Here we show current out-of-the-box techniques for training on a single GPU.

Fundamentals of Efficient Training on a Single GPU

How to fine tune and serve your own Mistral 7B model.

Finetuning Mistral 7B

How fast you can train your model depends on your hardware and your parallelism strategy. Knowing your hardware will guide you on which strategy you should use.

How Data Parallelism & Hardware Affect Speed

Use your favorite LLM application frameworks like Llama Index or Langchain with your self-served open source model deployment.

Drop-In Replacement for GPT with Llama 2 for OpenAI API

LLM-ATC handles orchestration for finetuning and serving LLMs. Just bring your data and let LLM-ATC work under the hood on the cloud provider of your choice.

LLM Air Traffic Controller: Easy Finetuning and Serving of Open Source LLMs

Most ML teams use GPU Utilization as their main performance metric, but we found this can be quite misleading.

GPU Utilization is a Misleading Metric

Automatically isolate faulty nodes and schedule only on healthy ones.

Automatic GPU Node Health and Pod Scheduling

Zero Instrumentation GPU Metrics with DCGM on Trainy: Konduktor

Run these benchmarks before training on multiple machines.

GPU Fabric Preflight Checks for MultiNode Training

Function Calling with Gorilla LLM

Generate traces of your entire cluster during distributed training, with little to no impact on training speed.

Feature Update: Trainy Monitoring Daemon

Trainy is your ultimate training optimization companion.

All Posts

Fundamentals of Efficient Training on a Single GPU

Finetuning Mistral 7B

How Data Parallelism & Hardware Affect Speed

Drop-In Replacement for GPT with Llama 2 for OpenAI API

LLM Air Traffic Controller: Easy Finetuning and Serving of Open Source LLMs