Add warning for fused_tasks_and_states compute mode for 2D/List state tensors

[ge0405]

Summary:
I visited all child classes that use RecMetricComputation to see if any is incompatible with the added FUSED_TASKS_AND_STATES_COMPUTATION in D72010614.

As of Apr 16, 2025, searching "(RecMetricComputation" in fbcode resulted in 47 results.

• 1 is rec_metric in comment (not show in the following table)
• 42 metric classes defined in torchrec, 29 in OSS, 13 in FB
• 4 metric classes in customer codebase, e.g. MVAI, admarket (see row 43-46 in the table below)

RecMetricComputation uses state tensors to compute/update results. I looked through all following metrics' state tensors and see if any if them are (1) not Tensor (e.g. List) or (2) 2D Tensor. If so, when these metrics use FUSED_TASKS_AND_STATES_COMPUTATION, the init should show warning (and simply FUSED_TASKS_COMPUTATION will be used) or raise Exception (if these metrics can't allow any fuse mode).

| | dir | metric | type of state tensors | tests/warning |
|1 | oss | auc.py | List | warning |
|2 | oss | tower_qps.py | | |
|3 | oss | precision_session.py | not allow fuse, code | both fuse modes raise exception |
|4 | oss | serving_ne.py | | |
|5 | oss | recall_session.py | not allow fuse, code | both fuse modes raise exception |
|6 | oss | calibration.py | | |
|7 | oss | multiclass_recall.py | 2D tensor| warning |
|8 | oss | ndcg.py | | |
|9 | oss | gauc.py | | |
|10 | oss | tensor_weighted_avg.py | | |
|11 | oss | ne.py | | |
|12 | oss | serving_calibration.py | | |
|13 | oss | segmented_ne.py | 2D tensor| warning |
|14 | oss | scalar.py | | |
|15 | oss | mae.py | | |
|16 | oss | ne_positive.py | | |
|17 | oss | weighted_avg.py | | |
|18 | oss | output.py | | |
|20 | oss | cali_free_ne.py | | |
|21 | oss | unweighted_ne.py | | |
|22 | oss | hindsight_target_pr.py | 1D but not n_tasks, code | can still fuse states |
|23 | oss | rauc.py | List | warning |
|24 | oss | precision.py | | |
|25 | oss | recall.py | | |
|26 | oss | auprc.py | List| warning |
|27 | oss | mse.py | | |
|28 | oss | ctr.py | | |
|29 | oss | accuracy.py | | |
|30 | fb | log_normal_cnll.py | | |
|31 | fb | coarse_grained_multiclass_ne.py | 2D tensor| warning |
|32 | fb | regression_huber.py | | |
|33 | fb | modified_poisson_nll.py | | |
|34 | fb | res_ne.py | | |
|35 | fb | dist_shift.py | | |
|36 | fb | serving_ne.py | | |
|37 | fb | unjoined_calibration.py | | |
|38 | fb | unjoined_ne.py | | |
|39 | fb | serving_calibration.py | | |
|40 | fb | multiclass_ne.py | 2D tensor | warning |
|41 | fb | bucket_metric.py [1] | not allow fuse, code | both fuse modes raise exception |
|42 | fb | bucket_weighted_average_metric.py [2] | shouldn't allow fuse, code | both fuse modes should raise exception |
|43 | MVAI | metrics.py | List | |
|44 | MVAI | ndcg_metrics.py | | |
|45 | admarket | metrics.py | 2D or 3D? | |
|46 | mrs/fm | metrics.py | | |

[1] There are 7 bucket metrics (bucket_calibration, bucket_ctr, bucket_hindsight_target_pr, bucket_mse, bucket_ne, bucket_precision, bucket_recall) inherit BucketMetricComputation defined in bucket_metric.py. All of them have 2D tensors and most shapes are (n_tasks, num_buckets).

[2] There is 1 bucket metric (bucket_weighted_average_logloss) inherit BucketWeightedAverageMetricComputation defined in bucket_weighted_average_metric.py. All the state tensors are 2D (n_tasks, num_buckets).

Differential Revision: D73293593

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D73293593