Add eval_drop_last flag to fix TE eval bug

llmfoundry/eval/datasets/in_context_learning_evaluation.py

@@ -1311,6 +1311,7 @@ def build_icl_dataloader(
     generation_kwargs: Dict,
     early_stopping_criteria: Optional[List[str]] = None,
     do_normalization: bool = True,
+    eval_drop_last: bool = False,
 ) -> DataSpec:
     """Factory method that builds the specific dataset for the specified.
 
@@ -1421,6 +1422,7 @@ def build_icl_dataloader(
             batch_size=effective_batchsize,
             sampler=sampler,
             collate_fn=dataset.collate_fn,
+            drop_last=eval_drop_last,
         ),
         device_transforms=None,
         get_num_samples_in_batch=dataset.get_num_samples_in_batch,
@@ -1532,6 +1534,7 @@ def get_icl_task_dataloader(
     generation_kwargs: Optional[Dict] = None,
     early_stopping_criteria: Optional[List[str]] = None,
     do_normalization: bool = True,
+    eval_drop_last: bool = False,
 ) -> Union[DataSpec, Dict[str, DataSpec]]:
     r"""Constructs a dataloader (or dataloaders if has_categories is True)
 
@@ -1656,6 +1659,7 @@ def get_icl_task_dataloader(
                 generation_kwargs=generation_kwargs,
                 early_stopping_criteria=early_stopping_criteria,
                 do_normalization=do_normalization,
+                eval_drop_last=eval_drop_last,
             )
         return result_dls
     else:
@@ -1681,4 +1685,5 @@ def get_icl_task_dataloader(
             generation_kwargs=generation_kwargs,
             early_stopping_criteria=early_stopping_criteria,
             do_normalization=do_normalization,
+            eval_drop_last=eval_drop_last,
         )

llmfoundry/utils/builders.py

@@ -65,6 +65,7 @@ def build_evaluators(
     device_eval_batch_size: int,
     icl_seq_len: int,
     icl_subset_num_batches: Optional[int],
+    eval_drop_last: bool = False,
 ) -> Tuple[List[Evaluator], List[str], Optional[EvalGauntlet]]:
 
     evaluators = []
@@ -85,6 +86,7 @@ def build_evaluators(
             device_eval_batch_size,
             icl_seq_len,
             icl_subset_num_batches,
+            eval_drop_last=eval_drop_last,
         )
         evaluators.extend(icl_evaluators)
 
@@ -150,13 +152,15 @@ def build_icl_data_and_gauntlet(
     device_eval_batch_size: int,
     icl_seq_len: int,
     icl_subset_num_batches: Optional[int] = None,
+    eval_drop_last: bool = False,
 ) -> Tuple[List[Evaluator], List[str], Optional[EvalGauntlet]]:
     icl_evaluators, logger_keys = build_icl_evaluators(
         icl_tasks_config,
         tokenizer,
         icl_seq_len,
         device_eval_batch_size,
         icl_subset_num_batches=icl_subset_num_batches,
+        eval_drop_last=eval_drop_last,
     )
     eval_gauntlet_cb = None
     if eval_gauntlet_config is not None:
@@ -503,6 +507,7 @@ def build_icl_evaluators(
     default_batch_size: int,
     destination_dir: Optional[str] = None,
     icl_subset_num_batches: Optional[int] = None,
+    eval_drop_last: bool = False,
 ) -> Tuple[List[Evaluator], List[str]]:
     if destination_dir is None:
         destination_dir = os.getcwd()
@@ -621,10 +626,16 @@ def _validate_cfg(icl_cfg: Dict[str, Any]):
                 generation_kwargs=icl_cfg.get('generation_kwargs', {}),
                 early_stopping_criteria=early_stopping_criteria,
                 do_normalization=icl_cfg.get('do_normalization', True),
+                eval_drop_last=eval_drop_last,
             )
             if 'has_categories' in icl_cfg and icl_cfg[
                 'has_categories'] and isinstance(dataloaders, dict):
                 for category in dataloaders.keys():
+                    if len(dataloaders[category].dataloader) == 0:
+                        log.warning(
+                            f'No data for {label}/{category}, skipping. May have been filtered out by eval_drop_last={eval_drop_last} and batch size={default_batch_size}.',
+                        )
+                        continue
                     logger_keys.extend([
                         f'metrics/{label}/{category}/{m}' for m in metric_names
                     ])
@@ -636,6 +647,11 @@ def _validate_cfg(icl_cfg: Dict[str, Any]):
                         ),
                     )
             else:
+                if len(dataloaders.dataloader) == 0:  # type: ignore
+                    log.warning(
+                        f'No data for {label}, skipping. May have been filtered out by eval_drop_last={eval_drop_last} and batch size={default_batch_size}.',
+                    )
+                    continue
                 logger_keys.extend([
                     f'metrics/{label}/{m}' for m in metric_names
                 ])

llmfoundry/utils/config_utils.py

@@ -131,6 +131,7 @@ class TrainConfig:
     eval_gauntlet_str: Optional[str] = None  # should not be set by the user
     icl_subset_num_batches: Optional[int] = None
     icl_seq_len: Optional[int] = None
+    eval_drop_last: Optional[bool] = False
 
     # Logging
     loggers: Optional[Dict[str, Any]] = None