perf(nemotron): add Nemotron 3.5 Lightning GB300 recipes

[malay-nagda]

What does this PR do?

Adds canonical Nemotron 3.5 Lightning throughput-benchmark recipes for an 8-GPU GB300 system.

• Adds BF16, MXFP8, NVFP4, and MXFP8 Megatron FSDP variants.
• Exports the recipes through megatron.bridge.perf_recipes.nemotronh and registers them in the performance-config inventory.
• Extends the existing Nemotron 3.5 recipe tests to cover GB300 model finalization, MTP settings, topology, precision-specific tuning, and FSDP defaults.
• Corrects the copied FSDP comment: the GB300 FSDP microbatch is 3, while the non-FSDP GB300 recipes use 4.

These are benchmark recipes for upper-bound throughput measurement, not production convergence recipes.

Testing

• uv run --with pre-commit --no-project pre-commit run --all-files — passed.
• Targeted pytest was attempted but could not run in the local checkout: project sync fails while building fast-hadamard-transform because its isolated build environment does not declare torch, and the existing .venv does not contain pytest or torch.
• Static inventory validation confirmed that all four GB300 inventory entries resolve to defined recipe factories.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[claude]

LGTM

Clean mirror of the existing GB200 Nemotron 3.5 Lightning perf recipes onto GB300, with matching unit-test coverage. Verified the four new config functions differ from their GB200 twins only in the intended env-var/topology set, and that the test registries (_GB300_RECIPES, _GB_FSDP_RECIPES, base-recipe pairs, finalize factories, and the MTP / topology / FSDP-defaults / CutDSL assertions) all reference the new functions. The dump_perf_configs.py entries also include the fsdp variant.

Minor (non-blocking): in _build_nemotron_3_5_lightning_gb300_mxfp8, moe_hybridep_num_sms = 16 is redundant. The GB300 nano MXFP8 builder already sets it to 16 (unlike the GB200 nano MXFP8 builder, which sets 32, where the override is meaningful). Harmless, just a copy artifact.

Suggested test cases:

• nemotron_3_5_lightning_pretrain_8gpu_gb300_bf16_config
• nemotron_3_5_lightning_pretrain_8gpu_gb300_fp8mx_config
• nemotron_3_5_lightning_pretrain_8gpu_gb300_fp8mx_fsdp_config
• nemotron_3_5_lightning_pretrain_8gpu_gb300_nvfp4_config
• nemotron_3_nano_pretrain_8gpu_gb300_bf16_config
• nemotron_3_nano_pretrain_8gpu_gb300_fp8mx_config
• nemotron_3_nano_pretrain_8gpu_gb300_nvfp4_config

[cuichenx]

LGTM