Production

• Packages
  • {inference-perf} (JOSS) - A production-scale GenAI inference performance benchmarking tool that allows you to benchmark and analyze the performance of inference deployments. It is agnostic of model servers and can be used to measure performance and compare different systems.
• Resources
  • LLM pricing calculator and model prices (per million tokens)
• Cost
  • LLMs are priced in terms of input and output tokens, with output generally charged at significantly higher rates
  • Reasoning queries (∼5,000 output tokens) raise energy use ∼13× vs. standard queries (source)
  • Hosted Model: input tokens × input-token price + output tokens × output-token price
    • Reasoning causes the model to generate many more internal/output tokens (e.g. 1000s). Then, the API provider can charge substantially more.
• Track per-step accuracy separately from overall task completion.
  • When per-step accuracy drops from 90% to 87% on a 10-step task, overall success rate drops from 35% to 24%.
  • You want to catch that degradation in monitoring, not in a post-incident review
• The Docker Model Runner (DMR) - A new Docker Desktop feature that enables the run of Large Language Models natively with Docker Desktop.
  • Follows the common Docker workflow:
    • Pull models from registries (e.g., Docker Hub)
    • Run models locally with GPU acceleration
    • Integrate models into the development workflows
  • Resources
    • Docs
    • Krispin Series
      • Getting Started with Docker Model Runner
      • Running LLMs Locally with Docker Model Runner and Python
      • Docker Model Runner - Pull LLMs from Hugging Face
      • Running LLMs Locally with Docker Model Runner and R