refactor(recipe)!: rename vr200 GPU type and perf recipes to vrnvl72

[malay-nagda]

What does this PR do ?

Renames the Vera Rubin NVL72 GPU type from vr200 to vrnvl72 in the performance recipes and launchers. vr200 stays a deprecated CLI alias so existing callers keep working while they move over.

Changelog

• Rename src/megatron/bridge/perf_recipes/<family>/vr200/ to .../vrnvl72/ for the 7 families (deepseek, gpt_oss, kimi, llama, nemotronh, qwen, qwen_vl) and the 53 *_vr200_* recipe functions to *_vrnvl72_*. Family __init__ exports, dump_perf_configs.py, docstrings, READMEs and tests follow.
• GPU handling uses vrnvl72: NUM_GPUS_PER_NODE_MAP, the 4752 MHz --peak_mem_clk default, the NUMA divisor, override and topology sets, and the training launcher's hardware check.
• --gpu vr200 is mapped to vrnvl72 by an argparse type= callable (LEGACY_GPU_ALIASES in scripts/performance/argument_parser.py). Only vrnvl72 is listed in the choices and help. The deprecation warning is logged on the launcher and global rank 0 only, because every rank re-parses the forwarded argv.
• Training launcher: a benchmark recipe name containing _vr200_ resolves to the renamed _vrnvl72_ recipe, with a warning, when that recipe exists (LEGACY_HARDWARE_ALIASES in scripts/training/recipe_metadata.py).
• Tests:
  • New tests/unit_tests/scripts/performance/test_argument_parser.py.
  • Alias cases in training test_recipe_metadata.py, test_recipe_runner.py and test_setup_experiment.py.
  • The performance dry-run test now also runs --gpu vr200. It asserts the canonical experiment name and the 4752,4752 memory-clock lock.

Breaking change (Python API only). Code that imports the old names directly must update:

• modules megatron.bridge.perf_recipes.<family>.vr200.* → megatron.bridge.perf_recipes.<family>.vrnvl72.*;
• the 53 functions *_vr200_*_config → *_vrnvl72_*_config;
• direct calls that pass gpu="vr200" to script helpers without argparse, and NUM_GPUS_PER_NODE_MAP["vr200"].

Nothing in this repo imports the old names. The CLI keeps accepting --gpu vr200 and _vr200_ recipe names during the deprecation window. Removing the alias later means deleting the two alias tables (each comment points to the other), the help and README notes, and the alias test cases.

Validation

• ruff 0.9.9 (check, --select I, format --check): clean on all 56 changed Python files.
• Torch-free unit tests, run locally with --noconftest: 138 passed, 1 failed in tests/unit_tests/scripts/performance/test_argument_parser.py and tests/unit_tests/scripts/training/{test_recipe_metadata,test_recipe_runner,test_setup_experiment}.py.
  • The failure, test_known_cross_package_collisions_have_explicit_benchmark_precedence, also fails on main.
  • tests/unit_tests/scripts is excluded from CI collection (norecursedirs), so these tests ran locally only.
  • The recipe tests that need torch or megatron.core run in CI.
• Static checks:
  • all 53 renamed recipes are exported;
  • the benchmark recipe index has 53 vrnvl72 names and no vr200 names;
  • all 414 dump_perf_configs.py combinations resolve;
  • every family __init__ import resolves.
• Name resolution for the 37 VR performance jobs in our CI: --gpu vrnvl72 and --gpu vr200 both resolve to the same existing *_vrnvl72_* recipes. The renamed recipe bodies are unchanged apart from the name.

GitHub Actions CI

See the CI section in the Contributing doc for how to trigger the CI.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)

Additional Information

• Downstream CI keeps passing --gpu vr200 until its own rename lands, and the alias covers that period.

🤖 Generated with Claude Code

[github-actions]

Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing:

/review

Add model=claude to use a Claude reviewer (the default is model=codex). mode=light|strict selects the review depth; for example, /review model=claude mode=strict. Comment /review help for all options.