Amazon Web Services says Lambda Event Functions can now run on EC2 G-series GPU instances through Lambda Managed Instances. Developers can deploy a container image for work such as model inference, video processing or rendering. AWS says Lambda handles GPU provisioning and scaling.
The new support is aimed at inference rather than model training, according to AWS. It can run custom models, including smaller language models, as well as GPU tasks outside machine learning. Functions can be triggered by AWS event sources such as Amazon S3, Amazon SQS and Amazon EventBridge.
Developers choose GPU instance types and memory
Developers choose the GPU instance types a capacity provider can use and set the minimum accelerator memory a function needs. AWS says Lambda places the function on an available GPU that meets that requirement. Multiple requests can share an execution environment and one copy of a model in GPU memory.
AWS provides container base images with tools including CUDA, PyTorch and vLLM. Developers can also supply their own image. In an example using a small instruction-tuned model from Hugging Face, AWS says four concurrent requests completed in about the same time as one request in a single execution environment.

GPU functions support 90-minute timeouts
GPU functions support up to 90-minute timeouts for asynchronous and event source mapping invocations. AWS lists limits of 32 GB per container image, 32 GB of temporary storage and up to 48 concurrent requests per execution environment.
GPU support for Event Functions on Lambda Managed Instances is available now. AWS advises users to check their G-series instance quota before setting up a capacity provider, because new accounts may have a quota of zero.



