KV cache capacity budget (new)
Offline single-GPU MHA/GQA memory planner. Enter your own measured or estimated model-weight and runtime reservations. Accounts for KV block rounding.
Small, inspectable Python utilities accompanying our technical guides. Download the source, review it, and run it in your own environment. Every tool has explicit assumptions and boundaries. These are not claims of physical GPU testing or independent performance benchmarks.
Inspect memory assumptions and debug runtime signals without running GPU workloads.
Offline single-GPU MHA/GQA memory planner. Enter your own measured or estimated model-weight and runtime reservations. Accounts for KV block rounding.
Classifies common NCCL failure signals in existing logs without executing GPU commands.
Inspects local path candidates for CUDA Toolkit header files. Does not install dependencies.
Bounded diagnostics and synthetic checks for common engineering failures.
Reads previously captured Docker state JSON from standard input; does not invoke Docker.
Inspects a Docker command as plain text for port exposure and privileged options. Does not execute it.
Uses fixed synthetic child processes to test local timeout cleanup behavior.
Sample MCP server for inspected local development. Review source and dependencies before use.
Example local MCP stdio client check. Review command and environment requirements before use.
Creates a disposable local fixture database using synthetic vectors.
Compare previously collected evidence, not fabricated benchmarks.
Compares supplied, previously saved JSON responses from search providers; makes no API calls.
Found an outdated API, incorrect assumption, or broken tool? Report it here. For testing methodology and AI-assisted editorial review, see our standards.