import argparse
import os
import json
import torch
import safetensors.torch

def per_tensor_quantize(tensor):
    """Quantize a tensor to FP8 using per-tensor static scaling factor."""
    finfo = torch.finfo(torch.float8_e4m3fn)
    if tensor.numel() == 0:
        min_val, max_val = torch.tensor(-16.0, dtype=tensor.dtype), torch.tensor(16.0, dtype=tensor.dtype)
    else:
        min_val, max_val = tensor.aminmax()
    amax = torch.maximum(min_val.abs(), max_val.abs())
    scale = finfo.max / amax.clamp(min=1e-12)
    qweight = (tensor * scale).clamp(min=finfo.min, max=finfo.max).to(torch.float8_e4m3fn)
    scale = scale.float().reciprocal()
    return qweight, scale

def process_safetensors_file(file_path):
    """Process a single safetensors file in-place, quantizing weights to FP8."""
    print(f"Processing {file_path}")
    tensors = safetensors.torch.load_file(file_path)
    
    modified_tensors = {}
    for name, tensor in tensors.items():
        if name.endswith('_proj.weight'):
            print("Quantizing", name)
            qweight, scale = per_tensor_quantize(tensor)
            modified_tensors[name] = qweight
            modified_tensors[f"{name}_scale"] = scale
        else:
            modified_tensors[name] = tensor

    safetensors.torch.save_file(modified_tensors, file_path)
    print(f"Updated {file_path} with quantized tensors")

def update_index_file(index_file_path):
    """Update the index file for the quantized model."""
    print(f"Updating index file: {index_file_path}")
    with open(index_file_path, 'r') as f:
        index = json.load(f)
    
    new_weight_map = {}
    for tensor_name, file_name in index['weight_map'].items():
        new_weight_map[tensor_name] = file_name
        if tensor_name.endswith('_proj.weight'):
            new_weight_map[f"{tensor_name}_scale"] = file_name
    
    index['weight_map'] = new_weight_map
    
    # Recalculate total_size
    total_size = sum(os.path.getsize(os.path.join(os.path.dirname(index_file_path), file)) 
                     for file in set(index['weight_map'].values()))
    index['metadata']['total_size'] = total_size
    
    with open(index_file_path, 'w') as f:
        json.dump(index, f, indent=2)
    print(f"Updated index file {index_file_path}")

def process_directory(directory):
    """Process all safetensors files in the given directory."""
    for filename in os.listdir(directory):
        file_path = os.path.join(directory, filename)
        if filename.endswith('.safetensors'):
            process_safetensors_file(file_path)
        elif filename == 'model.safetensors.index.json':
            index_file_path = file_path

    update_index_file(index_file_path)

if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Convert safetensors model to FP8 in-place.')
    parser.add_argument('directory', type=str, help='The directory containing the safetensors files and index file.')
    
    args = parser.parse_args()
    process_directory(args.directory)

Nemotron-4-340B-Instruct

Model Overview

Nemotron-4-340B-Instruct is a large language model (LLM) that can be used as part of a synthetic data generation pipeline to create training data that helps researchers and developers build their own LLMs. It is a fine-tuned version of the Nemotron-4-340B-Base model, optimized for English-based single and multi-turn chat use-cases. It supports a context length of 4,096 tokens.

The base model was pre-trained on a corpus of 9 trillion tokens consisting of a diverse assortment of English based texts, 50+ natural languages, and 40+ coding languages. Subsequently the Nemotron-4-340B-Instruct model went through additional alignment steps including:

Supervised Fine-tuning (SFT)
Direct Preference Optimization (DPO)
Reward-aware Preference Optimization (RPO) (Additional in-house alignment technique)

Throughout the alignment process, we relied on only approximately 20K human-annotated data while our data generation pipeline synthesized over 98% of the data used for supervised fine-tuning and preference fine-tuning (DPO & RPO). We provide comprehensive details about our synthetic data generation pipeline in the technical report.

This results in a model that is aligned for human chat preferences, improvements in mathematical reasoning, coding and instruction-following, and is capable of generating high quality synthetic data for a variety of use cases.

Under the NVIDIA Open Model License, NVIDIA confirms:

Models are commercially usable.
You are free to create and distribute Derivative Models.
NVIDIA does not claim ownership to any outputs generated using the Models or Derivative Models.

License:

NVIDIA Open Model License

Intended use

Nemotron-4-340B-Instruct is a chat model intended for use for the English language.

Nemotron-4-340B-Instruct is designed for Synthetic Data Generation to enable developers and enterprises for building and customizing their own large language models and LLM applications.

The instruct model itself can be further customized using the NeMo Framework suite of customization tools including Parameter-Efficient Fine-Tuning (P-tuning, Adapters, LoRA, and more), and Model Alignment (SFT, SteerLM, RLHF, and more) using NeMo-Aligner.

Model Developer: NVIDIA

Model Dates: Nemotron-4-340B-Instruct was trained between December 2023 and May 2024.

Data Freshness: The pretraining data has a cutoff of June 2023.

Required Hardware

BF16 Inference:

8x H200 (1x H200 node)
16x H100 (2x H100 nodes)
16x A100 80GB (2x A100 80GB nodes)

Model Architecture:

Nemotron-4-340B-Instruct is standard decoder-only Transformer, trained with a sequence length of 4096 tokens, uses Grouped-Query Attention (GQA), and Rotary Position Embeddings (RoPE).

Architecture Type: Transformer Decoder (auto-regressive language model)

Network Architecture: Nemotron-4

Prompt Format

Note: For Nemotron-4-340B-Instruct we recommend keeping the system prompt empty.

Single Turn

text

<extra_id_0>System

<extra_id_1>User
{prompt}
<extra_id_1>Assistant

Multi-Turn or Few-shot

text

<extra_id_0>System

<extra_id_1>User
{prompt 1}
<extra_id_1>Assistant
{response 1}
<extra_id_1>User
{prompt 2}
<extra_id_1>Assistant
{response 2}
...
<extra_id_1>User
{prompt N}
<extra_id_1>Assistant

An example of a formattable prompt template is available in the following section.

Usage

Deployment and inference with Nemotron-4-340B-Instruct can be done in three steps using NeMo Framework:

Create a Python script to interact with the deployed model. Create a Bash script to start the inference server Schedule a Slurm job to distribute the model across 2 nodes and associate them with the inference server.

Define the Python script call_server.py

python

import json
import requests

headers = {"Content-Type": "application/json"}

def text_generation(data, ip='localhost', port=None):
    resp = requests.put(f'http://{ip}:{port}/generate', data=json.dumps(data), headers=headers)
    return resp.json()


def get_generation(prompt, greedy, add_BOS, token_to_gen, min_tokens, temp, top_p, top_k, repetition, batch=False):
    data = {
        "sentences": [prompt] if not batch else prompt,
        "tokens_to_generate": int(token_to_gen),
        "temperature": temp,
        "add_BOS": add_BOS,
        "top_k": top_k,
        "top_p": top_p,
        "greedy": greedy,
        "all_probs": False,
        "repetition_penalty": repetition,
        "min_tokens_to_generate": int(min_tokens),
        "end_strings": ["<|endoftext|>", "<extra_id_1>", "\x11", "<extra_id_1>User"],
    }
    sentences = text_generation(data, port=1424)['sentences']
    return sentences[0] if not batch else sentences

PROMPT_TEMPLATE = """<extra_id_0>System

<extra_id_1>User
{prompt}
<extra_id_1>Assistant
"""

question = "Write a poem on NVIDIA in the style of Shakespeare"
prompt = PROMPT_TEMPLATE.format(prompt=question)
print(prompt)

response = get_generation(prompt, greedy=True, add_BOS=False, token_to_gen=1024, min_tokens=1, temp=1.0, top_p=1.0, top_k=0, repetition=1.0, batch=False)
response = response[len(prompt):]
if response.endswith("<extra_id_1>"):
    response = response[:-len("<extra_id_1>")]
print(response)

Given this Python script, create a Bash script which spins up the inference server within the NeMo container (docker pull nvcr.io/nvidia/nemo:24.01.framework) and calls the Python script call_server.py. The Bash script nemo_inference.sh is as follows,

bash

NEMO_FILE=$1
WEB_PORT=1424

depends_on () {
    HOST=$1
    PORT=$2
    STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)
    while [ $STATUS -ne 0 ]
    do
         echo "waiting for server ($HOST:$PORT) to be up"
         sleep 10
         STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)
    done
    echo "server ($HOST:$PORT) is up running"
}


/usr/bin/python3 /opt/NeMo/examples/nlp/language_modeling/megatron_gpt_eval.py \
        gpt_model_file=$NEMO_FILE \
        pipeline_model_parallel_split_rank=0 \
        server=True tensor_model_parallel_size=8 \
        trainer.precision=bf16 pipeline_model_parallel_size=2 \
        trainer.devices=8 \
        trainer.num_nodes=2 \
        web_server=False \
        port=${WEB_PORT} &
    SERVER_PID=$!

    readonly local_rank="${LOCAL_RANK:=${SLURM_LOCALID:=${OMPI_COMM_WORLD_LOCAL_RANK:-}}}"
    if [ $SLURM_NODEID -eq 0 ] && [ $local_rank -eq 0 ]; then
        depends_on "0.0.0.0" ${WEB_PORT}

        echo "start get json"
        sleep 5

        echo "SLURM_NODEID: $SLURM_NODEID"
        echo "local_rank: $local_rank"
        /usr/bin/python3 /scripts/call_server.py
        echo "clean up dameons: $$"
        kill -9 $SERVER_PID
        pkill python
    fi
    wait

Launch nemo_inference.sh with a Slurm script defined like below, which starts a 2-node job for model inference.

#!/bin/bash
#SBATCH -A SLURM-ACCOUNT
#SBATCH -p SLURM-PARITION
#SBATCH -N 2
#SBATCH -J generation      
#SBATCH --ntasks-per-node=8   
#SBATCH --gpus-per-node=8
set -x

RESULTS=<PATH_TO_YOUR_SCRIPTS_FOLDER>
OUTFILE="${RESULTS}/slurm-%j-%n.out"
ERRFILE="${RESULTS}/error-%j-%n.out"
MODEL=<PATH_TO>/Nemotron-4-340B-Instruct
CONTAINER="nvcr.io/nvidia/nemo:24.01.framework"
MOUNTS="--container-mounts=<PATH_TO_YOUR_SCRIPTS_FOLDER>:/scripts,MODEL:/model"

read -r -d '' cmd <<EOF
bash /scripts/nemo_inference.sh /model
EOF

srun -o $OUTFILE -e $ERRFILE --container-image="$CONTAINER" $MOUNTS bash -c "${cmd}"

Evaluation Results

MT-Bench (GPT-4-Turbo)

Evaluated using MT-Bench judging by GPT-4-0125-Preview as described in Appendix H in the HelpSteer2 Dataset Paper

total	writing	roleplay	extraction	stem	humanities	reasoning	math	coding	turn 1	turn 2
8.22	8.70	8.70	9.20	8.75	8.95	6.40	8.40	6.70	8.61	7.84

IFEval

Evaluated using the Instruction Following Eval (IFEval) introduced in Instruction-Following Evaluation for Large Language Models.

Prompt-Strict Acc	Instruction-Strict Acc
79.9	86.1

MMLU

Evaluated using the Multi-task Language Understanding benchmarks as introduced in Measuring Massive Multitask Language Understanding.

MMLU 0-shot
78.7

GSM8K

Evaluated using the Grade School Math 8K (GSM8K) benchmark as introduced in Training Verifiers to Solve Math Word Problems.

GSM8K 0-shot
92.3

HumanEval

Evaluated using the HumanEval benchmark as introduced in Evaluating Large Language Models Trained on Code.

HumanEval 0-shot
73.2

MBPP

Evaluated using the MBPP Dataset as introduced in the Program Synthesis with Large Language Models.

MBPP 0-shot
75.4

Arena Hard

Evaluated using the Arena-Hard Pipeline from the LMSys Org.

Arena Hard
54.2

AlpacaEval 2.0 LC

Evaluated using the AlpacaEval 2.0 LC (Length Controlled) as introduced in the paper: Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators

AlpacaEval 2.0 LC
41.5

TFEval

Evaluated using the CantTalkAboutThis Dataset as introduced in the CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues.

Distractor F1	On-topic F1
81.7	97.7

Adversarial Testing and Red Teaming Efforts

The Nemotron-4 340B-Instruct model underwent safety evaluation including adversarial testing via three distinct methods:

Garak, is an automated LLM vulnerability scanner that probes for common weaknesses, including prompt injection and data leakage.
AEGIS, is a content safety evaluation dataset and LLM based content safety classifier model, that adheres to a broad taxonomy of 13 categories of critical risks in human-LLM interactions.
Human Content Red Teaming leveraging human interaction and evaluation of the models' responses.

Limitations

The model was trained on data that contains toxic language, unsafe content, and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards here. Please report security vulnerabilities or NVIDIA AI Concerns here.

mgoin/Nemotron-4-340B-Instruct-FP8-Dynamic