record function reproducer

add.py

import torch
import sys

device = torch.device('cpu')
left = torch.zeros(100, device=device, requires_grad=True)
right = torch.zeros(100, device=device, requires_grad=True)
grad = torch.zeros(100, device=device)

for _ in range(10):
    output = torch.add(left, right)
    output.backward(grad)

wrapper.cc

#include <torch/extension.h>
#include <iostream>

int driver_register() {
  at::addGlobalCallback(
    at::RecordFunctionCallback(
      [](const at::RecordFunction& fn)
      -> std::unique_ptr<at::ObserverContext> {
        std::cout << fn.forwardThreadId() << std::endl;
        return nullptr;
      },
      [](const at::RecordFunction& fn, at::ObserverContext* ctx_ptr) {
        return;
      })
    .needsInputs(false)   // TODO(Keren): monitor inputs if needed?
    .needsOutputs(false)  // TODO(Keren): monitor outputs if needed?
    .scopes({}));
  return 0;
}

int _ret = driver_register();

Just to consolidate the discussion here, @Jokeren said I always get fwd_thread_id = 18446744073709551615, which should be 0 or 1