Error referencing float4_e2m1fn in ml_dtypes when loading an ASR model #15226
Description
Describe the bug
Running pip install nemo_toolkit["asr"] or pip install -U nemo_toolkit["asr"] works, but loading a model fails e.g.:
import nemo.collections.asr as nemo_asr
model_name = 'nvidia/parakeet-tdt-0.6b-v2'
asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name=model_name)
with the error:
AttributeError: module 'ml_dtypes' has no attribute 'float4_e2m1fn'. Did you mean: 'float8_e4m3fn'?
after running pip install "ml_dtypes >= 0.5" I get:
ImportError: Numba needs NumPy 2.3 or less. Got NumPy 2.4.
so the fix is to run:
pip install nemo_toolkit["asr"]
pip install "ml_dtypes >= 0.5" "numpy == 2.3.0"
The nemo dependencies should be updated to reference these version requirements.
Steps/Code to reproduce bug
python3 -m venv .venv
source .venv/bin/activate
pip install nemo_toolkit["asr"]
python3 test.py
deactivate
Expected behavior
The model loads without error.
Environment overview (please complete the following information)
- Environment location: Bare-metal
- Method of NeMo install:
pip install nemo_toolkit["asr"]
Environment details
- OS version: Ubuntu 25.10
- PyTorch version: 2.9.1+cu128
- Python version: 3.13.7
Additional context
n/a