This Splunk Lantern article outlines the steps to monitor Gen AI applications with Splunk Observability Cloud, covering setup with OpenTelemetry, NVIDIA GPU metrics, Python instrumentation, and OpenLIT integration to monitor GenAI applications built with technologies like Python, LLMs (OpenAI's GPT-4o, Anthropic's Claude 3.5 Haiku, Meta’s Llama), NVIDIA GPUs, Langchain, and vector databases (Pinecone, Chroma) using Splunk Observability Cloud. It outlines a six-step process:
- Access Splunk Observability Cloud: Sign up for a free trial if needed.
- Deploy Splunk Distribution of OpenTelemetry Collector: Use a Helm chart to install the collector in Kubernetes.
- Capture NVIDIA GPU Metrics: Utilize the NVIDIA GPU Operator and Prometheus receiver in the OpenTelemetry Collector.
- Instrument Python Applications: Use the Splunk Distribution of OpenTelemetry Python agent for automatic instrumentation and enable Always On Profiling.
- Enhance with OpenLIT: Install and initialize OpenLIT to capture detailed trace data, including LLM calls and interactions with vector databases (with options to disable PII capture).
- Start Using the Data: Leverage the collected metrics and traces, including features like Tag Spotlight, to identify and resolve performance issues (example given: OpenAI rate limits).
The article emphasizes OpenTelemetry's role in GenAI observability and highlights how Splunk Observability Cloud facilitates monitoring these complex applications, providing insights into performance, cost, and potential bottlenecks. It also points to resources for help and further information on specific aspects of the process.
A discussion post on Reddit's LocalLLaMA subreddit about logging the output of running models and monitoring performance, specifically for debugging errors, warnings, and performance analysis. The post also mentions the need for flags to output logs as flat files, GPU metrics (GPU utilization, RAM usage, TensorCore usage, etc.) for troubleshooting and analytics.
