NoPE-GPT-Small-Chat

NoPE GPT

NoPE GPT is a generative pretrained Transformer (GPT) language model with no positional embeddings (NoPE). Built using PyTorch and trained on HuggingFace's Fineweb, SmolTalk, and UltraFeedback datasets, NoPE GPT can answer questions, follow instructions, summarize documents, chat, and more. Best of all, the model weights and code are fully open-source for you to customize, improve upon, and share with the world.

Features

• No positional embeddings (NoPE): NoPE GPT aims to be a more parsimonious model by completely removing positional embeddings from the architecture. This allows for a variable context length without complex model surgery. Despite having no positional embeddings, NoPE GPT performs better at context length generalization than the best relative embeddings (ALiBi, RoPE, T5) offering good performance even when operating within 2X the trained context window.

• Low Memory Utilization: NoPE GPT lets you progressively employ training-time memory optimizations such as fully-sharded data-parallel (FSDP), activation checkpointing, mixed precision, and low-memory optimizer updates that allow you to train larger models on smaller hardware.

• Fully Open-source: Unlike closed-source LLMs, NoPE GPT provides both the model weights and the source code to train, fine-tune, export, and generate text from the model using your own hardware. With the help of the open-source software community, we aim to democratize access to AI and continually improve the models.

Recommended Pretraining Configurations

Below is a table of some recommended model pretraining configurations but feel free to experiment with settings on your own. See the model_sizing.ipynb notebook to estimate the memory and compute requirements for your model configuration.

We typically recommend a training block size (also referred to as context length) of between 1024 to 4096 for standard models and 4096 or higher for long-context applications such as conversational chat bots, retrieval augmented generation (RAG), and chain-of-thought (CoT) prompting a.k.a "reasoning" models.

Note: NoPE GPT can be trained using variable block sizes since the architecture does not depend on any discrete positional embeddings. This flexibility allows you to progressively extend the context window during training.

Install Project Dependencies

Project dependencies are specified in the requirements.txt file. You can install them with pip using the following command from the project root. We recommend using a virtual environment such as venv to keep package dependencies on your system tidy.

python -m venv ./.venv

source ./.venv/bin/activate

pip install -r requirements.txt

Pretraining

When we pre-train NoPE GPT we are focused on building a foundation of language and general knowledge to use as a base for further specialized training. The training objective is to predict the next token in a sample of text. It is a self-supervised form of training because the model learns from masked inputs of unsupervised data. The For the pretraining corpus we use the Fineweb dataset which consists of about 15T high-quality tokens gathered from the worldwide web. The dataset has been split into 3 subsets (10BT, 100BT, and 350BT versions) for training smaller models. If you'd like to start training right away, the default settings should work on most single-GPU systems with 12G of VRAM or more.

python pretrain.py

Note that it will take a while to download and pre-process the dataset the first time that the training script is run.

To customize the default "Small" architecture you can adjust the embedding_dimensions, num_attention_heads, num_hidden_layers, and feed_forward_ratio arguments of the pretraining script.

python pretrain.py --embedding_dimensions=4096 --num_attention_heads=64 --num_hidden_layers=48 --feed_forward_ratio=4

You can also adjust the batch_size, learning_rate, and gradient_accumulation_steps to suite your training setup.

python pretrain.py --batch_size=32 --learning_rate=0.01 --gradient_accumulation_steps=128

For distributed training, use PyTorch's torchrun extension to launch a distributed data parallel (DDP) session. The example below is for executing the training script on a single node with 8 individual GPUs.

torchrun --standalone --nnodes=1 --nproc-per-node=8 pretrain.py --batch_size=16 --gradient_accumulation_steps=128

Note that when training in data-parallel mode it's important that the gradient_accumulation_steps divides evenly into the world size for maximum performance. For example, if we have an 8 GPU cluster, we could perform 32 gradient accumulation steps in exactly 4 passes over the network.

Pretraining Arguments

Instruction-tuning

Instruction-tuning is a supervised training technique focused on developing specialized objectives such as chatting, text summarization, chain-of-thought, and prompt rewriting. The overall objective is still to predict the next token but the dataset has been curated for these more specialized objectives. In addition, we introduce three new special tokens (<|pad|>, <|im_start|> and <|im_end|>) that demarcate padding tokens and system, user, and assistant messages for use in the ChatML format. We use the SmolTalk and UltraFeedback datasets by HuggingFace as fine-tuning corpora because they include a broad range of training objectives such as conversation, instruction following, summarization, and human preference alignment.

Unlike pre-training, fine-tuning is not as resource intensive due to training much fewer parameters. The default arguments will work for most GPUs with 12G of VRAM or more.

python instruction-tune.py

To pick which dataset subsets to train on you can specify them in a comma-separated list like in the example below.

python instruction-tune.py --dataset_subsets=smol-magpie-ultra,smol-summarize,ultra-feedback

You can also adjust the batch_size, learning_rate, and gradient_accumulation_steps just like we did with pre-training.

python instruction-tune.py --batch_size=32 --learning_rate=0.01 --gradient_accumulation_steps=32

To adjust the number of trainable LoRA parameters as well as the strength of the LoRA and Dropout signals you can change the --rank, --alpha, and --dropout arguments respectively.

python instruction-tune.py --rank=4 --alpha=0.8 --dropout=0.1

Instruction-tuning Arguments

Training Dashboard

We use TensorBoard to capture and display pretraining events such as loss and gradient norm updates. To launch the dashboard server run the following command from the terminal.

tensorboard --logdir=./runs

Then navigate to the dashboard using your favorite web browser.

Text Generation

After pre-training, you can generate text from the model by running the generate.py script from the commandline. This inference script samples tokens from the model one at a time conditioned on a prompt and any previously generated tokens, together referred to as the context window. In the example below we are choosing to only sample from the top_k predicted tokens that have at least top_p cumulative probability mass when ordered descending by predicted probability.

python generate.py --temperature=0.4 --top_k=500 --top_p=0.9

Generation Arguments

Chatting

Once properly instruction-tuned you can use the chat script to hold back-and-forth conversations with the model. In addition, you can provide a custom system message that serves to focus activations for a particular tone or task. To start chatting, run the chat script like in the example below.

python chat.py

Since the chat script uses the same sampling technique as the generate.py script, you can use the same arguments to control the generation process such as --temperature and top_k.

python chat.py --temperature=0.8 --top_k=300

Chat Arguments

References:

• G. Penedo, et al. The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale, 38th Conference on Neural Information Processing Systems (NeurIPS 2024) Track on Datasets and Benchmarks.
• L. B. Allal, et al. SmolLM2 - with great data, comes great performance, 2024.
• A. Radford, et al. Language Models are Unsupervised Multitask Learners, OpenAI, 2019.
• T. Brown, et al. Language Models are Few-Shot Learners. OpenAI, 2020.
• A. Kazemnejad, et al. The Impact of Positional Encoding on Length Generalization in Transformers, 37th Conference on Neural Information Processing Systems (NeurIPS 2023).
• S. Rajbhandari, et al. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models, 2020.
• J. R. Hermans, et al. Accumulated Gradient Normalization, JMLR: Workshop and Conference Proceedings, 2017.
• T. Chen, et al. Training Deep Nets with Sublinear Memory Cost. MIT, 2019.
• B. Zhang, et al. Root Mean Square Layer Normalization. 33rd Conference on Neural Information Processing Systems, NeurIPS 2019.
• J. Kaplan, et al. Scaling Laws for Neural Language Models, OpenAI, 2020.
• J. Hoffman, et al. Training Compute-Optimal Large Language Models, Deep Mind, 2022.