Llama-3.1-Nemotron-Ultra-253B-CPT-v1 is a large language model (LLM) which is a derivative of Meta Llama-3.1-405B-Instruct (AKA the parent model). This model supports a context length of 128K tokens and fits on a single 8xH100 node for inference. While the model was derived from Llama-3.1-405B-Instruct it has undergone a substantial continual pre-training (CPT). This CPT variant can be viewed as a “re-based” derivative of Llama-3.1-405B-Instruct.
Llama-3.1-Nemotron-Ultra-253B-v1 is a model which offers a great tradeoff between model accuracy and efficiency. Efficiency (throughput) directly translates to savings. Using a novel Neural Architecture Search (NAS) approach, we greatly reduce the model’s memory footprint, enabling larger workloads, as well as reducing the number of GPUs required to run the model in a data center environment. This NAS approach enables the selection of a desired point in the accuracy-efficiency tradeoff. Furthermore, by using a novel method to vertically compress the model (see details here), it also offers a significant improvement in latency.
This model served as the basis model for creating Llama-3.1-Nemotron-Ultra-253B-v1, which is part of the Llama Nemotron Collection. You can find the other models in this family here:
Architecture Type: Dense decoder-only Transformer model Network Architecture: Llama-3.1-405B-Instruct, customized through Neural Architecture Search (NAS)
**This model was developed based on Llama-3.1-405B-Instruct
** This model has 253B model parameters.
The model is a derivative of Llama 3.1-405B-Instruct, using Neural Architecture Search (NAS). The NAS algorithm results in non-standard and non-repetitive blocks. This includes the following:
Skip attention: In some blocks, the attention is skipped entirely, or replaced with a single linear layer.
Variable FFN: The expansion/compression ratio in the FFN layer is different between blocks.
FFN Fusion: When several consecutive attention layers are skipped, which can result in a sequence of multiple FFNs, that sequence of FFNs are fused into a smaller number of wider FFN layers.
For each block of the parent model, we created multiple variants providing different tradeoff profiles of quality vs. computational complexity, discussed in more depth here. We then search over the blocks to create a model which meets the required throughput and memory constraints while minimizing the quality degradation. To recover performance, the model initially underwent knowledge distillation (KD) for 65 billion tokens. This is followed by a continual pretraining (CPT) phase for 88 billion tokens. While the initial model was infused from the instruct version of Llama 3.1-405B-Instruct, the substantial CPT it has undergone can be viewed as “re-basing” the model. Still, however, the re-based CPT final model should keep some of its instruction following capabilities.
Intended use
Llama-3.1-Nemotron-Ultra-253B-CPT-v1 can be used as a base model intended to be used mainly in English and coding languages.
Input
Input Type: Text
Input Format: String
Input Parameters: One-Dimensional (1D)
Other Properties Related to Input: Context length up to 131,072 tokens
Output
Output Type: Text
Output Format: String
Output Parameters: One-Dimensional (1D)
Other Properties Related to Output: Context length up to 131,072 tokens
We recommend using the transformers package with version 4.48.3.
py
1import torch
2import transformers
34model_id ="nvidia/Llama-3_1-Nemotron-Ultra-253B-CPT-v1"5model_kwargs ={"torch_dtype": torch.bfloat16,"trust_remote_code":True,"device_map":"auto"}6tokenizer = transformers.AutoTokenizer.from_pretrained(model_id)7tokenizer.pad_token_id = tokenizer.eos_token_id
89pipeline = transformers.pipeline(10"text-generation",11 model=model_id,12 tokenizer=tokenizer,13 max_new_tokens=32768,14 do_sample=False,15**model_kwargs
16)1718print(pipeline("Hey how are you?"))# Base model usage19print(pipeline([{"role":"user","content":"Hey how are you?"}]))# Chat model usage
Inference:
Engine:
Transformers
Test Hardware:
BF16:
8x NVIDIA H100-80GB
4x NVIDIA B100
FP 8
4x NVIDIA H100-80GB
Training Datasets
A large variety of training data was used for the knowledge distillation phase before post-training pipeline, 3 of which included: FineWeb, Buzz-V1.2, and Dolma.
Data Collection for Training Datasets:
Hybrid: Automated, Human, Synthetic
Data Labeling for Training Datasets:
Hybrid: Automated, Human, Synthetic
Evaluation Datasets
We used the datasets listed in the next section to evaluate Llama-3.1-Nemotron-Ultra-253B-CPT-v1.
Data Collection for Evaluation Datasets:
Hybrid: Human/Synthetic
Data Labeling for Evaluation Datasets:
Hybrid: Human/Synthetic/Automatic
Evaluation Results
Benchmark
Metric
Score
GSM-8K
strict-match
84.99
MMLU
macro
88.09
MATH500
micro
80.4
HumanEval
pass@1
88.41
RULER
128K
83.21
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.
