" alt="Screenshot 2023-03-18 at 6 20 34 PM" src="https://user-images.githubusercontent.com/32754868/226148687-b4e6cfe1-8be1-4657-9ebc-d134f697dd37.png">

<img width="254" alt="Screenshot 2023-03-18 at 6 20 43 PM" src="https://user-images.githubusercontent.com/32754868/226148690-caca6f76-d873-419e-b5e4-6bb403b3d179.png">

I changed it so to override the variable formatting instead of having to call a seperate command you can just do `print <tensor>`

I also add one for sizes
<img width="309" alt="Screenshot 2023-03-19 at 1 05 49 PM" src="https://user-images.githubusercontent.com/32754868/226206458-e3f0111b-6a97-4d75-8125-48455aa2cf43.png">

Last one:
<img width="815" alt="Screenshot 2023-03-19 at 1 39 23 PM" src="https://user-images.githubusercontent.com/32754868/226207687-20bd014f-9e0e-4c01-b2c8-190b7365aa70.png">

If you use the codelldb extension be sure to add:
    `"lldb.launch.initCommands": ["command source ${env:HOME}/.lldbinit"]`
    To your setttings .json

Pull Request resolved: https://github.com/pytorch/pytorch/pull/97101
Approved by: https://github.com/ngimel

.lldbinit

@@ -0,0 +1,17 @@
+# automatically load the pytorch_lldb extension.
+#
+# lldb automatically tries to load this file whenever it is executed from the
+# root of the pytorch repo, but by default it is not allowed to do so due to
+# security reasons. If you want to use pytorch_lldb, please add the following
+# line to your ~/.lldbinit (i.e., the .lldbinit file which is in your home
+# directory, NOT this file):
+#    settings set target.load-cwd-lldbinit true
+#    setting set escape-non-printables false
+#
+# Alternatively, you can manually load the pytorch_lldb  commands into your
+# existing lldb session by doing the following:
+#    (lldb) command script import tools/lldb/pytorch_lldb.py
+
+command script import tools/lldb/pytorch_lldb.py
+setting set escape-non-printables false
+type category enable torch

tools/lldb/pytorch_lldb.py

@@ -0,0 +1,97 @@
+from typing import Any
+
+import lldb  # type: ignore[import]
+
+
+def get_target() -> Any:
+    target = lldb.debugger.GetSelectedTarget()
+    if not target:
+        print("[-] error: no target available. please add a target to lldb.")
+        return None
+    return target
+
+
+class DisableBreakpoints:
+    """
+    Context-manager to temporarily disable all lldb breakpoints, useful if
+    there is a risk to hit one during the evaluation of one of our custom
+    commands
+    """
+
+    def __enter__(self) -> None:
+        target = get_target()
+
+        if target.DisableAllBreakpoints() is False:
+            print("[-] error: failed to disable all breakpoints.")
+
+    def __exit__(self, etype: Any, evalue: Any, tb: Any) -> None:
+        target = get_target()
+
+        if target.EnableAllBreakpoints() is False:
+            print("[-] error: failed to enable all breakpoints.")
+
+
+def IntArrayRef_summary(valobj: Any, internal_dict: Any, options: Any) -> str:
+    """Print human readable representation of c10::IntArrayRef"""
+    with DisableBreakpoints():
+        target = get_target()
+        tensor = valobj.GetName()
+        result = target.EvaluateExpression(
+            f"torch::gdb::int_array_ref_string({tensor})"
+        )
+        str_result = str(result)
+        str_result = str_result[str_result.find('"') + 1 : -1]
+        return str_result
+
+
+def DispatchKeyset_summary(valobj: Any, internal_dict: Any, options: Any) -> str:
+    """Print human readable representation of c10::DispatchKeyset"""
+    with DisableBreakpoints():
+        target = get_target()
+        keyset = valobj.GetName()
+        result = target.EvaluateExpression(
+            f"torch::gdb::dispatch_keyset_string({keyset})"
+        )
+        str_result = str(result)
+        str_result = str_result[str_result.find('"') + 1 : -1]
+        return str_result
+
+
+def Tensor_summary(valobj: Any, internal_dict: Any, options: Any) -> str:
+    """Print a human readable representation of the given at::Tensor.
+
+    at::Tensor instances do not have a C++ implementation of a repr method: in
+    pytorch, this is done by pure-Python code. As such, print <tensor>
+    internally creates a Python wrapper for the given tensor and call repr()
+    on it.
+    Usage:
+        print self
+    """
+    with DisableBreakpoints():
+        target = get_target()
+        tensor = valobj.GetName()
+        result = target.EvaluateExpression(f"torch::gdb::tensor_repr({tensor})")
+        str_result = str(result)
+        target.EvaluateExpression(f"(void)free({result.GetValue()})")
+        str_result = "\n" + str_result[str_result.find("tensor") : -1]
+        return str_result
+
+
+# And the initialization code to add your commands
+def __lldb_init_module(debugger: Any, internal_dict: Any) -> Any:
+    debugger.HandleCommand(
+        "type summary add c10::IntArrayRef -F pytorch_lldb.IntArrayRef_summary -w torch"
+    )
+    debugger.HandleCommand(
+        "type summary add c10::DispatchKeySet -F pytorch_lldb.DispatchKeyset_summary -w torch"
+    )
+    debugger.HandleCommand(
+        "type summary add at::Tensor -F pytorch_lldb.Tensor_summary -w torch"
+    )
+    print(
+        "Pretty Printing lldb summary for PyTorch AT types has been installed and is ready for use. "
+        "This category is enabled by default. To disable run: `type category disable torch`"
+    )
+    print(
+        "Usage:\n\tprint <at::tensor>\n\tprint <c10::IntArrayRef>\n\tprint <c10::DispatchKeySet>"
+    )

torch/csrc/utils.cpp

@@ -305,6 +305,18 @@ error:
   return nullptr;
 }
 
+std::string int_array_ref_string(at::IntArrayRef sizes) {
+  std::stringstream ss;
+  ss << sizes;
+  return ss.str();
+}
+
+std::string dispatch_keyset_string(c10::DispatchKeySet keyset) {
+  std::stringstream ss;
+  ss << keyset;
+  return ss.str();
+}
+
 } // namespace gdb
 } // namespace torch