f.write(msg)
                sys.exit(1)

            def check_result(completed_process):
                if completed_process.returncode:
                    log_failure(f"Command failed: {{' '.join(completed_process.args)}}")
                return completed_process

            # =============================================================================
            # == Check that subprocess matches parent =====================================
            # =============================================================================
            if os.path.realpath(sys.executable) != "{parent_interpreter}":
                log_failure(
                    "Interpreter mismatch:\n"
                    f"  {{os.path.realpath(sys.executable)}}\n    vs.\n  {parent_interpreter}"
                )

            if torch.__file__ != "{torch_file}":
                log_failure(
                    "PyTorch does not match expected file:\n"
                    f"  {{torch.__file__}}\n    vs.\n  {torch_file}"
                )

            # =============================================================================
            # == User specified setup =====================================================
            # =============================================================================
            # Load serialized globals
            {load_globals}

            # User setup str
            {setup}

            for _ in range({warmup_number}):
            {indented_stmt}

            # =============================================================================
            # == Callgrind management =====================================================
            # =============================================================================
            with open("{stat_log}", "wb") as stat_file:
                # If many instances of callgrind are running at once, the output of
                # `callgrind_control` may exceed 16kb which would cause `subprocess.PIPE`
                # to deadlock. So instead we use a file.
                callgrind_stat = check_result(subprocess.run(
                    ["callgrind_control", "--stat"],
                    stdout=stat_file,
                    stderr=subprocess.STDOUT,
                ))

            with open("{stat_log}", "rt") as stat_file:
                stat_lines = stat_file.read().splitlines()

            if f"PID {{PID}}: python {{__file__}}" not in stat_lines:
                log_failure("Process does not appear to be running callgrind.")

            gc.collect()
            time.sleep(0.01)

            # =============================================================================
            # == User code block ==========================================================
            # =============================================================================
            for _ in range({repeats}):
                callgrind_bindings._valgrind_toggle()
            {blocked_stmt}
                callgrind_bindings._valgrind_toggle_and_dump_stats()
                gc.collect()

            {baseline}
        r