fix(recipes): MB-922 reduce DeepSeek VR200 memory pressure

[svcnemo-autobot]

Background

MB-922 reports Linux/SLURM OOM kills before iteration 1 in the 256-GPU VR200 BF16 and NVFP4 DeepSeek-V3 jobs. Each rank inherited eight persistent dataloader workers, multiplying host processes and pinned-memory pressure per node.

What changed

• Keep both recipes derived from their GB300 configurations.
• Set dataset.num_workers = 0 for only the affected VR200 recipes.
• Add regression tests for the GB300 topology and worker limit.

Details

The PP/VPP/EP topology, micro-batch sizes, recompute modules, HybridEP domains, and precision settings remain unchanged. Removing dataloader subprocesses targets host RAM—the resource implicated by SLURM oom_kill—without replacing the tuned GB300 model configuration.

Tested

• Ruff format check: passed.
• Ruff lint: passed.
• Focused recipe suites: 423 passed.

The exact 256-GPU VR200 jobs remain for CI hardware validation.

[svcnemo-autobot]

/ok to test a998cbd

[svcnemo-autobot]

The VR200 GB200 → GB300 reassignment was introduced by bf07e5678cb96433467d0d495fc898de4a6849c2 (vr200 cfgs match w/ gb300, PR #4574). It changed the DeepSeek-V3 VR200 BF16 and NVFP4 aliases from their GB200 recipes to GB300. This PR restores the GB200-derived layouts for those two affected recipes.

[ko3n1g]

The VR200 GB200 → GB300 reassignment was introduced by bf07e5678cb96433467d0d495fc898de4a6849c2 (vr200 cfgs match w/ gb300, PR #4574). It changed the DeepSeek-V3 VR200 BF16 and NVFP4 aliases from their GB200 recipes to GB300. This PR restores the GB200-derived layouts for those two affected recipes.

@svcnemo-autobot lets keep the gb300 recipe and find another fix

[svcnemo-autobot]

💬 Amended PR #5086 as requested. It now preserves the GB300-derived topology and reduces host memory by disabling 8 persistent dataloader subprocesses per rank for only the affected VR200 recipes. Formatting, Ruff, and 423 focused recipe tests pass.

[svcnemo-autobot]

/ok to test 7124bd6