Update DeepSeek-V3 B300 recipe

[scsudhakaran]

This PR updates B300 perf configs to align with #4458

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[claude]

Review

Light review of the DeepSeek V3 B300 perf-recipe retuning. Two issues, one CI-blocking.

Critical - breaks perf-recipe env unit tests

deepseek_v3_pretrain_256gpu_b300_fp8mx_large_scale_config was refactored to derive from the fp8mx config and no longer assigns cfg.env_vars inline. tests/unit_tests/recipes/test_perf_recipe_environment.py requires every canonical flat recipe builder to declare its env via an inline cfg.env_vars assignment:

• test_every_flat_recipe_builder_declares_its_environment_inline (via _explicit_environment, asserts exactly one inline cfg.env_vars assignment)
• test_explicit_environment_invariants_across_all_flat_recipes

The builder name matches the canonical regex, so both tests will fail. Every other large_scale config (b200/gb200/gb300) still keeps the inline env dict for this reason. Keep an inline cfg.env_vars here.

Low - stale docstring on nvfp4 config

deepseek_v3_pretrain_256gpu_b300_nvfp4_config docstring still says (PP=16 matching base layout), but the refactor dropped the PP=16/VPP=None overrides and the layout call, so it now inherits PP=8/VPP=2 from the bf16 base. Update the docstring.

Notes (non-blocking)

• bf16/fp8cs/fp8mx configs changed MBS 2 to 1 and VPP None to 2; nvfp4 keeps MBS=2. Layout map has a (8, 2) entry so the new VPP=2 is covered. Intentional retunings - verify against measured baselines.
• fp8mx now adds NVTE_NORM_FWD/BWD_USE_CUDNN=1 (both present, satisfies the all-or-nothing cuDNN invariant) and switches cuda_graph_impl to full_iteration via _enable_deepseek_full_iteration_mxfp8.

Suggested test cases

• tests/unit_tests/recipes/test_perf_recipe_environment.py::test_every_flat_recipe_builder_declares_its_environment_inline
• tests/unit_tests/recipes/test_perf_recipe_environment.py::test_explicit_environment_invariants_across_all_flat_recipes
• deepseek_v3_pretrain_256gpu_b300_bf16_config
• deepseek_v3_pretrain_256gpu_b300_fp8cs_config
• deepseek_v3_pretrain_256gpu_b300_fp8mx_config
• deepseek_v3_pretrain_256gpu_b300_nvfp4_config
• deepseek_v3_pretrain_256gpu_b300_fp8mx_large_scale_config

(Perf configs touched are the B300 DeepSeek V3 recipes above; exercised by test_all_perf_recipe_factories and test_perf_recipe_environment offline construction tests rather than named _perf functional cases.)

[scsudhakaran]

/ok to test f7f22a1

[scsudhakaran]

/ok to test bec3e1d

[scsudhakaran]

/ok to test 03d9d29