fix: lazily import soundfile in datasets

nemo_rl/data/datasets/eval_datasets/mmau.py

@@ -18,10 +18,10 @@
 from typing import Any
 
 import numpy as np
-import soundfile as sf
 from datasets import Audio, load_dataset
 
 from nemo_rl.data.datasets.response_datasets.avqa import _resample_audio
+from nemo_rl.data.datasets.utils import read_audio
 from nemo_rl.data.interfaces import TaskDataSpec
 from nemo_rl.data.processors import vlm_hf_data_processor
 
@@ -58,7 +58,7 @@ def __init__(
     def format_data(self, data: dict[str, Any]) -> dict[str, Any]:
         """Convert a raw MMAU item into messages format for vlm_hf_data_processor."""
         audio_raw = data["audio"]
-        audio_array, orig_sr = sf.read(io.BytesIO(audio_raw["bytes"]))
+        audio_array, orig_sr = read_audio(io.BytesIO(audio_raw["bytes"]))
 
         # Convert to mono if stereo
         if audio_array.ndim > 1:

nemo_rl/data/datasets/response_datasets/audiomcq.py

@@ -17,13 +17,12 @@
 from typing import Any
 
 import numpy as np
-import soundfile as sf
 from datasets import Dataset, load_dataset
 from huggingface_hub import snapshot_download
 from scipy.signal import resample_poly
 
 from nemo_rl.data.datasets.raw_dataset import RawDataset
-from nemo_rl.data.datasets.utils import get_huggingface_cache_path
+from nemo_rl.data.datasets.utils import get_huggingface_cache_path, read_audio
 
 DEFAULT_TEMPLATE = (
     "{question} Please choose the answer from the following options: {choices}. "
@@ -165,7 +164,7 @@ def format_data(self, data: dict[str, Any]) -> dict[str, Any]:
                 f"(source_dataset={source!r}, audio_path={audio_path!r})."
             )
 
-        audio_array, orig_sr = sf.read(absolute_path)
+        audio_array, orig_sr = read_audio(absolute_path)
 
         # Mono downmix for multi-channel waveforms before resampling.
         if audio_array.ndim > 1:

nemo_rl/data/datasets/response_datasets/avqa.py

@@ -18,11 +18,11 @@
 from typing import Any
 
 import numpy as np
-import soundfile as sf
 from datasets import Audio, Dataset, load_dataset
 from scipy.signal import resample_poly
 
 from nemo_rl.data.datasets.raw_dataset import RawDataset
+from nemo_rl.data.datasets.utils import read_audio
 
 DEFAULT_TEMPLATE = (
     "{question} Please choose the answer from the following options: {choices}. "
@@ -109,7 +109,7 @@ def __init__(
 
     def format_data(self, data: dict[str, Any]) -> dict[str, Any]:
         audio_raw = data["audio"]
-        audio_array, orig_sr = sf.read(io.BytesIO(audio_raw["bytes"]))
+        audio_array, orig_sr = read_audio(io.BytesIO(audio_raw["bytes"]))
 
         # Resample to 16kHz if needed
         if orig_sr != 16000:

nemo_rl/data/datasets/utils.py

@@ -51,6 +51,25 @@ def load_audio_from_file(path: str, sampling_rate: int = 16000) -> np.ndarray:
     return waveform.squeeze(0).numpy().astype(np.float32)
 
 
+def read_audio(source: Union[str, io.BytesIO]) -> tuple[np.ndarray, int]:
+    """Read an audio file or buffer as ``(samples, sample_rate)``.
+
+    ``soundfile`` is imported lazily because it dlopens ``libsndfile`` at import
+    time, so images that strip the bundled codecs can still import the audio
+    dataset modules and run text-only training.
+    """
+    try:
+        import soundfile as sf
+    except (ImportError, OSError) as e:
+        raise RuntimeError(
+            "Reading audio requires the 'soundfile' package and its bundled "
+            "libsndfile library, which are unavailable in this environment. "
+            "Install them with `uv pip install --reinstall soundfile`."
+        ) from e
+
+    return sf.read(source)
+
+
 def assert_no_double_bos(token_ids: torch.Tensor, tokenizer: TokenizerType) -> None:
     """Assert that there are no double starting BOS tokens in the message.