Text Summarization using LLAMA2-70b

Dataset

The benchmark implementation run command will automatically download the validation and calibration datasets and do the necessary preprocessing. In case you want to download only the datasets, you can use the below commands.

Nvidia Preprocessed DatasetMLCommons Preprocessed DatasetUnprocessed Dataset

ValidationCalibration

LLAMA2-70b validation run uses the Open ORCA dataset.

Get Preprocessed Validation Dataset

mlcr get,dataset,preprocessed,openorca,_validation,_mlcommons,_nvidia -j

Get Preprocessed Calibration dataset

mlcr get,dataset,preprocessed,openorca,_calibration,_mlcommons,_nvidia -j

ValidationCalibration

LLAMA2-70b validation run uses the Open ORCA dataset.

Get Preprocessed Validation Dataset

mlcr get,dataset,preprocessed,openorca,_validation -j

Get Preprocessed Calibration dataset

mlcr get,dataset,preprocessed,openorca,_calibration -j

ValidationCalibration

LLAMA2-70b validation run uses the Open ORCA dataset.

Get Unprocessed Validation Dataset

mlcr get,dataset,openorca,_validation -j

Get Unprocessed Validation Dataset

mlcr get,dataset,openorca,_validation -j

• --outdirname=<PATH_TO_DOWNLOAD_OPENORCA_DATASET> could be provided to download the dataset to a specific location.

Model

The benchmark implementation run command will automatically download the required model and do the necessary conversions. In case you want to only download the official model, you can use the below commands.

Full-Precision ModelsQuantized Models

Pytorch

From MLCOMMONS storageFrom Hugging Face repo

Note: One has to accept the MLCommons Llama 2 License Confidentiality Notice to access the model files in MLCOMMONS storage.

Get the Official MLPerf LLAMA2-70B model from MLCOMMONS storage

mlcr get,ml-model,llama2-70b,_fp32,_pytorch -j

Note: Access to the HuggingFace model could be requested here.

Get model from HuggingFace repo

mlcr get,ml-model,llama2-70b,_hf --hf_token=<huggingface access token> -j

Using TensorRT-LLM

Round v5.1Round v5.0

Quantize Locally

Note: One has to accept the MLCommons Llama 2 License Confidentiality Notice to access the full precision model files in MLCOMMONS storage which are needed for quantization process.

mlcr get,ml-model,llama2-70b,_nvidia,_fp8,_v5.1 -j

• Use --checkpoint=<Full Precision model path> if model is already downloaded to a specific location.

Quantize LocallyPre-Quantized Model from MLCOMMONS Storage

Note: One has to accept the MLCommons Llama 2 License Confidentiality Notice to access the full precision model files in MLCOMMONS storage which are needed for quantization process.

mlcr get,ml-model,llama2-70b,_nvidia,_fp8,_v5.0 -j

• Use --checkpoint=<Full Precision model path> if model is already downloaded to a specific location.

Note: One has to accept the MLCommons Llama 2 License Confidentiality Notice to access the full precision model files and pre-quantized model files in MLCOMMONS storage.

mlcr get,ml-model,llama2-70b,_nvidia,_fp8,_v5.0,_pre-quantized -j

• Use --checkpoint=<Full Precision model path> if full precision model is already downloaded to a specific location.

• --outdirname=<PATH_TO_DOWNLOAD_LLAMA2_70B_MODEL> could be provided to download the model to a specific location.