onboard-gb200-1node-tests

Creating 1-node GitHub MR functional tests involves deriving single-node variants from existing 2-node mr-scoped recipes by scanning gpt.yaml and moe.yaml, then adjusting GPU parallelism for world_size reduction from 8 to 4 without changing the global batch size.

Parallelism tuning for 8 to 4 GPU runs reads each candidate model_config.yaml and applies safe PP/EP reductions to adapt the TP/PP/EP/ETP/DP relationships, ensuring the global batch settings remain completely unchanged during the world_size reduction.

You do not need to manually update product scopes when onboarding 1-node GB200 tests, because the process automatically produces gpt-1node.yaml entries or appends to moe-1node.yaml and wires scope: [mr-github, mr-github-slim] where appropriate.

Expanding CI coverage for GB200 GPT and MoE models requires scanning recipe product specs in gpt.yaml and moe.yaml for candidate mr and mr-slim entries, excluding those already present as *-1node.yaml, and generating the missing *_1node test cases.

Existing 2-node MR tests are excluded during 1-node test generation if matching *-1node.yaml files are already present, preventing duplicate recipe entries and ensuring only missing *_1node test cases are generated for the CI expansion.