perf(recipe): add DeepSeek V4 Flash GB300 FP8MX recipes

[malay-nagda]

What does this PR do ?

Adds three DeepSeek V4 Flash 128-GPU GB300 MXFP8 performance recipes corresponding to benchmark rows 17–19.

Changelog

• Add the MBS=1 DeepSeek V4 Flash GB300 MXFP8 recipe.
• Add the MBS=2 MoE activation-recompute variant.
• Add the MBS=2 optimizer-state and expert-FC1 offload variant.
• Export the new recipes from the DeepSeek performance recipe package.
• Add focused unit coverage for topology, batching, CUDA graphs, precision, recompute, and offload settings.

GitHub Actions CI

Validation performed:

• uv run --no-sync pre-commit run --all-files — passed.
• Focused pytest collection could not run in the local macOS environment because torch is not installed. A normal uv project sync is blocked earlier while building the CUDA-only fast-hadamard-transform dependency.

Before your PR is "Ready for review"

Pre checks:

• Read and followed the contributor guidelines.
• Added focused unit tests.
• No additional documentation is required for these exported recipe helpers.
• This PR does not change optional-component import behavior.

Additional Information

The benchmark sheet lists PP=1 and VPP=4, while all three saved W&B runtime configs resolve VPP to null. Megatron-Core requires PP greater than 1 whenever VPP is enabled, so these recipes preserve PP=1 and set VPP to None.

The branch is based on r0.6.0 commit ff3795155a582cc1b7acc5c48c10098cafc123fe.

[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.