NeMo-Speech.cpp/docs/nmt/models.md at main · NVIDIA/NeMo-Speech.cpp

GitHub

NMT models and conversion

The NMT pipeline runs a Riva-Translate decoder (a Mistral/Llama-architecture text model) through llama.cpp. The model is a standard Hugging Face checkpoint, not a NeMo checkpoint; the root converter delegates this architecture to the pinned llama.cpp converter.

Hugging Face:

nvidia/Riva-Translate-4B-Instruct-v2

Convert

Conversion is run from a source checkout; the Python tools are not included in native release archives. Install the root conversion requirements and use the unified converter. See

Model conversion

for the base environment setup.

python3 -m pip install -r requirements.txt python3 convert_model.py nvidia/Riva-Translate-4B-Instruct-v2 \ --outfile riva-translate-4b-instruct-v2.q8_0.gguf --outtype q8_0The converter initializes the pinned llama.cpp source automatically. A local Hugging Face checkpoint directory can be passed instead of the repository ID.

Notes

The language pair is selected per request, not baked into the model: see

configuration

.

The underlying llama.cpp converter may print a fix_mistral_regex warning while reading the Hugging Face tokenizer. It is benign for this model.

Precision

The converter accepts f32, f16, bf16, and q8_0; f16 and q8_0 are the tested runtime precisions on CPU and GPU backends, and q8_0 is smaller and the example-config default.

Once converted, point the server at it: see

NMT configuration

.