Training massive deep-learning models requires a balance of efficiency and scalability. In the context of the Transformers architecture, Mixture of Experts (MoE) models are massive machine learning architectures characterized for dividing tasks among multiple specialized sub-networks or “experts”. A gating network determines the expert to which a given input should be routed, enabling the model to handle complex tasks more efficiently by using the specialized capabilities of each expert. This dynamic routing mechanism allows MoE models to scale efficiently, activating only a subset of the network for each input, therefore reducing computational load while maintaining high model capacity.
AMD is excited to announce the integration of Google’s Gemma 3 models with AMD Instinct MI300X GPUs, optimized for high-performance inference using the vLLM framework. This collaboration empowers developers to harness advanced AMD AI hardware for scalable, efficient deployment of state-of-the-art language models. In this blog we will walk you through a step-by-step guide on deploying Google’s Gemma 3 model using vLLM on AMD Instinct GPUs, covering Docker setup, dependencies, authentication, and inference testing. Remember, the Gemma 3 model is gated—ensure you request access before beginning deployment.
As AI models continue to scale in size and complexity, deploying them efficiently requires strategic resource allocation. Tensor parallelism (TP) is a valuable technique for distributing workloads across multiple GPUs, reducing memory constraints, and enabling inference for large-scale models. However, the choice of TP configuration isn’t one-size-fits-all—it directly impacts performance, networking overhead, and cost efficiency.
As part of AMD’s newly released Instella family we are thrilled to introduce Instella-VL-1B, the first AMD vision language model for image understanding trained on AMD Instinct™ MI300X GPUs. Our journey with Instella-VL builds upon our previous 1-billion-parameter language models, AMD OLMo SFT. We further extend the language model’s visual understanding abilities by connecting it with a vision encoder (which is initialized from CLIP ViT-L/14-336). During training, we jointly finetune vision encoder and language model with vision-language data in three stages: Alignment, Pretraining and Supervised-Finetuning (SFT).
AMD is excited to announce Instella, a family of fully open state-of-the-art 3-billion-parameter language models (LMs) trained from scratch on AMD Instinct™ MI300X GPUs. Instella models outperform existing fully open models of similar sizes and achieve competitive performance compared to state-of-the-art open-weight models such as Llama-3.2-3B, Gemma-2-2B, and Qwen-2.5-3B, including their instruction-tuned counterparts.
Deploying Large Language Models (LLMs) in enterprise environments presents a multitude of challenges that organizations must navigate to harness their full potential. As enterprises expand their AI and HPC workloads, scaling the underlying compute and GPU infrastructure presents numerous challenges, including deployment complexities, resource optimization, and effective management of the compute resource fleet. In this blog, we will walk you through how to spin-up production-grade Serverless AI inference service on Kubernetes clusters by leveraging open source Knative/KServe technologies.
In this blog, we explore how DeepSeek-R1 achieves competitive performance on AMD Instinct™ MI300X GPUs, along with performance comparisons to H200 and a short demo application showcasing real-world usage. By leveraging MI300X, users can deploy DeepSeek-R1 and V3 models on a single node with impressive efficiency. In just two weeks, optimizations using SGLang have unlocked up to a 4X boost in inference speed, ensuring efficient scaling, lower latency, and optimized throughput. The MI300X’s high-bandwidth memory (HBM) and compute power enable execution of complex AI workloads, handling longer sequences and demanding reasoning tasks. With AMD and the SGLang community driving ongoing optimizations—including fused MoE kernels, MLA kernel fusion, and speculative decoding—MI300X is set to deliver an even more powerful AI inference experience.
In this blog you will learn the process of fine-tuning the Phi-3.5-mini-instruct Large Language Model (LLM) from Microsoft, using PyTorch in a multinode environment. The setup leverages the Hugging Face Accelerate library to handle the complexities of multi-GPU and multinode synchronization. Slurm is used to schedule and coordinate the job as a workload manager for high-performance computing environments. A custom Slurm Bash script launches the Docker containers on each node, ensuring the training environment is consistent across all machines. Inside the containers, PyTorch and the Accelerate library split the training data, synchronize the model updates, and optimize performance across the multinode setup. This approach lets you efficiently fine-tune large-scale models and reduce training time while maximizing hardware utilization across the entire cluster.
Kubernetes (often abbreviated as K8s) is an open-source platform designed for automating the deployment, scaling, and management of containerized applications. Developed by Google and now maintained by the Cloud Native Computing Foundation, Kubernetes enables developers to build, run, and manage applications across any infrastructure.
PyTorch Fully Sharded Data Parallel (FSDP) is a data parallelism technique that enables the training of large-scale models in a memory-efficient manner. FSDP achieves this memory efficiency by sharding model parameters, optimizer states, and/or gradients across GPUs, reducing the memory footprint required by each GPU. This enables the training of large-scale models with lower total GPU memory than DDP (Distributed Data Parallel), in which the model weights and optimizer states are replicated across all processes. To learn more about DDP, refer to Distributed Data Parallel (DDP) training on AMD GPU with ROCm.
Matrix multiplication underlies critical computational pathways in AI, with General Matrix Multiplication (GEMM) operations serving as performance-critical kernels in neural network architectures. From fully connected layers to convolutions and transformer attention mechanisms, GEMMs consume substantial computational and memory resources in large language models (LLMs). This blog explores GEMM optimization techniques for AMD GPUs, demonstrating methodologies to significantly enhance computational efficiency and performance scaling.
ROCm™ has emerged as a premier open software stack designed to address
the evolving needs of AI and machine learning workloads. Built for
inference and training, ROCm delivers leadership performance, empowering
developers and organizations to optimize their workloads for efficiency,
scalability, and cost-effectiveness.
Optimizing LLM performance on GPUs is challenging due to diverse model needs, memory constraints, and balancing latency and throughput. This document examines how hardware utilization, memory and communication bandwidth and scaling, contribute to inference performance, detailing optimal configurations for AMD Instinct™ MI300X GPUs.
Composer, developed by MosaicML, is an open-source deep learning training library built on top of PyTorch, designed to simplify and optimize distributed training workflows. It supports scalable training on multiple nodes and efficiently handles datasets of various sizes. Composer integrates advanced techniques such as PyTorch Fully Sharded Data Parallelism (FSDP), elastic sharded checkpointing, training callbacks, and speed-up algorithms to enhance training performance and flexibility. It closely resembles PyTorch’s torchrun and has demonstrated exceptional efficiency when scaling to hundreds of GPUs.
State Space Models (SSMs), such as Mamba, have emerged as a potential alternative to Transformer models. Vision backbones using only SSMs have yielded promising results. For more information about SSMs and Mamba’s performance on AMD hardware, see Mamba on AMD GPUs with ROCm.
This blog explores Vision Mamba (Vim), an innovative and efficient backbone for vision tasks and evaluate its performance on AMD GPUs with ROCm. We’ll start with a brief introduction to Vision Mamba, followed by a step-by-step guide on training and running inference with Vision Mamba on AMD GPUs using ROCm.
Triton Inference Server is an open-source platform designed to streamline AI inferencing. It supports the deployment, scaling, and inference of trained AI models from various machine learning and deep learning frameworks including Tensorflow, PyTorch, and vLLM, making it adaptable for diverse AI workloads. It is designed to work across multiple environments, including cloud, data centers and edge devices.
Image captioning, or the GenAI-based automatic generation of concise textual descriptions of images, has immensely important real-world applications. For example, image captioning can provide visually impaired users with textual descriptions of images for improved accessibility, image captioning can add textual descriptions to products in e-commerce applications and help children map images to their textual descriptions in early childhood educational apps. Image captioning can automatically describe objects and events in security camera footage in surveillance applications and can enable robots to auto-generate textual captions for objects and events they encountered in human-robot interaction (HRI) applications, and many more applications.
Image captioning is a sequence-to-sequence (seq2seq) machine learning task: a model converting a sequence from one domain (in this case, the image), to another (its textual description). In image captioning the image is partitioned into a sequence of patches. This sequence of image patches is then converted by the model to a corresponding sequence of text tokens.
In this blog post we will cover the bitsandbytes 8-bit representations. As you will see, the bitsandbytes 8-bit representations significantly help reduce the memory needed for fine-tuning and inferencing LLMs. There are many quantization techniques used in the field to decrease a model size, but bitsandbytes offers quantization to decrease the size of optimizer states as well. This post will help you understand the basic principles underlying the bitsandbytes 8-bit representations, explain the bitsandbytes 8-bit optimizer and LLM.int8 techniques, and show you how to implement these on AMD GPUs using ROCm.
With the increase in complexity and size of machine learning models, the demand for computational resources grows. Training on a single GPU can become a bottleneck for deep learning applications, especially with large datasets and models that are slow to train on a single GPU. Parallelized training addresses this challenge. Out of the various forms of parallelized training, this blog focuses on Distributed Data Parallel (DDP), a key feature in PyTorch that accelerates training across multiple GPUs and nodes.
This blog provides a thorough how-to guide on using Torchtune to fine-tune and scale large language models (LLMs) with AMD GPUs. Torchtune is a PyTorch library designed to let you easily fine-tune and experiment with LLMs. Using Torchtune’s flexibility and scalability, we show you how to fine-tune the Llama-3.1-8B model for summarization tasks using the EdinburghNLP/xsum dataset. Using LoRA(Low-Rank Adaptation), a parameter-efficient fine-tuning technique, Torchtune enables efficient training while maintaining performance across a different number of GPUs (2, 4, 6, and 8). This post also highlights how Torchtune’s distributed training capabilities allow users to scale up LLM fine-tuning on multiple GPUs to reduce training time while maintaining the quality of the trained model, demonstrating its potential and usage on modern AMD hardware using ROCm.
Transformer models have revolutionized natural language processing (NLP) by delivering high-performance results in tasks like machine translation, text summarization, text generation, and speech recognition. However, deploying these models in production can be challenging due to their high computational and memory requirements. CTranslate2 addresses these challenges by providing a custom runtime that implements various optimization techniques to accelerate Transformer models during inference.
Meta’s Llama models now support multimodal capabilities, expanding their functionality beyond traditional text-only applications. The Llama 3.2 models are available in a range of sizes, including medium-sized 11B and 90B multimodal models for vision-text reasoning tasks, and lightweight 1B and 3B text-only models designed for edge and mobile devices.
As the size of transformer models grow, so does the cost of conducting inference, impacting latency and throughput. Compression methods such as quantization and distillation, as well as hardware-aware optimizations such as Flash Attention and Triton, have been proposed to cut down the computation cost at different levels. However, these models either compromise on accuracy or require major changes to the model implementation.
As the scale and complexity of generative AI and deep learning models grow, multinode training, basically dividing a training job across several processors, has become an essential strategy to speed up training and fine-tuning processes of large generative AI models like SDXL. By distributing the training workload across multiple GPUs on multiple nodes, multinode setups can significantly accelerate the training process.
In this blog post we will show you, step-by step, how to set-up and fine-tune a Stable Diffusion XL (SDXL) model in a multinode Oracle Cloud Infrastructure’s (OCI) Kubernetes Engine (OKE) on AMD GPUs using ROCm.
In this blog, we’ll demonstrate the latest performance enhancements in vLLM inference on AMD Instinct accelerators using ROCm. In a nutshell, vLLM optimizes GPU memory utilization, allowing more efficient handling of large language models (LLMs) within existing hardware constraints, maximizing throughput and minimizing latency. We start the blog by briefly explaining how causal language models like Llama 3 and ChatGPT generate text, motivating the need to enhance throughput and reduce latency. If you’re new to vLLM, we also recommend reading our introduction to Inferencing and serving with vLLM on AMD GPUs.
ROCm 6.2 introduces support for the following vLLM features which we will use in this blog post.
Ready to supercharge your deep learning applications on AMD GPUs? In this blog, we’ll show you how to develop a custom fused dropout activation kernel for matrices in Triton, seamlessly call it from JAX, and benchmark its performance with ROCm. This powerful combination will take your model’s performance to the next level.
With the scale of large language models (LLMs) reaching hundred of billions of parameters, the ways we represent data within these enormous models dramatically impacts the resources required to train them (e.g. the number of GPUs needed for inference).
In our previous blogs (JAX mixed precision training; PyTorch AMP), we already demonstrated how mixed precision training can accelerate LLMs training process. In this blog post we will push things further and show you how quantization into an even lower precision data formats can speed up inference, saving time and memory, without sacrificing the overall performance of the model.
Quantization is a technique where the precision of a model’s parameters is reduced from a 32-bit floating point (FP32) or a 16-bit floating point (FP16) to an 8-bit integer (INT8). Standard models typically use 32-bit floating-point (FP32) precision. However, this higher precision is not always necessary for inference tasks. By converting model weights and activations to lower precision formats like INT8 (8-bit integer), we can achieve faster computations and lower memory usage, effectively reducing the model size by three-fourths (from 32-bit) or half (from 16-bit) with only a slight accuracy reduction, which is often outweighed by the speed gains.
Large Language Models (LLMs) have revolutionized the field of natural language processing, enabling machines to understand and generate human-like language. However, these models are often trained on vast amounts of general-purpose data, which can make them less effective for specific tasks or domains. Fine-tuning involves training a pre-trained LLM on a specialized dataset to enhance its performance on specific tasks. As Andrej Karpathy analogized, this process is akin to allowing someone to practice a particular skill. Just as a person might need to practice a skill in a specific context to become proficient, an LLM needs to be fine-tuned on a specific dataset to become proficient in a particular task. For instance, an LLM can be fine-tuned for tasks such as financial forecasting, technical support, legal advising, medical diagnosis, or even instruction following. By fine-tuning an LLM, organizations can achieve better results and improve information security by limiting the exposure of sensitive data.
In the rapidly evolving field of artificial intelligence, Large Language Models (LLMs) have emerged as powerful tools for understanding and generating human-like text. However, deploying these models efficiently at scale presents significant challenges. This is where vLLM comes into play. vLLM is an innovative open-source library designed to optimize the serving of LLMs using advanced techniques. Central to vLLM is PagedAttention, a novel algorithm that enhances the efficiency of the model’s attention mechanism by managing it as virtual memory. This approach optimizes GPU memory utilization, facilitating the processing of longer sequences and enabling more efficient handling of large models within existing hardware constraints. Additionally, vLLM incorporates continuous batching to maximize throughput and minimize latency. By leveraging these cutting-edge techniques, vLLM significantly improves the performance and scalability of LLM deployment, allowing organizations to harness the power of state-of-the-art AI models more effectively and economically.
Image classification is a key task in computer vision aiming at “understanding” an entire image. The outcome of an image classifier is a label or a category for the image as a whole, unlike object recognition where the task is to detect and classify multiple objects within an image.
Measuring the performance of new technologies is as old as human history, and often as intriguing (consider for example that we still compare the performance of new electric vehicle motors using horsepower). In the rapidly advancing field of machine learning (ML) MLPerf was established by MLCommons on May 2nd 2018 and quickly became the golden standard of measuring the accuracy, speed, and efficiency of AI. MLPerf provides benchmarks on training, HPC and Inference performance. Companies across the industry use MLPerf submissions to evaluate the performance of various GPUs and software platforms, and make their technology adoption decisions based on these results.
In this blog you will learn how to use ROCm, running on AMD’s Instinct GPUs, for a range of popular and useful natural language processing (NLP) tasks, using different large language models (LLMs). The blog includes a simple to follow hands-on guide that shows you how to implement LLMs for core NLP applications ranging from text generation and sentiment analysis to extractive question answering (QA), and solving a math problem.
Time series forecasting (TSF) is a key concept in fields such as signal processing, data science, and machine learning (ML). TSF involves predicting future behavior of a system by analyzing its past temporal patterns, using historical data to forecast future data points. Classical approaches to TSF relied on a variety of statistical methods. Recently, machine learning techniques have been increasingly used for TSF, generating discussions within the community about whether these modern approaches outperform the classical statistical ones (see: Are Transformers Effective for Time Series Forecasting? and Yes, Transformers are Effective for Time Series Forecasting (+ Autoformer)).
In this blog we explore how to fine-tune the Robustly Optimized BERT Pretraining Approach (RoBERTa) large language model, with emphasis on PyTorch’s mixed precision capabilities. Specifically, we explore using AMD GPUs for mixed precision fine-tuning to achieve faster model training without any major impacts on accuracy.
Graphs and graph analytics are related concepts that can help us understand complex
data and relationships. In this context, a graph is a mathematical model that represents entities
(called nodes or vertices) and their connections (called edges or links). And graph analytics
is a form of data analysis that uses graph structures and algorithms to reveal insights
from the data.
This blog provides a comprehensive guide on measuring and comparing the performance of various algorithms in a JAX-implemented generative AI model. Leveraging the JAX Profiler and statistical analysis, this blog demonstrates how to reliably evaluate key steps and compare algorithm performance on AMD GPUs.
PyTorch 2.0 introduces torch.compile(), a tool to vastly accelerate PyTorch code and models. By converting PyTorch code into highly optimized kernels, torch.compile delivers substantial performance improvements with minimal changes to the existing codebase. This feature allows for precise optimization of individual functions, entire modules, and complex training loops, providing a versatile and powerful tool for enhancing computational efficiency.
In this blog, we will show how to leverage PyTorch TunableOp to accelerate models using ROCm on AMD GPUs.
We will discuss the basics of General Matrix Multiplications (GEMMs), show an example of tuning a single GEMM, and finally, demonstrate real-world performance gains on an LLM (gemma) using TunableOp.
AI Voice agents, or voice bots, are designed to communicate with people using a spoken language. Voice bots are commonly deployed in customer service and personal assistant applications, and have the potential to enter and revolutionize almost every aspect of people’s interaction with technology that can benefit from the use of voice.
Automatic Speech Recognition (ASR), the technology that processes human speech into text, is essential for the creation of AI Voice agents. In this blog post we will provide you with a hands-on introduction to the deployment of three machine learning ASR models, using ROCm on AMD GPUs.
OpenLLM is an open-source platform designed to facilitate the deployment and utilization of large language models (LLMs), supporting a wide range of models for diverse applications, whether in cloud environments or on-premises. In this tutorial, we will guide you through the process of starting an LLM server using OpenLLM, enabling interaction with the server from your local machine, with special emphasis on leveraging the capabilities of AMD GPUs.
OpenAI has developed a powerful GPU focused programming language and compiler called Triton that works seamlessly with AMD GPUs. The goal of Triton is to enable AI engineers and scientists to write high-performant GPU code with minimal expertise. Triton kernels are performant because of their blocked program representation, allowing them to be compiled into highly optimized binary code. Triton also leverages Python for kernel development, making it both familiar and accessible. And the kernels can be easily compiled by simply declaring the triton.jit python decorator before the kernel.
As models increase in size, the time and memory needed to train them–and consequently, the cost–also increases. Therefore, any measures we take to reduce training time and memory usage can be highly beneficial. This is where Automatic Mixed Precision (AMP) comes in.