Posts in category: Amazon AWS

As generative artificial intelligence (AI) inference becomes increasingly critical for businesses, customers are seeking ways to scale their generative AI operations or integrate generative AI models into existing workflows. Model optimization has emerged as a crucial step, allowing organizations to balance cost-effectiveness and responsiveness, improving productivity. However, price-performance requirements vary widely across use cases. For chat applications, minimizing latency is key to offer an interactive experience, whereas real-time applications like recommendations require maximizing throughput. Navigating these trade-offs poses a significant challenge to rapidly adopt generative AI, because you must carefully select and evaluate different optimization techniques.

To overcome these challenges, we are excited to introduce the inference optimization toolkit, a fully managed model optimization feature in Amazon SageMaker. This new feature delivers up to ~2x higher throughput while reducing costs by up to ~50% for generative AI models such as Llama 3, Mistral, and Mixtral models. For example, with a Llama 3-70B model, you can achieve up to ~2400 tokens/sec on a ml.p5.48xlarge instance v/s ~1200 tokens/sec previously without any optimization.

This inference optimization toolkit uses the latest generative AI model optimization techniques such as compilation, quantization, and speculative decoding to help you reduce the time to optimize generative AI models from months to hours, while achieving the best price-performance for your use case. For compilation, the toolkit uses the Neuron Compiler to optimize the model’s computational graph for specific hardware, such as AWS Inferentia, enabling faster runtimes and reduced resource utilization. For quantization, the toolkit utilizes Activation-aware Weight Quantization (AWQ) to efficiently shrink the model size and memory footprint while preserving quality. For speculative decoding, the toolkit employs a faster draft model to predict candidate outputs in parallel, enhancing inference speed for longer text generation tasks. To learn more about each technique, refer to Optimize model inference with Amazon SageMaker. For more details and benchmark results for popular open source models, see Achieve up to ~2x higher throughput while reducing costs by up to ~50% for generative AI inference on Amazon SageMaker with the new inference optimization toolkit – Part 1.

In this post, we demonstrate how to get started with the inference optimization toolkit for supported models in Amazon SageMaker JumpStart and the Amazon SageMaker Python SDK. SageMaker JumpStart is a fully managed model hub that allows you to explore, fine-tune, and deploy popular open source models with just a few clicks. You can use pre-optimized models or create your own custom optimizations. Alternatively, you can accomplish this using the SageMaker Python SDK, as shown in the following notebook. For the full list of supported models, refer to Optimize model inference with Amazon SageMaker.

Using pre-optimized models in SageMaker JumpStart

The inference optimization toolkit provides pre-optimized models that have been optimized for best-in-class cost-performance at scale, without any impact to accuracy. You can choose the configuration based on the latency and throughput requirements of your use case and deploy in a single click.

Taking the Meta-Llama-3-8b model in SageMaker JumpStart as an example, you can choose Deploy from the model page. In the deployment configuration, you can expand the model configuration options, select the number of concurrent users, and deploy the optimized model.

Deploying a pre-optimized model with the SageMaker Python SDK

You can also deploy a pre-optimized generative AI model using the SageMaker Python SDK in just a few lines of code. In the following code, we set up a ModelBuilder class for the SageMaker JumpStart model. ModelBuilder is a class in the SageMaker Python SDK that provides fine-grained control over various deployment aspects, such as instance types, network isolation, and resource allocation. You can use it to create a deployable model instance, converting framework models (like XGBoost or PyTorch) or Inference Specs into SageMaker-compatible models for deployment. Refer to Create a model in Amazon SageMaker with ModelBuilder for more details.

sample_input = {
    "inputs": "Hello, I'm a language model,",
    "parameters": {"max_new_tokens":128, "do_sample":True}
}

sample_output = [
    {
        "generated_text": "Hello, I'm a language model, and I'm here to help you with your English."
    }
]
schema_builder = SchemaBuilder(sample_input, sample_output)

builder = ModelBuilder(
    model="meta-textgeneration-llama-3-8b", # JumpStart model ID
    schema_builder=schema_builder,
    role_arn=role_arn,
)

List the available pre-benchmarked configurations with the following code:

builder.display_benchmark_metrics()

Choose the appropriate instance_type and config_name from the list based on your concurrent users, latency, and throughput requirements. In the preceding table, you can see the latency and throughput across different concurrency levels for the given instance type and config name. If config name is lmi-optimized, that means the configuration is pre-optimized by SageMaker. Then you can call .build() to run the optimization job. When the job is complete, you can deploy to an endpoint and test the model predictions. See the following code:

# set deployment config with pre-configured optimization
bulder.set_deployment_config(
    instance_type="ml.g5.12xlarge", 
    config_name="lmi-optimized"
)

# build the deployable model
model = builder.build()

# deploy the model to a SageMaker endpoint
predictor = model.deploy(accept_eula=True)

# use sample input payload to test the deployed endpoint
predictor.predict(sample_input)

Using the inference optimization toolkit to create custom optimizations

In addition to creating a pre-optimized model, you can create custom optimizations based on the instance type you choose. The following table provides a full list of available combinations. In the following sections, we explore compilation on AWS Inferentia first, and then try the other optimization techniques for GPU instances.

Compilation from SageMaker JumpStart

For compilation, you can select the same Meta-Llama-3-8b model from SageMaker JumpStart and choose Optimize on the model page. In the optimization configuration page, you can choose ml.inf2.8xlarge for your instance type. Then provide an output Amazon Simple Storage Service (Amazon S3) location for the optimized artifacts. For large models like Llama 2 70B, for example, the compilation job can take more than an hour. Therefore, we recommend using the inference optimization toolkit to perform ahead-of-time compilation. That way, you only need to compile one time.

Compilation using the SageMaker Python SDK

For the SageMaker Python SDK, you can configure the compilation by changing the environment variables in the .optimize() function. For more details on compilation_config, refer to LMI NeuronX ahead-of-time compilation of models tutorial.

compiled_model = builder.optimize(
    instance_type="ml.inf2.8xlarge",
    accept_eula=True,
    compilation_config={
        "OverrideEnvironment": {
            "OPTION_TENSOR_PARALLEL_DEGREE": "2",
            "OPTION_N_POSITIONS": "2048",
            "OPTION_DTYPE": "fp16",
            "OPTION_ROLLING_BATCH": "auto",
            "OPTION_MAX_ROLLING_BATCH_SIZE": "4",
            "OPTION_NEURON_OPTIMIZE_LEVEL": "2",
        }
   },
   output_path=f"s3://{output_bucket_name}/compiled/"
)

# deploy the compiled model to a SageMaker endpoint
predictor = compiled_model.deploy(accept_eula=True)

# use sample input payload to test the deployed endpoint
predictor.predict(sample_input)

Quantization and speculative decoding from SageMaker JumpStart

For optimizing models on GPU, ml.g5.12xlarge is the default deployment instance type for Llama-3-8b. You can choose quantization, speculative decoding, or both as optimization options. Quantization uses AWQ to reduce the model’s weights to low-bit (INT4) representations. Finally, you can provide an output S3 URL to store the optimized artifacts.

With speculative decoding, you can improve latency and throughput by either using the SageMaker provided draft model or bringing your own draft model from the public Hugging Face model hub or upload from your own S3 bucket.

After the optimization job is complete, you can deploy the model or run further evaluation jobs on the optimized model. On the SageMaker Studio UI, you can choose to use the default sample datasets or provide your own using an S3 URI. At the time of writing, the evaluate performance option is only available through the Amazon SageMaker Studio UI.

Quantization and speculative decoding using the SageMaker Python SDK

The following is the SageMaker Python SDK code snippet for quantization. You just need to provide the quantization_config attribute in the .optimize() function.

optimized_model = builder.optimize(
    instance_type="ml.g5.12xlarge",
    accept_eula=True,
    quantization_config={
        "OverrideEnvironment": {
            "OPTION_QUANTIZE": "awq"
        }
    },
    output_path=f"s3://{output_bucket_name}/quantized/"
)

# deploy the optimized model to a SageMaker endpoint
predictor = optimized_model.deploy(accept_eula=True)

# use sample input payload to test the deployed endpoint
predictor.predict(sample_input)

For speculative decoding, you can change to a speculative_decoding_config attribute and configure SageMaker or a custom model. You may need to adjust the GPU utilization based on the sizes of the draft and target models to fit them both on the instance for inference.

optimized_model = builder.optimize(
    instance_type="ml.g5.12xlarge",
    accept_eula=True,
    speculative_decoding_config={
        "ModelProvider": "sagemaker"
    }
    # speculative_decoding_config={
        # "ModelProvider": "custom",
        # use S3 URI or HuggingFace model ID for custom draft model
        # note: using HuggingFace model ID as draft model requires HF_TOKEN in environment variables
        # "ModelSource": "s3://custom-bucket/draft-model", 
    # }
)

# deploy the optimized model to a SageMaker endpoint
predictor = optimized_model.deploy(accept_eula=True)

# use sample input payload to test the deployed endpoint
predictor.predict(sample_input)

Conclusion

Optimizing generative AI models for inference performance is crucial for delivering cost-effective and responsive generative AI solutions. With the launch of the inference optimization toolkit, you can now optimize your generative AI models, using the latest techniques such as speculative decoding, compilation, and quantization to achieve up to ~2x higher throughput and reduce costs by up to ~50%. This helps you achieve the optimal price-performance balance for your specific use cases with just a few clicks in SageMaker JumpStart or a few lines of code using the SageMaker Python SDK. The inference optimization toolkit significantly simplifies the model optimization process, enabling your business to accelerate generative AI adoption and unlock more opportunities to drive better business outcomes.

To learn more, refer to Optimize model inference with Amazon SageMaker and Achieve up to ~2x higher throughput while reducing costs by up to ~50% for generative AI inference on Amazon SageMaker with the new inference optimization toolkit – Part 1.

About the Authors

James Wu is a Senior AI/ML Specialist Solutions Architect
Saurabh Trikande is a Senior Product Manager
Rishabh Ray Chaudhury is a Senior Product Manager
Kumara Swami Borra is a Front End Engineer
Alwin (Qiyun) Zhao is a Senior Software Development Engineer
Qing Lan is a Senior SDE

Read More

Today, Amazon SageMaker announced a new inference optimization toolkit that helps you reduce the time it takes to optimize generative artificial intelligence (AI) models from months to hours, to achieve best-in-class performance for your use case. With this new capability, you can choose from a menu of optimization techniques, apply them to your generative AI models, validate performance improvements, and deploy the models in just a few clicks.

By employing techniques such as speculative decoding, quantization, and compilation, Amazon SageMaker’s new inference optimization toolkit delivers up to ~2x higher throughput while reducing costs by up to ~50% for generative AI models such as Llama 3, Mistral, and Mixtral models. For example, with a Llama 3-70B model, you can achieve up to ~2400 tokens/sec on a ml.p5.48xlarge instance v/s ~1200 tokens/sec previously without any optimization. Additionally, the inference optimization toolkit significantly reduces the engineering costs of applying the latest optimization techniques, because you don’t need to allocate developer resources and time for research, experimentation, and benchmarking before deployment. You can now focus on your business objectives instead of the heavy lifting involved in optimizing your models.

In this post, we discuss the benefits of this new toolkit and the use cases it unlocks.

Benefits of the inference optimization toolkit

“Large language models (LLMs) require expensive GPU-based instances for hosting, so achieving a substantial cost reduction is immensely valuable. With the new inference optimization toolkit from Amazon SageMaker, based on our experimentation, we expect to reduce deployment costs of our self-hosted LLMs by roughly 30% and to reduce latency by up to 25% for up to 8 concurrent requests” said FNU Imran, Machine Learning Engineer, Qualtrics.

Today, customers try to improve price-performance by optimizing their generative GenAI models with techniques such as speculative decoding, quantization, and compilation. Speculative decoding achieves speedup by predicting and computing multiple potential next tokens in parallel, thereby reducing the overall runtime, without loss in accuracy. Quantization reduces the memory requirements of the model by using a lower-precision data type to represent weights and activations. Compilation optimizes the model to deliver the best-in-class performance on the chosen hardware type, without loss in accuracy.

However, it takes months of developer time to optimally implement generative AI optimization techniques—you need to go through a myriad of open source documentation, iteratively prototype different techniques, and benchmark before finalizing on the deployment configurations. Additionally, there’s a lack of compatibility across techniques and various open source libraries, making it difficult to stack different techniques for best price-performance.

The inference optimization toolkit helps address these challenges by making the process simpler and more efficient. You can select from a menu of latest model optimization techniques and apply them to your models. You can also select a combination of techniques to create an optimization recipe for their models. Then you can run benchmarks using your custom data to evaluate the impact of the techniques on the output quality and the inference performance in just a few clicks. SageMaker will do the heavy lifting of provisioning the required hardware to run the optimization recipe using the most efficient set of deep learning frameworks and libraries, and provide compatibility with target hardware so that the techniques can be efficiently stacked together. You can now deploy popular models like Llama 3 and Mistral available on Amazon SageMaker JumpStart with accelerated inference techniques within minutes, either using the Amazon SageMaker Studio UI or the Amazon SageMaker Python SDK. A number of preset serving configurations are exposed for each model, along with precomputed benchmarks that provide you with different options to pick between lower cost (higher concurrency) or lower per-user latency.

Using the new inference optimization toolkit, we benchmarked the performance and cost impact of different optimization techniques. The toolkit allowed us to evaluate how each technique affected throughput and overall cost-efficiency for our question answering use case.

Speculative decoding

Speculative decoding is an inference technique that aims to speed up the decoding process of large and therefore slow LLMs for latency-critical applications without compromising the quality of the generated text. The key idea is to use a smaller, less powerful, but faster language model called the draft model to generate candidate tokens that are then validated by the larger, more powerful, but slower target model. At each iteration, the draft model generates multiple candidate tokens. The target model verifies the tokens and if it finds a particular token is not acceptable, it rejects it and regenerates that itself. Therefore, the larger model may be doing both verification and some small amount of token generation. The smaller model is significantly faster than the larger model. It can generate all the tokens quickly and then send batches of these tokens to the target models for verification. The target models evaluate them all in parallel, significantly speeding up the final response generation (verification is faster than auto-regressive token generation). For a more detailed understanding, refer to the paper from DeepMind, Accelerating Large Language Model Decoding with Speculative Sampling.

With the new inference optimization toolkit from SageMaker, you get out-of-the-box support for speculative decoding from SageMaker that has been tested for performance at scale for various popular open models. SageMaker offers a pre-built draft model that you can use out of the box, eliminating the need to invest time and resources in building your own draft model from scratch. If you prefer to use your own custom draft model, SageMaker also supports this option, providing flexibility to accommodate your specific custom draft models.

The following graph showcases the throughput (tokens per second) for a Llama3-70B model deployed on a ml.p5.48xlarge using speculative decoding provided through SageMaker compared to a deployment without speculative decoding.

While the results below use a ml.p5.48xlarge, you can also look at exploring deploying Llama3-70 with speculative decoding on a ml.p4d.24xlarge.

The dataset used for these benchmarks is based on a curated version of the OpenOrca question answering use case, where the payloads are between 500–1,000 tokens with mean 622 with respect to the Llama 3 tokenizer. All requests set the maximum new tokens to 250.

Given the increase in throughput that is realized with speculative decoding, we can also see the blended price difference when using speculative decoding vs. when not using speculative decoding.

Here we have calculated the blended price as a 3:1 ratio of input to output tokens.

The blended price is defined as follows:

• Total throughput (tokens per second) = (1/(p50 inter token latency)) x concurrency
• Blended price ($ per 1 million tokens) = (1−(discount rate)) × (instance per hour price) ÷ ((total token throughput per second)×60×60÷10^6)) ÷ 4

Check out the following notebook to learn how to enable speculative decoding using the optimization toolkit for a pre-trained SageMaker JumpStart model.

The following are not supported when using the SageMaker provided draft model for speculative decoding:

• If you have fine-tuned your model outside of SageMaker JumpStart
• The custom inference script is not supported when using SageMaker draft models
• Local testing
• Inference components

Quantization

Quantization is one of the most popular model compression methods to reduce your memory footprint and accelerate inference. By using a lower-precision data type to represent weights and activations, quantizing LLM weights for inference provides four main benefits:

• Reduced hardware requirements for model serving – A quantized model can be served using less expensive and more available GPUs or even made accessible on consumer devices or mobile platforms.
• Increased space for the KV cache – This enables larger batch sizes and sequence lengths.
• Faster decoding latency – Because the decoding process is memory bandwidth bound, less data movement from reduced weight sizes directly improves decoding latency, unless offset by dequantization overhead.
• A higher compute-to-memory access ratio (through reduced data movement) – This is also known as arithmetic intensity. This allows for fuller utilization of available compute resources during decoding.

For quantization, the inference optimization toolkit from SageMaker provides compatibility and supports Activation-aware Weight Quantization (AWQ) for GPUs. AWQ is an efficient and accurate low-bit (INT3/4) post-training weight-only quantization technique for LLMs, supporting instruction-tuned models and multi-modal LLMs. By quantizing the model weights to INT4 using AWQ, you can deploy larger models (like Llama 3 70B) on ml.g5.12xlarge, which is 79.88% cheaper than ml.p4d.24xlarge based on the 1 year SageMaker Savings Plan rate. The memory footprint of INT4 weights is four times smaller than that of native half-precision weights (35 GB vs. 140 GB for Llama 3 70B).

The following graph compares the throughput of an AWQ quantized Llama 3 70B instruct model on ml.g5.12xlarge against a Llama 3 70B instruct model on ml.p4d.24xlarge.There could be implications to the accuracy of the AWQ quantized model due to compression. Having said that, the price-performance is better on ml.g5.12xlarge and the throughput per instance is lower. You can add additional instances to your SageMaker endpoint according to your use case. We can see the cost savings realized in the following blended price graph.

In the following graph, we have calculated the blended price as a 3:1 ratio of input to output tokens. In addition, we applied the 1 year SageMaker Savings Plan rate for the instances.

Refer to the following notebook to learn more about how to enable AWQ quantization and speculative decoding using the optimization toolkit for a pre-trained SageMaker JumpStart model. If you want to deploy a fine-tuned model with SageMaker JumpStart using speculative decoding, refer to the following notebook.

Compilation

Compilation optimizes the model to extract the best available performance on the chosen hardware type, without any loss in accuracy. For compilation, the SageMaker inference optimization toolkit provides efficient loading and caching of optimized models to reduce model loading and auto scaling time by up to 40–60 % for Llama 3 8B and 70B.

Model compilation enables running LLMs on accelerated hardware, such as GPUs or custom silicon like AWS Trainium and AWS Inferentia, while simultaneously optimizing the model’s computational graph for optimal performance on the target hardware. On Trainium and AWS Inferentia, the Neuron Compiler ingests deep learning models in various formats such as PyTorch and safetensors, and then optimizes them to run efficiently on Neuron devices. When using the Large Model Inference (LMI) Deep Learning Container (DLC), the Neuron Compiler is invoked from within the framework and creates compiled artifacts. These compiled artifacts are unique for a combination of input shapes, precision of the model, tensor parallel degree, and other framework- or compiler-level configurations. Although the compilation process avoids overhead during inference and enables optimized inference, it can take a lot of time.

To avoid re-compiling every time a model is deployed onto a Neuron device, SageMaker introduces the following features:

• A cache of pre-compiled artifacts – This includes popular models like Llama 3, Mistral, and more for Trainium and AWS Inferentia 2. When using an optimized model with the compilation config, SageMaker automatically uses these cached artifacts when the configurations match.
• Ahead-of-time compilation – The inference optimization toolkit enables you to compile your models with the desired configurations before deploying them on SageMaker.

The following graph illustrates the improvement in model loading time when using pre-compiled artifacts with the SageMaker LMI DLC. The models were compiled with a sequence length of 8,192 and a batch size of 8, with Llama 3 8B deployed on an inf2.48xlarge (tensor parallel degree = 24) and Llama 3 70B on a trn1.32xlarge (tensor parallel degree = 32).

Refer to the following notebook for more information on how to enable Neuron compilation using the optimization toolkit for a pre-trained SageMaker JumpStart model.

Pricing

For compilation and quantization jobs, SageMaker will optimally choose the right instance type, so you don’t have to spend time and effort. You will be charged based on the optimization instance used. To learn model, see Amazon SageMaker pricing. For speculative decoding, there is no optimization involved; QuickSilver will package the right container and parameters for the deployment. Therefore, there are no additional costs to you.

Conclusion

Refer to Achieve up to 2x higher throughput while reducing cost by up to 50% for GenAI inference on SageMaker with new inference optimization toolkit: user guide – Part 2 blog to learn to get started with the inference optimization toolkit when using SageMaker inference with SageMaker JumpStart and the SageMaker Python SDK. You can use the inference optimization toolkit on any supported models on SageMaker JumpStart. For the full list of supported models, refer to Optimize model inference with Amazon SageMaker.

About the authors

Raghu Ramesha is a Senior GenAI/ML Solutions Architect
Marc Karp is a Senior ML Solutions Architect
Ram Vegiraju is a Solutions Architect
Pierre Lienhart is a Deep Learning Architect
Pinak Panigrahi is a Senior Solutions Architect Annapurna ML
Rishabh Ray Chaudhury is a Senior Product Manager

Read More

Anthropic Claude 3.5 Sonnet currently ranks at the top of S&P AI Benchmarks by Kensho, which assesses large language models (LLMs) for finance and business. Kensho is the AI Innovation Hub for S&P Global. Using Amazon Bedrock, Kensho was able to quickly run Anthropic Claude 3.5 Sonnet through a challenging suite of business and financial tasks. We discuss these tasks and the capabilities of Anthropic Claude 3.5 Sonnet in this post.

Limitations of LLM evaluations

It is a common practice to use standardized tests, such as Massive Multitask Language Understanding (MMLU, a test consisting of multiple-choice questions that cover 57 disciplines like math, philosophy, and medicine) and HumanEval (testing code generation), to evaluate LLMs. Although these evaluations are useful in giving LLM users a sense of an LLM’s relative performance, they have limitations. For example, there could be leakage of benchmark datasets’ questions and answers into training data. Additionally, today’s LLMs work well for general tasks, such as question answering tasks and code generation. However, these capabilities don’t always translate to domain-specific tasks. In the financial services industry, we hear customers ask which model to choose for their financial domain generative artificial intelligence (AI) applications. These applications require the LLMs to have requisite domain knowledge and be able to reason about numeric data to calculate metrics and extract insights. We have also heard from customers that highly ranked general benchmark LLMs don’t necessarily provide them with the best performance for their given finance and business applications.

Our customers often ask us if we have a benchmark of LLMs just for the financial industry that could help them pick the right LLMs faster.

S&P AI Benchmarks by Kensho

When Kensho’s R&D lab began to research and develop useful, challenging datasets for finance and business, it quickly became clear that within the finance industry, there was a scarcity of such realistic evaluations. To address this challenge, the lab created S&P AI Benchmarks, which aims to serve as the industry standard for benchmarking models for finance and business.

“By offering a robust and independent benchmarking solution, we want to help the financial services industry make smart decisions about which models to implement for which use cases.”

– Bhavesh Dayalji, Chief AI Officer of S&P Global and CEO of Kensho.

S&P AI Benchmarks focuses on measuring models’ ability to perform tasks that center around three categories of capabilities and knowledge: domain knowledge, quantity extraction, and quantitative reasoning (more details can be found in this paper). This publicly available resource includes a corresponding leaderboard, which allows everyone to see the performance of every state-of-the-art language model that has been evaluated on these rigorous tasks. Anthropic Claude 3.5 Sonnet is currently ranked number one (as of July 2024), demonstrating Anthropic’s strengths in the business and finance domain.

Kensho chose to test their benchmark with Amazon Bedrock because of its ease of use and enterprise-ready security and privacy controls.

The evaluation tasks

S&P AI Benchmarks evaluates LLMs using a wide range of questions concerning finance and business. The evaluation comprises 600 questions spanning three categories: domain knowledge, quantity extraction, and quantitative reasoning. Each question has been verified by domain experts and finance professionals with over 5 years of experience.

Quantitative reasoning

This task determines if, given a question and lengthy documents, the model can perform complex calculations and correctly reason to produce an accurate answer. The questions are written by financial professionals using real-world data and financial knowledge. As such, they are closer to the kinds of questions that business and financial professionals would ask in a generative AI application. The following is an example:

Question: The market price of K-T-Lew Corporation’s common stock is $60 per share, and each share gives its owner one subscription right. Four rights are required to purchase an additional share of common stock at the subscription price of $54 per share. If the common stock is currently selling rights-on, what is the theoretical value of a right? Answer to the nearest cent.

To answer the question, LLMs must resolve complex quantity references and use implicit financial background knowledge. For example, “subscription right,” “selling rights-on,” and “subscription price” in the preceding question require financial background knowledge to understand the terms. To generate the answer, LLMs need to have the financial knowledge of calculating the “theoretical value of a right.”

Quantity extraction

Given financial reports, an LLM can extract the pertinent numerical information. Many business and finance workflows require high-precision quantity extraction. In the following example, for an LLM to answer the question correctly, it needs to understand the table row represents location and the column represents year, and then extract the correct quantity (total amount) from the table based on the asked location and year:

Question: What was the Total Americas amount in 2019? (thousand)

Given Context: The Company’s top ten clients accounted for 42.2%, 44.2% and 46.9% of 
its consolidated revenues during the years ended December 31, 2019, 2018 and 2017, respectively.
The following table represents a disaggregation of revenue from contracts with customers by 
delivery location (in thousands):

Domain knowledge

Models must demonstrate an understanding of business and financial terms, practices, and formulae. The task is to answer multiple-choice questions collected from CFA practice exams and the business ethics, microeconomics, and professional accounting exams from the MMLU dataset. In the following example question, the LLM needs to understand what a fixed-rate system is:

Question: A fixed-rate system is characterized by:
A: Explicit legislative commitment to maintain a specified parity.
B: Monetary independence being subject to the maintenance of an exchange rate peg.
C: Target foreign exchange reserves bearing a direct relationship to domestic monetary aggregates.

Anthropic Claude 3.5 Sonnet on Amazon Bedrock

In addition to ranking at the top on S&P AI Benchmarks, Anthropic Claude 3.5 Sonnet yields state-of-the-art performance on a wide range of other tasks, including undergraduate-level expert knowledge (MMLU), graduate-level expert reasoning (GPQA), code (HumanEval), and more. As pointed out in Anthropic’s Claude 3.5 Sonnet model now available in Amazon Bedrock: Even more intelligence than Claude 3 Opus at one-fifth the cost, Anthropic Claude 3.5 Sonnet made key improvements in visual processing and understanding, writing and content generation, natural language processing, coding, and generating insights.

Get started with Anthropic Claude 3.5 Sonnet on Amazon Bedrock

Anthropic Claude 3.5 Sonnet is generally available in Amazon Bedrock as part of the Anthropic Claude family of AI models. Amazon Bedrock is a fully managed service that offers quick access to a choice of industry-leading LLMs and other foundation models from AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon. It also offers a broad set of capabilities to build generative AI applications, simplifying development while supporting privacy and security. Tens of thousands of customers have already selected Amazon Bedrock as the foundation for their generative AI strategy. Customers from the financial industry such as Nasdaq, NYSE, Broadridge, Jefferies, NatWest, and more use Amazon Bedrock to build their generative AI applications.

“The Kensho team uses Amazon Bedrock to quickly evaluate models from several different providers. In fact, access to Amazon Bedrock allowed the team to benchmark Anthropic Claude 3.5 Sonnet within 24 hours.”

– Diana Mingels, Head of Machine Learning at Kensho.

Conclusion

In this post, we walked through the S&P AI Benchmarks task details for business and finance. The benchmark shows that Anthropic Claude 3.5 Sonnet is the leading performer in these tasks. To start using this new model, see Anthropic Claude models. With Amazon Bedrock, you get a fully managed service offering access to leading AI models from companies like AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon through a single API, along with a broad set of capabilities to build generative AI applications. Learn more and get started today at Amazon Bedrock.

About the authors

Qingwei Li is a Machine Learning Specialist at Amazon Web Services. He received his Ph.D. in Operations Research after he broke his advisor’s research grant account and failed to deliver the Nobel Prize he promised. Currently he helps customers in the financial service and insurance industry build machine learning solutions on AWS. In his spare time, he likes reading and teaching.

Joe Dunn is an AWS Principal Solutions Architect in Financial Services with over 20 years of experience in infrastructure architecture and migration of business-critical loads to AWS. He helps financial services customers to innovate on the AWS Cloud by providing solutions using AWS products and services.

Raghvender Arni (Arni) is a part of the AWS Generative AI GTM team and leads the Cross-Portfolio team which is a multidisciplinary group of AI specialists dedicated to accelerating and optimizing generative AI adoption across industries.

Simon Zamarin is an AI/ML Solutions Architect whose main focus is helping customers extract value from their data assets. In his spare time, Simon enjoys spending time with family, reading sci-fi, and working on various DIY house projects.

Scott Mullins is Managing Director and General Manger of AWS’ Worldwide Financial Services organization. In this role, Scott is responsible for AWS’ relationships with systemically important financial institutions, and for leading the development and execution of AWS’ strategic initiatives across Banking, Payments, Capital Markets, and Insurance around the world. Prior to joining AWS in 2014, Scott’s 28-year career in financial services included roles at JPMorgan Chase, Nasdaq, Merrill Lynch, and Penson Worldwide. At Nasdaq, Scott was the Product Manager responsible for building the exchange’s first cloud-based solution, FinQloud. Before joining NASDAQ, Scott ran Surveillance and Trading Compliance for one of the nation’s largest clearing broker-dealers, with responsibility for regulatory response, emerging regulatory initiatives, and compliance matters related to the firm’s trading and execution services divisions. Prior to his roles in regulatory compliance, Scott spent 10 years as an equity trader. A graduate of Texas A&M University, Scott is a subject matter expert quoted in industry media, and a recognized speaker at industry events..

Read More

This blog post is co-written with Qaish Kanchwala  from The Weather Company.

As industries begin adopting processes dependent on machine learning (ML) technologies, it is critical to establish machine learning operations (MLOps) that scale to support growth and utilization of this technology. MLOps practitioners have many options to establish an MLOps platform; one among them is cloud-based integrated platforms that scale with data science teams. AWS provides a full-stack of services to establish an MLOps platform in the cloud that is customizable to your needs while reaping all the benefits of doing ML in the cloud.

In this post, we share the story of how The Weather Company (TWCo) enhanced its MLOps platform using services such as Amazon SageMaker, AWS CloudFormation, and Amazon CloudWatch. TWCo data scientists and ML engineers took advantage of automation, detailed experiment tracking, integrated training, and deployment pipelines to help scale MLOps effectively. TWCo reduced infrastructure management time by 90% while also reducing model deployment time by 20%.

The need for MLOps at TWCo

TWCo strives to help consumers and businesses make informed, more confident decisions based on weather. Although the organization has used ML in its weather forecasting process for decades to help translate billions of weather data points into actionable forecasts and insights, it continuously strives to innovate and incorporate leading-edge technology in other ways as well. TWCo’s data science team was looking to create predictive, privacy-friendly ML models that show how weather conditions affect certain health symptoms and create user segments for improved user experience.

TWCo was looking to scale its ML operations with more transparency and less complexity to allow for more manageable ML workflows as their data science team grew. There were noticeable challenges when running ML workflows in the cloud. TWCo’s existing Cloud environment lacked transparency for ML jobs, monitoring, and a feature store, which made it hard for users to collaborate. Managers lacked the visibility needed for ongoing monitoring of ML workflows. To address these pain points, TWCo worked with the AWS Machine Learning Solutions Lab (MLSL) to migrate these ML workflows to Amazon SageMaker and the AWS Cloud. The MLSL team collaborated with TWCo to design an MLOps platform to meet the needs of its data science team, factoring present and future growth.

Examples of business objectives set by TWCo for this collaboration are:

• Achieve quicker reaction to the market and faster ML development cycles
• Accelerate TWCo migration of their ML workloads to SageMaker
• Improve end user experience through adoption of manage services
• Reduce time spent by engineers in maintenance and upkeep of the underlying ML infrastructure

Functional objectives were set to measure the impact of MLOps platform users, including:

• Improve the data science team’s efficiency in model training tasks
• Decrease the number of steps required to deploy new models
• Reduce the end-to-end model pipeline runtime

Solution overview

The solution uses the following AWS services:

• AWS CloudFormation – Infrastructure as code (IaC) service to provision most templates and assets.
• AWS CloudTrail – Monitors and records account activity across AWS infrastructure.
• Amazon CloudWatch – Collects and visualizes real-time logs that provide the basis for automation.
• AWS CodeBuild – Fully managed continuous integration service to compile source code, runs tests, and produces ready-to-deploy software. Used to deploy training and inference code.
• AWS CodeCommit – Managed sourced control repository that stores MLOps infrastructure code and IaC code.
• AWS CodePipeline – Fully managed continuous delivery service that helps automate the release of pipelines.
• Amazon SageMaker – Fully managed ML platform to perform ML workflows from exploring data, training, and deploying models.
• AWS Service Catalog – Centrally manages cloud resources such as IaC templates used for MLOps projects.
• Amazon Simple Storage Service (Amazon S3) – Cloud object storage to store data for training and testing.

The following diagram illustrates the solution architecture.

This architecture consists of two primary pipelines:

• Training pipeline – The training pipeline is designed to work with features and labels stored as a CSV-formatted file on Amazon S3. It involves several components, including Preprocess, Train, and Evaluate. After training the model, its associated artifacts are registered with the Amazon SageMaker Model Registry through the Register Model component. The Data Quality Check part of the pipeline creates baseline statistics for the monitoring task in the inference pipeline.
• Inference pipeline – The inference pipeline handles on-demand batch inference and monitoring tasks. Within this pipeline, SageMaker on-demand Data Quality Monitor steps are incorporated to detect any drift when compared to the input data. The monitoring results are stored in Amazon S3 and published as a CloudWatch metric, and can be used to set up an alarm. The alarm is used later to invoke training, send automatic emails, or any other desired action.

The proposed MLOps architecture includes flexibility to support different use cases, as well as collaboration between various team personas like data scientists and ML engineers. The architecture reduces the friction between cross-functional teams moving models to production.

ML model experimentation is one of the sub-components of the MLOps architecture. It improves data scientists’ productivity and model development processes. Examples of model experimentation on MLOps-related SageMaker services require features like Amazon SageMaker Pipelines, Amazon SageMaker Feature Store, and SageMaker Model Registry using the SageMaker SDK and AWS Boto3 libraries.

When setting up pipelines, resources are created that are required throughout the lifecycle of the pipeline. Additionally, each pipeline may generate its own resources.

The pipeline setup resources are:

• Training pipeline:
  • SageMaker pipeline
  • SageMaker Model Registry model group
  • CloudWatch namespace
• Inference pipeline:
  • SageMaker pipeline

The pipeline run resources are:

• Training pipeline:
  • SageMaker model

You should delete these resources when the pipelines expire or are no longer needed.

SageMaker project template

In this section, we discuss the manual provisioning of pipelines through an example notebook and automatic provisioning of SageMaker pipelines through the use of a Service Catalog product and SageMaker project.

By using Amazon SageMaker Projects and its powerful template-based approach, organizations establish a standardized and scalable infrastructure for ML development, allowing teams to focus on building and iterating ML models, reducing time wasted on complex setup and management.

The following diagram shows the required components of a SageMaker project template. Use Service Catalog to register a SageMaker project CloudFormation template in your organization’s Service Catalog portfolio.

To start the ML workflow, the project template serves as the foundation by defining a continuous integration and delivery (CI/CD) pipeline. It begins by retrieving the ML seed code from a CodeCommit repository. Then the BuildProject component takes over and orchestrates the provisioning of SageMaker training and inference pipelines. This automation delivers a seamless and efficient run of the ML pipeline, reducing manual intervention and speeding up the deployment process.

Dependencies

The solution has the following dependencies:

• Amazon SageMaker SDK – The Amazon SageMaker Python SDK is an open source library for training and deploying ML models on SageMaker. For this proof of concept, pipelines were set up using this SDK.
• Boto3 SDK – The AWS SDK for Python (Boto3) provides a Python API for AWS infrastructure services. We use the SDK for Python to create roles and provision SageMaker SDK resources.
• SageMaker Projects – SageMaker Projects delivers standardized infrastructure and templates for MLOps for rapid iteration over multiple ML use cases.
• Service Catalog – Service Catalog simplifies and speeds up the process of provisioning resources at scale. It offers a self-service portal, standardized service catalog, versioning and lifecycle management, and access control.

Conclusion

In this post, we showed how TWCo uses SageMaker, CloudWatch, CodePipeline, and CodeBuild for their MLOps platform. With these services, TWCo extended the capabilities of its data science team while also improving how data scientists manage ML workflows. These ML models ultimately helped TWCo create predictive, privacy-friendly experiences that improved user experience and explains how weather conditions impact consumers’ daily planning or business operations. We also reviewed the architecture design that helps maintain responsibilities between different users modularized. Typically data scientists are only concerned with the science aspect of ML workflows, whereas DevOps and ML engineers focus on the production environments. TWCo reduced infrastructure management time by 90% while also reducing model deployment time by 20%.

This is just one of many ways AWS enables builders to deliver great solutions. We encourage to you to get started with Amazon SageMaker today.

About the Authors

Qaish Kanchwala is a ML Engineering Manager and ML Architect at The Weather Company. He has worked on every step of the machine learning lifecycle and designs systems to enable AI use cases. In his spare time, Qaish likes to cook new food and watch movies.

Chezsal Kamaray is a Senior Solutions Architect within the High-Tech Vertical at Amazon Web Services. She works with enterprise customers, helping to accelerate and optimize their workload migration to the AWS Cloud. She is passionate about management and governance in the cloud and helping customers set up a landing zone that is aimed at long-term success. In her spare time, she does woodworking and tries out new recipes while listening to music.

Anila Joshi has more than a decade of experience building AI solutions. As an Applied Science Manager at the AWS Generative AI Innovation Center, Anila pioneers innovative applications of AI that push the boundaries of possibility and guides customers to strategically chart a course into the future of AI.

Kamran Razi is a Machine Learning Engineer at the Amazon Generative AI Innovation Center. With a passion for creating use case-driven solutions, Kamran helps customers harness the full potential of AWS AI/ML services to address real-world business challenges. With a decade of experience as a software developer, he has honed his expertise in diverse areas like embedded systems, cybersecurity solutions, and industrial control systems. Kamran holds a PhD in Electrical Engineering from Queen’s University.

Shuja Sohrawardy is a Senior Manager at AWS’s Generative AI Innovation Center. For over 20 years, Shuja has utilized his technology and financial services acumen to transform financial services enterprises to meet the challenges of a highly competitive and regulated industry. Over the past 4 years at AWS, Shuja has used his deep knowledge in machine learning, resiliency, and cloud adoption strategies, which has resulted in numerous customer success journeys. Shuja holds a BS in Computer Science and Economics from New York University and an MS in Executive Technology Management from Columbia University.

Francisco Calderon is a Data Scientist at the Generative AI Innovation Center (GAIIC). As a member of the GAIIC, he helps discover the art of the possible with AWS customers using generative AI technologies. In his spare time, Francisco likes playing music and guitar, playing soccer with his daughters, and enjoying time with his family.

Read More

Eviden is a next-gen technology leader in data-driven, trusted, and sustainable digital transformation. With a strong portfolio of patented technologies and worldwide leading positions in advanced computing, security, AI, cloud, and digital platforms, Eviden provides deep expertise for a multitude of industries in more than 47 countries. Eviden is an AWS Premier partner, bringing together 47,000 world-class talents and expanding the possibilities of data and technology across the digital continuum, now and for generations to come. Eviden is an Atos Group company with an annual revenue of over €5 billion.

We are passionate about our people improving their skills, and support the development of the next generation of cloud-centered talent. Although fundamental knowledge gained through training and certification is important, there’s no substitute for getting hands-on. We complement individual learning with hands-on opportunities, including Immersion Days, Gamedays, and using AWS DeepRacer.

AWS DeepRacer empowers users to train reinforcement learning models on the AWS Cloud and race them around a virtual track. Unlike traditional programming, where you define the desired output, AWS DeepRacer allows you to define rewards for specific behaviors, such as going faster or staying centered on the track. This hands-on experience gives learners an excellent opportunity to engage with the AWS Management Console and develop Python-based reward functions, fostering valuable skills in cloud-centered technologies and machine learning (ML).

To elevate the event experience and streamline the management of their global AWS DeepRacer series, Eviden adopted the open source AWS DeepRacer Event Manager (DREM) solution. In this post, we discuss the benefits of DREM and the experience for racers, event staff, and spectators.

Introducing AWS DeepRacer Event Manager

AWS DeepRacer Event Manager is an innovative application that Eviden has deployed within its own AWS environment. Comprised of AWS Cloud-centered services, DREM is designed to simplify the process of hosting in-person AWS DeepRacer events, while also delivering a more engaging and immersive experience for both participants and spectators.

With their prior experience hosting AWS DeepRacer events in the UK, Eviden sought to expand the reach of this exciting initiative globally. By adopting the DREM solution, Eviden’s experienced event staff in the UK were able to seamlessly support their counterparts in hosting AWS DeepRacer events for the first time in locations such as Bydgoszcz, Paris, and Pune. The DREM solution empowers Eviden to seamlessly configure and manage their global AWS DeepRacer events. Within the platform, each event location’s AWS DeepRacer cars are registered, enabling remote configuration and model uploads powered by AWS Systems Manager. Additionally, a Raspberry Pi device is registered at each location to serve as an integrated timing solution, which uses DREM’s data-driven racing capabilities to capture and report critical performance metrics for each racer, such as best lap time, average lap time, and total laps completed.

Furthermore, the deep integration between DREM’s timing solution and the event’s streaming overlay has enabled Eviden to deliver a significantly more engaging experience for both in-person attendees and remote viewers, so everyone can stay fully informed and immersed in the action throughout the event.

The following diagram illustrates the DREM architecture and components.

Racer experience

For racers, the DREM experience begins with registration, during which they are encouraged to upload their AWS DeepRacer models in advance of the event. Authentication is seamlessly handled by Amazon Cognito, with out-of-the-box support for a local Amazon Cognito identity store, as well as the flexibility to integrate with a corporate identity provider solution if required. After the registered racers have uploaded their models, DREM automatically scans for any suspicious content, quarantining it as necessary, before making the verified models available for the racing competition.

Event staff experience

For the event staff, the DREM solution greatly simplifies the process of running an AWS DeepRacer competition. User management, model uploading, and naming conventions are all handled seamlessly within the platform, eliminating any potential confusion around model ownership. To further enforce the integrity of the competition, DREM applies MD5 hashing to the uploaded models, preventing any unauthorized model sharing between racers. Additionally, the DREM interface makes it remarkably straightforward and efficient to upload multiple models to the AWS DeepRacer cars, providing a far superior experience compared to the cars’ native graphical UI. DREM also simplifies the management of the AWS DeepRacer car fleets and Raspberry Pi timing devices, empowering the event staff to remotely remove models, restart the AWS DeepRacer service, and even print user-friendly labels for the cars, making it effortless for racers to connect to them using the provided tablets.

To further streamline the event management process, DREM provides pre-built scripts that enable seamless registration of devices, which can then be fully managed remotely. The timekeeping functionality is automatically handled within DREM, using Raspberry Pi devices, pressure sensors, and pressure sensor trimming—either through custom DIY modifications or the use of the excellent Digital Racing Kings boards. The DREM timekeeping system represents a significant improvement over the alternative solutions Eviden has used in the past. It captures critical race metrics, including remaining time, all lap times, and the fastest lap. Additionally, the system provides an option to invalidate a lap, such as in cases where a car went off-track. After a racer has completed their runs, the data is securely stored in DREM, and the leaderboard is automatically updated to reflect the latest results.

Spectator experience

As a global systems integrator, Eviden was determined to deliver a truly spectacular experience, not only for the onsite participants in their AWS DeepRacer finals, but also for those in the room observing the event, as well as remote viewers racing using a proxy or simply interested in watching the competition unfold. To achieve this, Eviden took advantage of the DREM solution’s seamlessly integrated streaming overlay and leaderboard capabilities, which kept all attendees, both in-person and online, fully engaged and informed of the current standings throughout the event.

In previous AWS DeepRacer events, participants had to rely on someone in the room to verbally communicate lap times and remaining race time. However, with DREM, both the racers and spectators have instant access to all the critical timing information, keeping everyone fully up to date. This was especially beneficial for remote participants, who could now clearly see which cars were on the track and follow the progress of their fellow racers, with the streaming overlay dynamically updating to display the top positions on the leaderboard.

In addition, DREM offers a dedicated webpage displaying the complete leaderboard, which can be conveniently shown on screens in the event space, as well as allow remote attendees to follow the competition progress from other locations throughout the day.

The significant improvements to the event experience were clearly reflected in the feedback received from this year’s participants:

• “Joining remotely was superb.”
• “Remote interaction over teams was much improved this year.”
• “The physical event was excellent, well attended, and ran very smoothly.”
• “Loved it, even though I was attending remote.”
• “The event itself was fantastic, with each stage being well-planned and organized. The on-site atmosphere exceeded my expectations, making it an incredible event to be part of.”

Well-architected

The DREM solution has been meticulously designed with a well-architected approach. From the event organizer’s perspective, the peace of mind of knowing that DREM is secured using AWS WAF, Amazon CloudFront, and AWS Shield Standard, with user management seamlessly handled by Amazon Cognito, is invaluable. Additionally, the platform’s role-based access control (RBAC) is managed through AWS Identity and Access Management (IAM), applying least-privilege policies for enhanced security.

The DREM solution is built entirely using AWS Cloud-centered technologies, delivering inherent performance efficiency and reliability. When Eviden is not actively hosting events, there is minimal ongoing activity in the DREM environment, consisting primarily of standing items such as AWS WAF rules, CloudFront distributions, Amazon Simple Storage Service (Amazon S3) buckets, Amazon DynamoDB tables, and Systems Manager fleet configurations. However, during event periods, DREM seamlessly scales to use AWS Lambda, Amazon EventBridge, AWS Step Functions, and other serverless services, dynamically meeting the demands of the event hosting requirements.

Owing to the well-architected nature of the DREM solution, the platform is remarkably cost-effective. When not actively hosting events, the DREM environment incurs a minimal cost of around $6–8 per month, the majority of which is attributed to the AWS WAF protection. During event periods, the costs increase based on the number of users and models uploaded, but typically only rise to around $15 per month. To further optimize ongoing costs, DREM incorporates measures such as an Amazon S3 lifecycle policy that automatically removes uploaded models after a period of 2 weeks.

Conclusion

Are you interested in elevating your own AWS DeepRacer events and delivering a more engaging experience for your participants? We encourage you to explore the AWS DeepRacer Event Manager solution and see how it can transform your event management process.

To get started, visit the GitHub repo to learn more about the solution’s features and architecture. You can also reach out to the Eviden team or your local AWS Solutions Architect to discuss how DREM can be tailored to your specific event requirements.

Don’t miss out on the opportunity to take your AWS DeepRacer initiatives to the next level. Explore DREM and join us at an upcoming AWS DeepRacer event today!

About the authors

Sathya Paduchuri is a Senior Partner Solution Architect(PSA) at Amazon Web Services. Sathya helps partners run optimised workloads on AWS, build and develop their cloud practice(s) and develop new offerings.

Mark Ross is a Chief Architect at Eviden and has specialised in AWS for the past 8 years, gaining and maintaining all AWS certifications since 2021. Mark is passionate about helping customers build, migrate to and exploit AWS.  Mark has created and grown a large AWS community within Eviden.

Read More

Stable Diffusion XL by Stability AI is a high-quality text-to-image deep learning model that allows you to generate professional-looking images in various styles. Managed versions of Stable Diffusion XL are already available to you on Amazon SageMaker JumpStart (see Use Stable Diffusion XL with Amazon SageMaker JumpStart in Amazon SageMaker Studio) and Amazon Bedrock (see Stable Diffusion XL in Amazon Bedrock), allowing you to produce creative content in minutes. The base version of Stable Diffusion XL 1.0 assists with the creative process using generic subjects in the image, which enables use cases such as game character design, creative concept generation, film storyboarding, and image upscaling. However, for use cases that require generating images with a unique subject, you can fine-tune Stable Diffusion XL with a custom dataset by using a custom training container with Amazon SageMaker. With this personalized image generation model, you can incorporate your custom subject into the powerful image generation process that is provided by the Stable Diffusion XL base model.

In this post, we provide step-by-step instructions to create a custom, fine-tuned Stable Diffusion XL model using SageMaker to generate unique images. This automated solution helps you get started quickly by providing all the code and configuration necessary to generate your unique images—all you need is images of your subject. This is useful for use cases across various domains such as media and entertainment, games, and retail. Examples include using your custom subject for marketing material for film, character creation for games, and brand-specific images for retail. To explore more AI use cases, visit the AI Use Case Explorer.

Solution overview

The solution is composed of three logical parts:

• The first part creates a Docker container image with the necessary framework and configuration for the training container.
• The second part uses the training container to perform model training on your dataset, and outputs a fine-tuned custom Low-Rank Adaptation (LoRA) model. LoRA is an efficient fine-tuning method that doesn’t require adjusting the base model parameters. Instead, it adds a smaller number of parameters that are applied to the base model temporarily.
• The third part takes the fine-tuned custom model and allows you to generate creative and unique images.

The following diagram illustrates the solution architecture.

The workflow to create the training container consists of the following services:

• SageMaker uses Docker containers throughout the ML lifecycle. SageMaker is flexible and allows you to bring your own container to use for model development, training, and inference. For this post, we build a custom container with the appropriate dependencies that will perform the fine-tuning.
• Kohya SS is a framework that allows you to train Stable Diffusion models. Kohya SS works with different host environments. This solution uses the Docker on Linux environment option. Kohya SS can be used with a GUI. However, this solution uses the equivalent GUI parameters as a pre-configured TOML file to automate the entire Stable Diffusion XL fine-tuning process.
• AWS CodeCommit is a fully managed source control service that hosts private Git repositories. We use CodeCommit to store the code that is necessary to build the training container (Dockerfile, buildspec.yml), and the training script (train.py) that is invoked when model training is initiated.
• Amazon EventBridge is a serverless event bus, used to receive, filter, and route events. EventBridge captures any changes to the CodeCommit repository files, and invokes a new Docker container image to be built.
• Amazon Elastic Container Registry (Amazon ECR) is a fully managed container hosting registry. We use it to store the custom training container image.
• AWS CodeBuild is a fully managed continuous integration service that compiles source code, runs tests, and produces deployable software packages. We use it to build the custom training container image. CodeBuild then pushes this image to Amazon ECR.

Various methods exist to fine-tune your model. Compared to methods that require training a new full model, the LoRA fine-tuning method doesn’t modify the original model. Instead, think of it as a layer on top of the base model. Not having to train and produce a full model for each subject has its advantages. This lowers the compute requirements for training, reduces the storage size of the models, and decreases the training time required, making the process more cost-effective at scale. In this post, we demonstrate how to create a LoRA model, based on the Stable Diffusion XL 1.0 base model, using your own subject.

The training workflow uses the following services and features:

• Amazon Simple Storage Service (Amazon S3) is a highly durable and scalable object store. Your custom dataset and configuration file will be uploaded to Amazon S3, and then retrieved by the custom Docker container to train on those images.
• Amazon SageMaker Model Training is a feature of SageMaker that allows you to standardize and manage your training jobs at scale, without the need to manage infrastructure. When the container starts up as part of a training job, the train.py file is invoked. When the training process is complete, the output model that resides in the /opt/ml/model directory is automatically uploaded to the S3 bucket specified in the training job configuration.
• Amazon SageMaker Pipelines is a workflow orchestration service that allows you to automate ML processes, from data preprocessing to model monitoring. This allows you to initiate a training pipeline, taking in as input the Amazon S3 location of your dataset and configuration file, ECR container image, and infrastructure specifications for the training job.

Now you’re ready to prompt your fine-tuned model to generate unique images. SageMaker gives you the flexibility to bring your own container for inference. You can use SageMaker hosting services with your own custom inference container to configure an inference endpoint. However, to demonstrate the Automatic1111 Stable Diffusion UI, we show you how to run inference on an Amazon Elastic Compute Cloud (Amazon EC2) instance (or locally on your own machine).

This solution fully automates the creation of a fine-tuned LoRA model with Stable Diffusion XL 1.0 as the base model. In the following sections, we discuss how to satisfy the prerequisites, download the code, and use the Jupyter notebook in the GitHub repository to deploy the automated solution using an Amazon SageMaker Studio environment.

The code for this end-to-end solution is available in the GitHub repository.

Prerequisites

This solution has been tested in the AWS Region us-west-2, but applies to any Region where these services are available. Make sure you have the following prerequisites:

• AWS account
• SageMaker domain
• SageMaker domain user profile

Download the necessary code in SageMaker Studio

In this section, we walk through the steps to download the necessary code in SageMaker Studio and set up your notebook.

Navigate to the terminal in SageMaker Studio JupyterLab

Complete the following steps to open the terminal:

1. Log in to your AWS account and open the SageMaker Studio console.
2. Select your user profile and choose Open Studio to open SageMaker Studio.
3. Choose JupyterLab to open the JupyterLab application. This environment is where you will run the commands.
4. If you already have a space created, choose Run to open the space.
5. If you don’t have a space, choose Create JupyterLab space. Enter a name for the space and choose Create space. Leave the default values and choose Run space.
6. When the environment shows a status of Running, choose Open JupyterLab to open the new space.
7. In the JupyterLab Launcher window, choose Terminal.

Download the code to your SageMaker Studio environment

Run the following commands from the terminal. For this post, you check out just the required directories of the GitHub repo (so you don’t have to download the entire repository).

git clone --no-checkout https://github.com/aws/amazon-sagemaker-examples.git
cd amazon-sagemaker-examples/
git sparse-checkout set use-cases/text-to-image-fine-tuning
git checkout

If successful, you should see the output Your branch is up to date with 'origin/main'.

Open the notebook in SageMaker Studio JupyterLab

Complete the following steps to open the notebook:

1. In JupyterLab, choose File Browser in the navigation pane.
2. Navigate to the project directory named amazon-sagemaker-examples/use-cases/text-to-image-fine-tuning.
3. Open the Jupyter notebook named kohya-ss-fine-tuning.ipynb.
4. Choose your runtime kernel (it’s set to use Python 3 by default).
5. Choose Select.

You now have a kernel that is ready to run commands. In the following steps, we use this notebook to create the necessary resources.

Train a custom Stable Diffusion XL model

In this section, we walk through the steps to train a custom Stable Diffusion XL model.

Set up AWS infrastructure with AWS CloudFormation

For your convenience, an AWS CloudFormation template has been provided to create the necessary AWS resources. Before you create the resources, configure AWS Identity and Access Management (IAM) permissions for your SageMaker IAM role. This role is used by the SageMaker environment, and grants permissions to run certain actions. As with all permissions, make sure you follow the best practice of only granting the permissions necessary to perform your tasks.

1. On the IAM console, choose Roles in the navigation pane.
2. Choose the role named AmazonSageMaker-ExecutionRole-<id>. This should be the role that is assigned to your domain.
3. In the Permissions policies section, choose the policy named AmazonSageMaker-ExecutionPolicy-<id>.
4. Choose Edit to edit the customer managed policy.
5. Add the following permissions to the policy, then choose Next.
6. Choose Save changes to confirm your added permissions.

You now have the proper permissions to run commands in your SageMaker environment.

7. Navigate back to your notebook named kohya-ss-fine-tuning.ipynb in your JupyterLab environment.
8. In the notebook step labeled Step One – Create the necessary resources through AWS CloudFormation, run the code cell to create the CloudFormation stack.

Wait for the CloudFormation stack to finish creating before moving on. You can monitor the status of the stack creation on the AWS CloudFormation console. This step should take about 2 minutes.

Set up your custom images and fine-tuning configuration file

In this section, you first upload your fine-tuning configuration file to Amazon S3. The configuration file is specific to the Kohya program. Its purpose is to specify the configuration settings programmatically rather than manually using the Kohya GUI.

This file is provided with opinionated values. You can modify the configuration file with different values if desired. For information about what the parameters mean, refer to LoRA training parameters. You will need to experiment to achieve the desired result. Some parameters rely on underlying hardware and GPU (for example, mixed_precision=bf16 or xformers). Make sure your training instance has the proper hardware configuration to support the parameters you select.

You also need to upload a set of images to Amazon S3. If you don’t have your own dataset and decide to use images from public sources, make sure to adhere to copyright and license restrictions.

The structure of the S3 bucket is as follows:

bucket/0001-dataset/kohya-sdxl-config.toml

bucket/0001-dataset/<asset-folder-name>/     (images and caption files go here)

bucket/0002-dataset/kohya-sdxl-config.toml

bucket/0002-dataset/<asset-folder-name>/     (images and captions files go here)

...

The asset-folder-name uses a special naming convention, which is defined later in this post. Each xxxx-dataset prefix can contain separate datasets with different config file contents. Each pipeline takes a single dataset as input. The config file and asset folder will be downloaded by the SageMaker training job during the training step.

Complete the following steps:

1. Navigate back to your notebook named kohya-ss-fine-tuning.ipynb in your JupyterLab environment.
2. In the Notebook step labeled Step Two – Upload the fine-tuning configuration file, run the code cell to upload the config file to Amazon S3.
3. Verify that you have an S3 bucket named sagemaker-kohya-ss-fine-tuning-<account id>, with a 0001-dataset prefix containing the kohya-sdxl-config.tomlfile.

Next, you create an asset folder and upload your custom images and caption files to Amazon S3. The asset-folder-name must be named according to the required naming convention. This naming convention is what defines the number of repetitions and the trigger word for the prompt. The trigger word is what identifies your custom subject. For example, a folder name of 60_dwjz signifies 60 repetitions with the trigger prompt word dwjz. Consider using initials or abbreviations of your subject for the trigger word so it doesn’t collide with existing words. For example, if your subject is a tiger, you could use the trigger word tgr. More repetitions don’t always translate to better results. Experiment to achieve your desired result.

4. On the S3 console, navigate to the bucket named sagemaker-kohya-ss-fine-tuning-<account id>.
5. Choose the prefix named 0001-dataset.
6. Choose Create folder.
7. Enter a folder name for your assets using the naming convention (for example, 60_dwjz) and choose Create folder.
8. Choose the prefix. This is where your images and caption files go.
9. Choose Upload.
10. Choose Add files, choose your image files, then choose Upload.

When selecting images to use, favor quality over quantity. Some preprocessing of your image assets might be beneficial, such as cropping a person if you are fine-tuning a human subject. For this example, we used approximately 30 images for a human subject with great results. Most of them were high resolution, and cropped to include the human subject only—head and shoulders, half body, and full body images were included but not required.

Optionally, you can use caption files to assist your model in understanding your prompts better. Caption files have the .caption extension, and its contents describe the image (for example, dwjz wearing a vest and sunglasses, serious facial expression, headshot, 50mm). The image file names should match the corresponding (optional) caption file names. Caption files are highly encouraged. Upload your caption files to the same prefix as your images.

At the end of your upload, your S3 prefix structure should look similar to the following:

bucket/0001-dataset/kohya-sdxl-config.toml

bucket/0001-dataset/60_dwjz/

bucket/0001-dataset/60_dwjz/1.jpg

bucket/0001-dataset/60_dwjz/1.caption

bucket/0001-dataset/60_dwjz/2.jpg

bucket/0001-dataset/60_dwjz/2.caption

...

There are many variables to fine-tuning, and as of this writing there are no definitive recommendations for generating great results. To achieve good results, include enough steps in the training, good resolution assets, and enough images.

Set up the required code

The code required for this solution is provided and will be uploaded to the CodeCommit repository that was created by the CloudFormation template. This code is used to build the custom training container. Any updates to the code in this repository will invoke the container image to be built and pushed to Amazon ECR through an EventBridge rule.

The code consists of the following components:

• buildspec.yml – Creates the container image by using the GitHub repository for Kohya SS, and pushes the training image to Amazon ECR
• Dockerfile – Used to override the Dockerfile in the Kohya SS project, which is slightly modified to be used with SageMaker training
• train.py – Initiates the Kohya SS program to do the fine-tuning, and is invoked when the SageMaker training job runs

Complete the following steps to create the training container image:

1. Navigate back to your notebook named kohya-ss-fine-tuning.ipynb in your JupyterLab environment.
2. In the step labeled Step Three – Upload the necessary code to the AWS CodeCommit repository, run the code cell to upload the required code to the CodeCommit repository.

This event will initiate the process that creates the training container image and uploads the image to Amazon ECR.

3. On the CodeBuild console, locate the project named kohya-ss-fine-tuning-build-container.

Latest build status should display as In progress. Wait for the build to finish and the status to change to Succeeded. The build takes about 15 minutes.

A new training container image is now available in Amazon ECR. Every time you make a change to the code in the CodeCommit repository, a new container image will be created.

Initiate the model training

Now that you have a training container image, you can use SageMaker Pipelines with a training step to train your model. SageMaker Pipelines enables you to build powerful multi-step pipelines. There are many step types provided for you to extend and orchestrate your workflows, allowing you to evaluate models, register models, consider conditional logic, run custom code, and more. The following steps are used in this pipeline:

• Condition step – Evaluate input parameters. If successful, proceed with the training step. If not successful, proceed with the fail step. This step validates that the training volume size is at least 50 GB. You could extend this logic to only allow specific instance types, to only allow specific training containers, and add other guardrails if applicable.
• Training Step – Run a SageMaker training job, given the input parameters.
• Fail step – Stop the pipeline and return an error message.

Complete the following steps to initiate model training:

1. On the SageMaker Studio console, in the navigation pane, choose Pipelines.
2. Choose the pipeline named kohya-ss-fine-tuning-pipeline.
3. Choose Create to create a pipeline run.
4. Enter a name, description (optional), and any desired parameter values.
5. You can keep the default settings of using the 0001-dataset for the input data and an ml.g5.8xlarge instance type for training.
6. Choose Create to invoke the pipeline.

7. Choose the current pipeline run to view its details.
8. In the graph, choose the pipeline step named TrainNewFineTunedModel to access the pipeline run information.

The Details tab displays metadata, logs, and the associated training job. The Overview tab displays the output model location in Amazon S3 when training is complete (note this Amazon S3 location for use in later steps). SageMaker processes the training output by uploading the model in the /opt/ml/model directory of the training container to Amazon S3, in the location specified by the training job.

Wait for the pipeline status to show as Succeeded before proceeding to the next step.

Run inference on a custom Stable Diffusion XL model

There are many options for model hosting. For this post, we demonstrate how to run inference with Automatic1111 Stable Diffusion web UI running on an EC2 instance. This tool enables you to use various image generation features through a user interface. It’s a straightforward way to learn the parameters available in a visual format and experiment with supplementary features. For this reason, we demonstrate using this tool as part of this post. However, you can also use SageMaker to host an inference endpoint, and you have the option to use your own custom inference container.

Install the Automatic1111 Stable Diffusion web UI on Amazon EC2

Complete the following steps to install the web UI:

1. Create an EC2 Windows instance and connect to it. For instructions, see Get started with Amazon EC2.
2. Choose Windows Server 2022 Base Amazon Machine Image, a g5.8xlarge instance type, a key pair, and 100 GiB of storage. Alternatively, you can use your local machine.
3. Install NVIDIA drivers to enable the GPU. This solution has been tested with the Data Center Driver for Windows version 551.78.
4. Install the Automatic1111 Stable Diffusion web UI using the instructions in the Automatic Installation on Windows section in the GitHub repo. This solution has been tested with version 1.9.3. The last step of installation will ask you to run webui-user.bat, which will install and launch the Stable Diffusion UI in a web browser.

5. Download the Stable Diffusion XL 1.0 Base model from Hugging Face.
6. Move the downloaded file sd_xl_base_1.0.safetensors to the directory ../stable-diffusion-webui/models/Stable-diffusion/.
7. Scroll to the bottom of the page and choose Reload UI.
8. Choose sd_xl_base_1.0.safetensors on the Stable Diffusion checkpoint dropdown menu.
9. Adjust the default Width and Height values to 1024 x 1024 for better results.
10. Experiment with the remaining parameters to achieve your desired result. Specifically, try adjusting the settings for Sampling method, Sampling steps, CFG Scale, and Seed.

The input prompt is extremely important to achieve great results. You can add extensions to assist with your creative workflow. This style selector extension is great at supplementing prompts.

11. To install this extension, navigate to the Extensions tab, choose Install from URL, enter the style selector extension URL, and choose Install.
12. Reload the UI for changes to take effect.

You will notice a new section called SDXL Styles, which you can select from to add to your prompts.

13. Download the fine-tuned model that was created by the SageMaker pipeline training step.

The model is stored in Amazon S3 with the file name model.tar.gz.

14. You can use the Share with a presigned URL option to share as well.

15. Unzip the contents of the model.tar.gz file (twice) and copy the custom_lora_model.safetensors LoRA model file to the directory ../stable-diffusion-webui/models/Lora.
16. Choose the Refresh icon on the Lora tab to verify that your custom_lora_model is available.

17. Choose custom_lora_model, and it will populate the prompt input box with the text <lora:custom_lora_model:1>.
18. Append a prompt to the text (see examples in the next section).
19. You can decrease or increase the multiplier of your LoRA model by changing the 1 value. This adjusts the influence of your LoRA model accordingly.
20. Choose Generate to run inference against your fine-tuned LoRA model.

Example results

These results are from a fine-tuned model trained on 39 high-resolution images of the author, using the provided code and configuration files in this solution. Caption files were written for each of these images, using the trigger word aallzz.

	Prompt: concept art <lora:custom_lora_model:1.0> aallzz professional headshot, cinematic, bokeh, dramatic lighting, shallow depth of field, vignette, highly detailed, high budget, 8k, cinemascope, moody, epic, gorgeous, digital artwork, illustrative, painterly, matte painting Negative Prompt: photo, photorealistic, realism, anime, abstract, glitch Sampler: DPM2 Sampling Steps: 90 CFG Scale: 8.5 Width/Height: 1024×1024
	Prompt: cinematic film still <lora:custom_lora_model:1> aallzz eating a burger, cinematic, bokeh, dramatic lighting, shallow depth of field, vignette, highly detailed, high budget, cinemascope, moody, epic, gorgeous, film grain, grainy Negative Prompt: anime, cartoon, graphic, painting, graphite, abstract, glitch, mutated, disfigured Sampler: DPM2 Sampling Steps: 70 CFG Scale: 8 Width/Height: 1024×1024
	Prompt: concept art <lora:custom_lora_model:1> aallzz 3D profile picture avatar, vector icon, character, mountain background, sun backlight, digital artwork, illustrative, painterly, matte painting, highly detailed Negative Prompt: photo, photorealistic, realism, glitch, mutated, disfigured, glasses Sampler: DPM2 Sampling Steps: 100 CFG Scale: 9 Width/Height: 1024×1024
	Prompt: concept art <lora:custom_lora_model:1> aallzz 3D profile picture avatar, vector icon, vector illustration, vector art, realistic cartoon character, professional attire, digital artwork, illustrative, painterly, matte painting, highly detailed Negative Prompt: photo, photorealistic, realism, glitch, mutated, disfigured, glasses, hat Sampler: DPM2 Sampling Steps: 100 CFG Scale: 10 Width/Height: 1024×1024
	Prompt: cinematic photo <lora:custom_lora_model:1> aallzz portrait, sitting, magical elephant with large tusks, wearing safari clothing, majestic scenery in the background, river, natural lighting, 50mm, highly detailed, photograph, film, bokeh, professional, 4k, highly detailed Negative Prompt: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, glitch, mutated, disfigured, glasses, hat Sampler: DPM2 Sampling Steps: 100 CFG Scale: 9.5 Width/Height: 1024×1024

Clean up

To avoid incurring charges, delete the resources you created as part of this solution:

1. Delete the objects in your S3 bucket. You must delete the objects before deleting the stack.
2. Delete your container image in Amazon ECR. You must delete the image before deleting the stack.
3. On the AWS CloudFormation console, delete the stack named kohya-ss-fine-tuning-stack.
4. If you created an EC2 instance for running inference, stop or delete the instance.
5. Stop or delete your SageMaker Studio instances, applications, and spaces.

Conclusion

Congratulations! You have successfully fine-tuned a custom LoRA model to be used with Stable Diffusion XL 1.0. We created a custom training Docker container, fine-tuned a custom LoRA model to be used with Stable Diffusion XL, and used the resulting model to generate creative and unique images. The end-to-end training solution was fully automated with a CloudFormation template to help you get started quickly. Now, try creating a custom model with your own subject. To explore more AI use cases, visit the AI Use Case Explorer.

About the Author

Alen Zograbyan is a Sr. Solutions Architect at Amazon Web Services. He currently serves media and entertainment customers, and has expertise in software engineering, DevOps, security, and AI/ML. He has a deep passion for learning, teaching, and photography.

Read More

Foreigners and expats living outside of their home country deal with a large number of emails in various languages daily. They often find themselves struggling with language barriers when it comes to setting up reminders for events like business gatherings and customer meetings. To solve this problem, this post shows you how to apply AWS services such as Amazon Bedrock, AWS Step Functions, and Amazon Simple Email Service (Amazon SES) to build a fully-automated multilingual calendar artificial intelligence (AI) assistant. It understands the incoming messages, translates them to the preferred language, and automatically sets up calendar reminders.

Amazon Bedrock is a fully managed service that makes foundation models (FMs) from leading AI startups and Amazon available through an API, so you can choose from a wide range of FMs to find the model that’s best suited for your use case. With Amazon Bedrock, you can get started quickly, privately customize FMs with your own data, and easily integrate and deploy them into your applications using AWS tools without having to manage any infrastructure.

AWS Step Functions is a visual workflow service that helps developers build distributed applications, automate processes, orchestrate microservices, and create data and machine learning (ML) pipelines. It lets you orchestrate multiple steps in the pipeline. The steps could be AWS Lambda functions that generate prompts, parse foundation models’ output, or send email reminders using Amazon SES. Step Functions can interact with over 220 AWS services, including optimized integrations with Amazon Bedrock. Step Functions pipelines can contain loops, map jobs, parallel jobs, conditions, and human interaction, which can be useful for AI-human interaction scenarios.

This post shows you how to quickly combine the flexibility and capability of both Amazon Bedrock FMs and Step Functions to build a generative AI application in a few steps. You can reuse the same design pattern to implement more generative AI applications with low effort. Both Amazon Bedrock and Step Functions are serverless, so you don’t need to think about managing and scaling the infrastructure.

The source code and deployment instructions are available in the Github repository.

Overview of solution

Figure 1: Solution architecture

As shown in Figure 1, the workflow starts from the Amazon API Gateway, then goes through different steps in the Step Functions state machine. Pay attention to how the original message flows through the pipeline and how it changes. First, the message is added to the prompt. Then, it is transformed into structured JSON by the foundation model. Finally, this structured JSON is used to carry out actions.

1. The original message (example in Norwegian) is sent to a Step Functions state machine using API Gateway.
2. A Lambda function generates a prompt that includes system instructions, the original message, and other needed information such as the current date and time. (Here’s the generated prompt from the example message).
   • Sometimes, the original message might not specify the exact date but instead says something like “please RSVP before this Friday,” implying the date based on the current context. Therefore, the function inserts the current date into the prompt to assist the model in interpreting the correct date for this Friday.
3. Invoke the Bedrock FM to run the following tasks, as outlined in the prompt, and pass the output to the next step to the parser:
   • Translate and summarize the original message in English.
   • Extract events information such as subject, location, and time from the original message.
   • Generate an action plan list for events. For now, the instruction only asks the FM to generate action plan for sending calendar reminder emails for attending an event.
4. Parse the FM output to ensure it has a valid schema. (Here’s the parsed result of the sample message.)
   • Anthropic Claude on Amazon Bedrock can control the output format and generate JSON, but it might still produce the result as “this is the json {…}.” To enhance robustness, we implement an output parser to ensure adherence to the schema, thereby strengthening this pipeline.
5. Iterate through the action-plan list and perform step 6 for each item. Every action item follows the same schema:

   {
             "tool_name": "create-calendar-reminder",
             "parameters": {
               "body": "Jeff, the CEO, invites employees to ...",
               "raw_body": "Kjære ansatte,nnVi ..",
               "subject": "Winter fun and team building event",
               "start_datetime": "2024-05-30T10:00:00Z",
               "end_datetime": "2024-05-30T23:00:00Z",
               "location": "Holmenkollbakken, Oslo"
             }
   }

6. Choose the right tool to do the job:
   • If the tool_name equals create-calendar-reminder, then run sub-flow A to send out a calendar reminder email using Lambda Function.
   • For future support of other possible jobs, you can expand the prompt to create a different action plan (assign different values to tool_name), and run the appropriate action outlined in sub-flow B.
7.  Done.

Prerequisites

To run this solution, you must have the following prerequisites:

• An AWS account
• Enable model access to the Anthropic Claude 3 Sonnet model on Amazon Bedrock in the deployment AWS Region by following the steps in Amazon Bedrock access setup.
• Create and verify identities in Amazon SES: If your Amazon SES is in sandbox mode, you must verify the email addresses of the sender and recipient.

Deployment and testing

Thanks to AWS Cloud Development Kit (AWS CDK), you can deploy the full stack with a single command line by following the deployment instructions from the Github repository. The deployment will output the API Gateway endpoint URL and an API key.

Use a tool such as curl to send messages in different languages to API Gateway for testing:

apigw=[THE_VALUE_OF_GenaiCalendarAgentStack.APIUrl]
apikey=[THE_VALUE_OF_GenaiCalendarAgentStack.GeneratedAPIKeyValue]
curl -v $apigw --header "Content-Type: application/json" --header "x-api-key:$apikey" -d @./doc/sample-inputs/norsk1.json 

Within 1–2 minutes, email invitations should be sent to the recipient from your sender email address, as shown in Figure 2.

Figure 2: Email generated by the solution

Cleaning up

To avoid incurring future charges, delete the resources by running the following command in the root path of the source code:

$ cdk destroy

Future extension of the solution

In the current implementation, the solution only sends out calendar reminder emails; the prompt only instructs the foundation model to generate action items where tool_name equals create-calendar-reminder. You can extend the solution to support more actions. For example, automatically send an email to the event originator and politely decline it if the event is in July (summer vacation for many):

1. Modify the prompt instruction: If the event date is in July, create an action item and set the value of tool_name to send-decline-mail.
2. Similar to the sub-flow A, create a new sub-flow C where tool_name matches send-decline-mail:
   1. Invoke the Amazon Bedrock FM to generate email content explaining that you cannot attend the event because it’s in July (summer vacation).
   2. Invoke a Lambda function to send out the decline email with the generated content.

In addition, you can experiment with different foundation models on Amazon Bedrock, such as Meta Llma 3 or Mistral AI, for better performance or lower cost. You can also explore Agents for Amazon Bedrock, which can orchestrate and run multistep tasks.

Conclusion

In this post, we explored a solution pattern for using generative AI within a workflow. With the flexibility and capabilities offered by both Amazon Bedrock FMs and AWS Step Functions, you can build a powerful generative AI assistant in a few steps. This assistant can streamline processes, enhance productivity, and handle various tasks efficiently. You can easily modify or upgrade its capacity without being burdened by the operational overhead of managed services.

You can find the solution source code in the Github repository and deploy your own multilingual calendar assistant by following the deployment instructions.

Check out the following resources to learn more:

• Visit aws community to discover how our builder communities are using Amazon Bedrock in their solutions.
• Learn more about Generative AI on AWS
• Learn more about Amazon Bedrock

About the Author

Feng Lu is a Senior Solutions Architect at AWS with 20 years professional experience. He is passionate about helping organizations to craft scalable, flexible, and resilient architectures that address their business challenges. Currently, his focus lies in leveraging Artificial Intelligence (AI) and Internet of Things (IoT) technologies to enhance the intelligence and efficiency of our physical environment.

Read More

Generative AI and transformer-based large language models (LLMs) have been in the top headlines recently. These models demonstrate impressive performance in question answering, text summarization, code, and text generation. Today, LLMs are being used in real settings by companies, including the heavily-regulated healthcare and life sciences industry (HCLS). The use cases can range from medical information extraction and clinical notes summarization to marketing content generation and medical-legal review automation (MLR process). In this post, we explore how LLMs can be used to design marketing content for disease awareness.

Marketing content is a key component in the communication strategy of HCLS companies.  It’s also a highly non-trivial balance exercise, because the technical content should be as accurate and precise as possible, yet engaging and empowering for the target audience. The main goal of the marketing content is to raise awareness about certain health conditions and disseminate knowledge of possible therapies among patients and healthcare providers. By accessing up-to-date and accurate information, healthcare providers can adapt their patients’ treatment in a more informed and knowledgeable way. However, medical content being highly sensitive, the generation process can be relatively slow (from days to weeks), and may go through numerous peer-review cycles, with thorough regulatory compliance and evaluation protocols.

Could LLMs, with their advanced text generation capabilities, help streamline this process by assisting brand managers and medical experts in their generation and review process?

To answer this question, the AWS Generative AI Innovation Center recently developed an AI assistant for medical content generation. The system is built upon Amazon Bedrock and leverages LLM capabilities to generate curated medical content for disease awareness. With this AI assistant, we can effectively reduce the overall generation time from weeks to hours, while giving the subject matter experts (SMEs) more control over the generation process. This is accomplished through an automated revision functionality, which allows the user to interact and send instructions and comments directly to the LLM via an interactive feedback loop. This is especially important since the revision of content is usually the main bottleneck in the process.

Since every piece of medical information can profoundly impact the well-being of patients, medical content generation comes with additional requirements and hinges upon the content’s accuracy and precision. For this reason, our system has been augmented with additional guardrails for fact-checking and rules evaluation. The goal of these modules is to assess the factuality of the generated text and its alignment with pre-specified rules and regulations. With these additional features, you have more transparency and control over the underlying generative logic of the LLM.

This post walks you through the implementation details and design choices, focusing primarily on the content generation and revision modules. Fact-checking and rules evaluation require special coverage and will be discussed in an upcoming post.

Image 1: High-level overview of the AI-assistant and its different components

Architecture

The overall architecture and the main steps in the content creation process are illustrated in Image 2. The solution has been designed using the following services:

• Amazon Elastic Container Service (ECS): to deploy and manage our Streamlit UI.
• Amazon Lambda: to run the backend code, which encompasses the generative logic.
• Amazon Textract: for documents parsing, text, and layout extraction.
• Amazon Bedrock: to interact with supported LLMs and embedding models.
• Amazon Translate: for content translation.
• Amazon Simple Storage Service (S3): for documents and processed data caching.

Image 2: Content generation steps

The workflow is as follows:

• In step 1, the user selects a set of medical references and provides rules and additional guidelines on the marketing content in the brief.
• In step 2, the user interacts with the system through a Streamlit UI, first by uploading the documents and then by selecting the target audience and the language.
• In step 3, the frontend sends the HTTPS request via the WebSocket API and API gateway and triggers the first Amazon Lambda function.
• In step 5, the lambda function triggers the Amazon Textract to parse and extract data from pdf documents.
• The extracted data is stored in an S3 bucket and then used as in input to the LLM in the prompts, as shown in steps 6 and 7.
• In step 8, the Lambda function encodes the logic of the content generation, summarization, and content revision.
• Optionally, in step 9, the content generated by the LLM can be translated to other languages using the Amazon Translate.
• Finally, the LLM generates new content conditioned on the input data and the prompt. It sends it back to the WebSocket via the Lambda function.

Preparing the generative pipeline’s input data

To generate accurate medical content, the LLM is provided with a set of curated scientific data related to the disease in question, e.g. medical journals, articles, websites, etc. These articles are chosen by brand managers, medical experts and other SMEs with adequate medical expertise.

The input also consists of a brief, which describes the general requirements and rules the generated content should adhere to (tone, style, target audience, number of words, etc.). In the traditional marketing content generation process, this brief is usually sent to content creation agencies.

It is also possible to integrate more elaborate rules or regulations, such as the HIPAA privacy guidelines for the protection of health information privacy and security. Moreover, these rules can either be general and universally applicable or they can be more specific to certain cases. For example, some regulatory requirements may apply to some markets/regions or a particular disease. Our generative system allows a high degree of personalization so you can easily tailor and specialize the content to new settings, by simply adjusting the input data.

The content should be carefully adapted to the target audience, either patients or healthcare professionals. Indeed, the tone, style, and scientific complexity should be chosen depending on the readers’ familiarity with medical concepts. The content personalization is incredibly important for HCLS companies with a large geographical footprint, as it enables synergies and yields more efficiencies across regional teams.

From a system design perspective, we may need to process a large number of curated articles and scientific journals. This is especially true if the disease in question requires sophisticated medical knowledge or relies on more recent publications. Moreover, medical references contain a variety of information, structured in either plain text or more complex images, with embedded annotations and tables. To scale the system, it is important to seamlessly parse, extract, and store this information. For this purpose, we use Amazon Textract, a machine learning (ML) service for entity recognition and extraction.

Once the input data is processed, it is sent to the LLM as contextual information through API calls. With a context window as large as 200K tokens for Anthropic Claude 3, we can choose to either use the original scientific corpus, hence improving the quality of the generated content (though at the price of increased latency), or summarize the scientific references before using them in the generative pipeline.

Medical reference summarization is an essential step in the overall performance optimization and is achieved by leveraging LLM summarization capabilities. We use prompt engineering to send our summarization instructions to the LLM. Importantly, when performed, summarization should preserve as much article’s metadata as possible, such as the title, authors, date, etc.

Image 3: A simplified version of the summarization prompt

To start the generative pipeline, the user can upload their input data to the UI. This will trigger the Textract and optionally, the summarization Lambda functions, which, upon completion, will write the processed data to an S3 bucket. Any subsequent Lambda function can read its input data directly from S3. By reading data from S3, we avoid throttling issues usually encountered with Websockets when dealing with large payloads.

Image 4: A high-level schematic of the content generation pipeline

Content Generation

Our solution relies primarily on prompt engineering to interact with Bedrock LLMs. All the inputs (articles, briefs and rules) are provided as parameters to the LLM via a LangChain PrompteTemplate object. We can guide the LLM further with few-shot examples illustrating, for instance, the citation styles. Fine-tuning – in particular, Parameter-Efficient Fine-Tuning techniques – can specialize the LLM further to the medical knowledge and will be explored at a later stage.

Image 5: A simplified schematic of the content generation prompt

Our pipeline is multilingual in the sense it can generate content in different languages. Claude 3, for example, has been trained on dozens of different languages besides English and can translate content between them. However, we recognize that in some cases, the complexity of the target language may require a specialized tool, in which case, we may resort to an additional translation step using Amazon Translate.

Image 6: Animation showing the generation of an article on Ehlers-Danlos syndrome, its causes, symptoms, and complications

Content Revision

Revision is an important capability in our solution because it enables you to further tune the generated content by iteratively prompting the LLM with feedback. Since the solution has been designed primarily as an assistant, these feedback loops allow our tool to seamlessly integrate with existing processes, hence effectively assisting SMEs in the design of accurate medical content. The user can, for instance, enforce a rule that has not been perfectly applied by the LLM in a previous version, or simply improve the clarity and accuracy of some sections. The revision can be applied to the whole text.  Alternatively, the user can choose to correct individual paragraphs. In both cases, the revised version and the feedback are appended to a new prompt and sent to the LLM for processing.

Image 7: A simplified version of the content revision prompt

Upon submission of the instructions to the LLM, a Lambda function triggers a new content generation process with the updated prompt. To preserve the overall syntactic coherence, it is preferable to re-generate the whole article, keeping the other paragraphs untouched. However, one can improve the process by re-generating only those sections for which feedback has been provided. In this case, proper attention should be paid to the consistency of the text. This revision process can be applied recursively, by improving upon the previous versions, until the content is deemed satisfactory by the user.

Image 8: Animation showing the revision of the Ehlers-Danlos article. The user can ask, for example, for additional information

Conclusion

With the recent improvements in the quality of LLM-generated text, generative AI has become a transformative technology with the potential to streamline and optimize a wide range of processes and businesses.

Medical content generation for disease awareness is a key illustration of how LLMs can be leveraged to generate curated and high-quality marketing content in hours instead of weeks, hence yielding a substantial operational improvement and enabling more synergies between regional teams. Through its revision feature, our solution can be seamlessly integrated with existing traditional processes, making it a genuine assistant tool empowering medical experts and brand managers.

Marketing content for disease awareness is also a landmark example of a highly regulated use case, where precision and accuracy of the generated content are critically important. To enable SMEs to detect and correct any possible hallucination and erroneous statements, we designed a factuality checking module with the purpose of detecting potential misalignment in the generated text with respect to source references.

Furthermore, our rule evaluation feature can help SMEs with the MLR process by automatically highlighting any inadequate implementation of rules or regulations. With these complementary guardrails, we ensure both scalability and robustness of our generative pipeline, and consequently, the safe and responsible deployment of AI in industrial and real-world settings.

Bibliography

• Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, & Illia Polosukhin. (2023). Attention Is All You Need.
• Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, & Dario Amodei. (2020). Language Models are Few-Shot Learners.
• Mesko, B., & Topol, E. (2023). The imperative for regulatory oversight of large language models (or generative AI) in healthcare. NPJ digital medicine, 6, 120.
• Clusmann, J., Kolbinger, F.R., Muti, H.S. et al. The future landscape of large language models in medicine. Commun Med 3, 141 (2023). https://doi.org/10.1038/s43856-023-00370-1
• Kai He, Rui Mao, Qika Lin, Yucheng Ruan, Xiang Lan, Mengling Feng, & Erik Cambria. (2023). A Survey of Large Language Models for Healthcare: from Data, Technology, and Applications to Accountability and Ethics.
• Mu W, Muriello M, Clemens JL, Wang Y, Smith CH, Tran PT, Rowe PC, Francomano CA, Kline AD, Bodurtha J. Factors affecting quality of life in children and adolescents with hypermobile Ehlers-Danlos syndrome/hypermobility spectrum disorders. Am J Med Genet A. 2019 Apr;179(4):561-569. doi: 10.1002/ajmg.a.61055. Epub 2019 Jan 31. PMID: 30703284; PMCID: PMC7029373.
• Berglund B, Nordström G, Lützén K. Living a restricted life with Ehlers-Danlos syndrome (EDS). Int J Nurs Stud. 2000 Apr;37(2):111-8. doi: 10.1016/s0020-7489(99)00067-x. PMID: 10684952.

About the authors

Sarah Boufelja Y. is a Sr. Data Scientist with 8+ years of experience in Data Science and Machine Learning. In her role at the GenAII Center, she worked with key stakeholders to address their Business problems using the tools of machine learning and generative AI. Her expertise lies at the intersection of Machine Learning, Probability Theory and Optimal Transport.

Liza (Elizaveta) Zinovyeva is an Applied Scientist at AWS Generative AI Innovation Center and is based in Berlin. She helps customers across different industries to integrate Generative AI into their existing applications and workflows. She is passionate about AI/ML, finance and software security topics. In her spare time, she enjoys spending time with her family, sports, learning new technologies, and table quizzes.

Nikita Kozodoi is an Applied Scientist at the AWS Generative AI Innovation Center, where he builds and advances generative AI and ML solutions to solve real-world business problems for customers across industries. In his spare time, he loves playing beach volleyball.

Marion Eigner is a Generative AI Strategist who has led the launch of multiple Generative AI solutions. With expertise across enterprise transformation and product innovation, she specializes in empowering businesses to rapidly prototype, launch, and scale new products and services leveraging Generative AI.

Nuno Castro is a  Sr. Applied Science Manager at AWS Generative AI Innovation Center. He leads Generative AI customer engagements, helping AWS customers find the most impactful use case from ideation, prototype through to production. He’s has 17 years experience in the field in industries such as finance, manufacturing, and travel, leading ML teams for 10 years.

Aiham Taleb, PhD, is an Applied Scientist at the Generative AI Innovation Center, working directly with AWS enterprise customers to leverage Gen AI across several high-impact use cases. Aiham has a PhD in unsupervised representation learning, and has industry experience that spans across various machine learning applications, including computer vision, natural language processing, and medical imaging.

Read More