fix: TP config for nemotron-flash-1b + super-49B vllm_deploy cascade

[adil-a]

What

Fixes two *_vllm_deploy CI tests that cascade from a missing checkpoint:

• nemotron-flash-1b PEFT (job 337980668)
• llama-3.3-nemotron-super-49B SFT (job 337980592)

Root cause

Both deploys error No checkpoint found under .../robustness_checkpoint because the upstream finetune/robustness producer dies first: a custom-code (trust_remote_code) architecture with no registered TP plan now hard-errors at tp_size>1 (newer torch DTensor shard_order assert; the #2244 fail-fast in parallelizer.py). These previously rode AutoModel's default base plan on older torch.

Fixes

• nemotron-flash-1b (nemotron_flash_1b_squad_peft.yaml): NemotronFlash (hybrid mamba2/deltanet) has no TP plan in any transformers version (verified 5.5.0 and latest 5.12.1), in the model's Hub code, or in AutoModel — and its hybrid SSM/conv layers aren't expressible with the standard colwise/rowwise TP styles. The robustness cross-TP phase ran at tp_size=2 and aborted at setup, before the checkpoint was saved. It's a 1B model that doesn't need TP → set the robustness reload to tp_size: 1. (Train→save→AutoModel-reload→HF-reload still validate; only cross-TP-at-2, which never had a real plan, is dropped.)
• super-49B / DeciLM-nemotron-nas (llama3_3_nemotron_super_49B_squad.yaml): AutoModel already ships get_decilm_nemotron_tp_plan (named llama_nemotron_super_tp_plan, since fix: tp plan for nemotron super #1487); the recipe never selected it, so the finetune fell through to the broken default plan at tp_size=4. Wire distributed.tp_plan: llama_nemotron_super_tp_plan. (All 49 real-attention blocks have 8 KV heads → divisible by tp=4 finetune and tp=8 robustness; the 31 no-op attention blocks stay replicated.)

Validation

• Targeted GitLab pipeline (both recipes + their _vllm_deploy, on this commit): https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/54962684 (running)

Pre-checks

• YAML parses; DCO signed off.

🤖 Generated with Claude Code

[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.