Posts tagged Tuning
Table Question-Answering with TaPas
- 26 April 2024
Conventionally, the question-answering task is framed as a semantic parsing task where the question is translated to a full logical form that can be executed against the table to retrieve the correct answer. However, this requires a lot of annotated data, which can be expensive to acquire.
Multimodal (Visual and Language) understanding with LLaVA-NeXT
- 26 April 2024
LLaVa (Large Language And Vision Assistant) was introduced in 2023 and became a milestone for multimodal models. It combines a pretrained vision encoder and a pretrained LLM for general purpose visual and language understanding. In January 2024, LLaVa-NeXT was released, which boasts significant enhancements, including higher input’s visual resolution and improved logical reasoning and world knowledge.
Unlocking Vision-Text Dual-Encoding: Multi-GPU Training of a CLIP-Like Model
- 24 April 2024
In this blog, we will build a vision-text dual encoder model akin to CLIP and fine-tune it with the COCO dataset on AMD GPU with ROCm. This work is inspired by the principles of CLIP and the Hugging Face example. The idea is to train a vision encoder and a text encoder jointly to project the representation of images and their descriptions into the same embedding space, such that the text embeddings are located near the embeddings of the images they describe. The objective during training is to maximize the similarity between the embeddings of image and text pairs in the batch while minimizing the similarity of embeddings for incorrect pairs. The model achieves this by learning a multimodal embedding space. A symmetric cross entropy loss is optimized over these similarity scores.
Text Summarization with FLAN-T5
- 16 April 2024
In this blog, we showcase the language model FLAN-T5 and how to fine-tune it on a summarization task with HuggingFace in an AMD GPUs + ROCm system.
Instruction fine-tuning of StarCoder with PEFT on multiple AMD GPUs
- 16 April 2024
In this blog, we will show you how to fine-tune the StarCoder base model on AMD GPUs with an instruction-answer pair dataset so that it can follow instructions to generate code and answer questions. We will also show you how to use parameter-efficient fine-tuning (PEFT) to minimize the computation cost for the fine-tuning process.
Affinity part 2 - System topology and controlling affinity
- 16 April 2024
In Part 1 of the Affinity blog series, we looked at the importance of setting affinity for High Performance Computing (HPC) workloads. In this blog post, our goals are the following:
Affinity part 1 - Affinity, placement, and order
- 16 April 2024
Modern hardware architectures are increasingly complex with multiple sockets, many cores in each Central Processing Unit (CPU), Graphical Processing Units (GPUs), memory controllers, Network Interface Cards (NICs), etc. Peripherals such as GPUs or memory controllers will often be local to a CPU socket. Such designs present interesting challenges in optimizing memory access times, data transfer times, etc. Depending on how the system is built, hardware components are connected, and the workload being run, it may be advantageous to use the resources of the system in a specific way. In this article, we will discuss the role of affinity, placement, and order in improving performance for High Performance Computing (HPC) workloads. A short case study is also presented to familiarize you with performance considerations on a node in the Frontier supercomputer. In a follow-up article, we also aim to equip you with the tools you need to understand your system’s hardware topology and set up affinity for your application accordingly.
Enhancing LLM Accessibility: A Deep Dive into QLoRA Through Fine-tuning Llama 2 on a single AMD GPU
- 15 April 2024
Building on the previous blog Fine-tune Llama 2 with LoRA blog, we delve into another Parameter Efficient Fine-Tuning (PEFT) approach known as Quantized Low Rank Adaptation (QLoRA). The focus will be on leveraging QLoRA for the fine-tuning of Llama-2 7B model using a single AMD GPU with ROCm. This task, made possible through the use of QLoRA, addresses challenges related to memory and computing limitations. The exploration aims to showcase how QLoRA can be employed to enhance accessibility to open-source large language models.
Small language models with Phi-2
- 08 April 2024
Like many other LLMs, Phi-2 is a transformer-based model with a next-word prediction objective that is trained on billions of tokens. At 2.7 billion parameters, Phi-2 is a relatively small language model, but it achieves outstanding performance on a variety of tasks, including common sense reasoning, language understanding, math, and coding. For reference, GPT 3.5 has 175 billion parameters and the smallest version of LLaMA-2 has 7 billion parameters. According to Microsoft, Phi-2 is capable of matching or outperforming models up to 25 times larger due to more carefully curated training data and model scaling.
Scale AI applications with Ray
- 01 April 2024
Most machine-learning (ML) workloads today require multiple GPUs or nodes to achieve the performance or scale required by applications. However, scaling workloads beyond single node/single GPU workloads is difficult and require some expertise in distributed processing.
Large language model inference optimizations on AMD GPUs
- 15 March 2024
Large language models (LLMs) have transformed natural language processing and comprehension, facilitating a multitude of AI applications in diverse fields. LLMs have various promising use cases, including AI assistants, chatbots, programming, gaming, learning, searching, and recommendation systems. These applications leverage the capabilities of LLMs to provide personalized and interactive experiences, which enhances user engagement.
Building a decoder transformer model on AMD GPU(s)
- 12 March 2024
In this blog, we demonstrate how to run Andrej Karpathy’s beautiful PyTorch re-implementation of GPT on single and multiple AMD GPUs on a single node using PyTorch 2.0 and ROCm. We use the works of Shakespeare to train our model, then run inference to see if our model can generate Shakespeare-like text.
Question-answering Chatbot with LangChain on an AMD GPU
- 11 March 2024
LangChain is a framework designed to harness the power of language models for building cutting-edge applications. By connecting language models to various contextual sources and providing reasoning abilities based on the given context, LangChain creates context-aware applications that can intelligently reason and respond. In this blog, we demonstrate how to use LangChain and Hugging Face to create a simple question-answering chatbot. We also demonstrate how to augment our large language model (LLM) knowledge with additional information using the Retrieval Augmented Generation (RAG) technique, then allow our bot to respond to queries based on the information contained within specified documents.
Simplifying deep learning: A guide to PyTorch Lightning
- 08 February 2024
PyTorch Lightning is a higher-level wrapper built on top of PyTorch. Its purpose is to simplify and abstract the process of training PyTorch models. It provides a structured and organized approach to machine learning (ML) tasks by abstracting away the repetitive boilerplate code, allowing you to focus more on model development and experimentation. PyTorch Lightning works out-of-the-box with AMD GPUs and ROCm.
Using LoRA for efficient fine-tuning: Fundamental principles
- 05 February 2024
Low-Rank Adaptation of Large Language Models (LoRA) is used to address the challenges of fine-tuning large language models (LLMs). Models like GPT and Llama, which boast billions of parameters, are typically cost-prohibitive to fine-tune for specific tasks or domains. LoRA preserves pre-trained model weights and incorporates trainable layers within each model block. This results in a significant reduction in the number of parameters that need to be fine-tuned and considerably reduces GPU memory requirements. The key benefit of LoRA is that it substantially decreases the number of trainable parameters–sometimes by a factor of up to 10,000–leading to a considerable decrease in GPU resource demands.
Fine-tune Llama 2 with LoRA: Customizing a large language model for question-answering
- 01 February 2024
In this blog, we show you how to fine-tune Llama 2 on an AMD GPU with ROCm. We use Low-Rank Adaptation of Large Language Models (LoRA) to overcome memory and computing limitations and make open-source large language models (LLMs) more accessible. We also show you how to fine-tune and upload models to Hugging Face.
Pre-training BERT using Hugging Face & TensorFlow on an AMD GPU
- 29 January 2024
This blog explains an end-to-end process for pre-training the Bidirectional Encoder Representations from Transformers (BERT) base model from scratch using Hugging Face libraries with a TensorFlow backend for English corpus text (WikiText-103-raw-v1).
Pre-training BERT using Hugging Face & PyTorch on an AMD GPU
- 26 January 2024
This blog explains an end-to-end process for pre-training the Bidirectional Encoder Representations from Transformers (BERT) base model from scratch using Hugging Face libraries with a PyTorch backend for English corpus text (WikiText-103-raw-v1).
LLM distributed supervised fine-tuning with JAX
- 25 January 2024
In this article, we review the process for fine-tuning a Bidirectional Encoder Representations from Transformers (BERT)-based large language model (LLM) using JAX for a text classification task. We explore techniques for parallelizing this fine-tuning procedure across multiple AMD GPUs, then evaluate our model’s performance on a holdout dataset. For this, we use a (BERT)-base-cased transformer model with a General Language Understanding Evaluation (GLUE) benchmark dataset on multiple AMD GPUs.
Pre-training a large language model with Megatron-DeepSpeed on multiple AMD GPUs
- 24 January 2024
In this blog, we show you how to pre-train a GPT-3 model using the Megatron-DeepSpeed framework on multiple AMD GPUs. We also demonstrate how to perform inference on the text-generation task with your pre-trained model.