Llama.cpp imatrix quantizations of nvidia/Llama-3.1-Minitron-4B-Width-Base

Using llama.cpp commit 2e59d61 for quantization.

Original model: https://huggingface.co/nvidia/Llama-3.1-Minitron-4B-Width-Base

All quants were made using the imatrix option and Bartowski's calibration file.

Perplexity table (the lower the better)

Quant	Size (MB)	PPL	Size (%)	Accuracy (%)	PPL error rate
IQ1_S	1158	81.7502	13.44	9.26	0.69197
IQ1_M	1227	40.8601	14.24	18.53	0.31979
IQ2_XXS	1343	16.5816	15.59	45.67	0.11466
IQ2_XS	1448	13.024	16.81	58.14	0.08768
IQ2_S	1551	12.6045	18	60.07	0.08478
IQ2_M	1643	11.0911	19.07	68.27	0.07374
Q2_K_S	1654	11.0796	19.2	68.34	0.07646
Q2_K	1755	10.3111	20.37	73.44	0.07045
IQ3_XXS	1794	9.342	20.82	81.05	0.0612
IQ3_XS	1934	9.4403	22.45	80.21	0.06137
Q3_K_S	2005	8.8949	23.27	85.13	0.05946
IQ3_S	2017	9.0714	23.41	83.47	0.05851
IQ3_M	2083	8.3352	24.18	90.84	0.0534
Q3_K_M	2191	8.1839	25.43	92.52	0.05408
Q3_K_L	2351	8.093	27.29	93.56	0.05352
IQ4_XS	2419	7.774	28.08	97.4	0.05097
Q4_0	2533	7.8479	29.4	96.49	0.05132
IQ4_NL	2538	7.7697	29.46	97.46	0.05091
Q4_K_S	2541	7.8125	29.49	96.92	0.05101
Q4_K_M	2650	7.7376	30.76	97.86	0.05038
Q4_1	2772	7.8155	32.17	96.89	0.05116
Q5_K_S	3017	7.6649	35.02	98.79	0.05021
Q5_0	3024	7.6407	35.1	99.1	0.0499
Q5_K_M	3081	7.6283	35.76	99.26	0.04985
Q5_1	3263	7.6439	37.87	99.06	0.04996
Q6_K	3539	7.587	41.07	99.8	0.04945
Q8_0	4581	7.5739	53.17	99.98	0.04941
F16	8616	7.5721	100	100	0.04942

Llama-3.1-Minitron-4B-Width-Base

Model Overview

Llama-3.1-Minitron-4B-Width-Base is a base text-to-text model that can be adopted for a variety of natural language generation tasks. It is obtained by pruning Llama-3.1-8B; specifically, we prune model embedding size, number of attention heads, and MLP intermediate dimension. Following pruning, we perform continued training with distillation using 94 billion tokens to arrive at the final model; we use the continuous pre-training data corpus used in Nemotron-4 15B for this purpose.

This model is ready for commercial use.

Model Developer: NVIDIA

Model Dates: Llama-3.1-Minitron-4B-Width-Base was trained between July 29, 2024 and Aug 3, 2024.

License

This model is released under the NVIDIA Open Model License Agreement.

Model Architecture

Llama-3.1-Minitron-4B-Width-Base uses a model embedding size of 3072, 32 attention heads, MLP intermediate dimension of 9216, with 32 layers in total. Additionally, it uses Grouped-Query Attention (GQA) and Rotary Position Embeddings (RoPE).

Architecture Type: Transformer Decoder (Auto-Regressive Language Model)

Network Architecture: Llama-3.1

Input Type(s): Text

Input Format(s): String

Input Parameters: None

Other Properties Related to Input: Works well within 8k characters or less.

Output Type(s): Text

Output Format: String

Output Parameters: 1D

Other Properties Related to Output: None

Usage

Pull requests to support this model in Hugging Face Transformers are currently under review (#32495 and #32502) and are expected to be merged soon. In the meantime, please follow the installation instructions below:

# Fetch PR 32502
$ git clone -b suhara/llama-kv-channels --single-branch https://github.com/suhara/transformers.git && cd transformers

# Fetch changes from PR 32495
$ git fetch https://github.com/suiyoubi/transformers.git aot/head_dim_rope && git cherry-pick FETCH_HEAD --strategy-option theirs

# Install transformers
$ pip install -e .

We can now run inference on this model:

import torch
from transformers import AutoTokenizer, LlamaForCausalLM

# Load the tokenizer and model
model_path = "nvidia/Llama3.1-Minitron-4B-Width-Base"
tokenizer = AutoTokenizer.from_pretrained(model_path)

device = 'cuda'
dtype = torch.bfloat16
model = LlamaForCausalLM.from_pretrained(model_path, torch_dtype=dtype, device_map=device)

# Prepare the input text
prompt = 'Complete the paragraph: our solar system is'
inputs = tokenizer.encode(prompt, return_tensors='pt').to(model.device)

# Generate the output
outputs = model.generate(inputs, max_length=20)

# Decode and print the output
output_text = tokenizer.decode(outputs[0])
print(output_text)

Software Integration

Runtime Engine(s):

NeMo 24.05

Supported Hardware Microarchitecture Compatibility:

NVIDIA Ampere
NVIDIA Blackwell
NVIDIA Hopper
NVIDIA Lovelace

[Preferred/Supported] Operating System(s):

Linux

Dataset & Training

Data Collection Method by Dataset: Automated

Labeling Method by Dataset: Not Applicable

Properties: The training corpus for Llama-3.1-Minitron-4B-Width-Base consists of English and multilingual text, as well as code. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. The corpus spans domains including legal, math, science, finance, and more. In our continued training set, we introduce a small portion of question-answering, and alignment style data to improve model performance.

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

Evaluation Results

Overview

5-shot performance. Language Understanding evaluated using Massive Multitask Language Understanding:

Average
60.5

Zero-shot performance. Evaluated using select datasets from the LM Evaluation Harness with additions:

HellaSwag	Winogrande	GSM8K	ARC-Challenge	XLSum
76.1	73.5	41.2	55.6	28.7

Code generation performance. Evaluated using MBPP:

Score
32.0

Inference

Engine: TensorRT-LLM

Test Hardware: NVIDIA A100

DType: BFloat16

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.

Please report security vulnerabilities or NVIDIA AI Concerns here.

References

Compact Language Models via Pruning and Knowledge Distillation