Set is_binarized to false for valid dataloaders.

src/fairseq2/checkpoint/_manager.py

@@ -1,3 +1,3 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
 #
@@ -13,14 +13,12 @@
 from os import scandir
 from pathlib import Path
 from shutil import Error
-from typing import Protocol, TypeAlias, cast, final
+from typing import cast, final, Protocol, TypeAlias
 
 import torch
-from torch import Tensor
-from typing_extensions import override
 
 from fairseq2.error import InternalError
-from fairseq2.gang import Gang, GangError, Gangs, all_sum
+from fairseq2.gang import all_sum, Gang, GangError, Gangs
 from fairseq2.nn.data_parallel import load_with_sdp_gang
 from fairseq2.typing import CPU
 from fairseq2.utils.file import (
@@ -33,6 +31,8 @@
 )
 from fairseq2.utils.state import Stateful
 from fairseq2.utils.threading import ThreadPool
+from torch import Tensor
+from typing_extensions import override
 
 
 class CheckpointManager(ABC):
@@ -872,7 +872,7 @@
         if keep_last_n is None and keep_best_n is None and keep_every_n_steps is None:
             return False
 
-        step_numbers = self.get_step_numbers()
+        step_numbers = self.get_step_numbers(exclude_model_only=True)
         if not step_numbers:
             return False
 

src/fairseq2/recipes/wav2vec2/_train.py

@@ -269,6 +269,7 @@ def load_wav2vec2_trainer(
     # Initialize the validation unit.
     if config.dataset.valid_split is not None:
 
+        config.dataset.extras["is_binarized"] = False
         read_options = SpeechReadOptions(
             batching=batching,
             dtype=config.trainer.dtype,