Prevent from calling compute

[Boltzmachine]

The metric(preds, targets) will call compute automatically. But I do not want it because compute is time-consuming in my problem. So I only want to call it at the end of the evaluation epoch.

[Borda]

I understand that the automatic call to compute inside metric(preds, targets) can be costly and you want to avoid it until the end of the evaluation epoch.
By design, TorchMetrics updates internal state with the call to the metric and calls compute to get a final value. To achieve your goal, you should avoid using the metric as a callable during every batch. Instead, call the .update(preds, targets) method to update the metric state without triggering computation, then call .compute() only once when you need the final result (e.g., at the end of the epoch).

metric = YourMetric(...)
for batch in dataloader:
    metric.update(preds, targets)  # just update state
final_result = metric.compute()  # call compute once after epoch

This approach avoids repeated calls to compute while maintaining the correct metric calculation.
If your specific metric class does not behave this way or compute is still triggered on update, please check the metric implementation or share a code snippet for more targeted advice.