feat: support Nemotron 3 Nano 4B BF16

[binaryaaron]

initial attempt at real nemotron support in nss, mostly aroudn getting mamba to work correctly and addressing toeknizer / prompt construction; this is not meant to merge right away. need to test on DP as well.

Summary

• add exact BF16 support for nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 across metadata, native Transformers training, vLLM generation, and grouped time-series assembly
• use one model-owned response-framing contract across prompt rendering, token budgeting, structured constraints, parsing, and inference evaluation
• add external Mamba kernel bootstrap plus an A100 selective-state-update tuner covering effective batches through 65,536. hopefully we can obviate this step before landing.
• reject unvalidated DP and quantized training combinations; FP8 remains deferred

A100 validation

A full financial time-series run completed on an A100 SXM4 80GB:

• 3,640 valid rows, 95.49% validity
• SQS 9.2 and AIA protection 9.5
• runtime 18m25s
• W&B: https://wandb.ai/nemo-llm-service/aagonzales-sandbox/runs/q5xip7lo
• generated-data artifact: nemotron3-financial-timeseries-confirmed-synthetic-data:v0

The lower-than-expected MIA result is under investigation and will be reported separately.

The A100 tuner generated and validated all nine intended selective-state-update entries from effective batch 128 through 65,536. Numerical validation passed for every shape, with measured kernel-level speedups of 1.07x to 1.43x over the vLLM heuristic. The tuned JSON remains machine-local and is activated with VLLM_TUNED_CONFIG_FOLDER.

todos

• FP8 support
• Differential privacy and quantized training support
• slurm benchmarks
• External kernel builds require a CUDA 12.9 toolchain and remain separate from the standard uv.lock environment.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: fd7a9224-1790-436d-b28c-57d84fb0a0a0

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch binaryaaron/feat/nemotron-support

---

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

❌ Patch coverage is 84.14634% with 130 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
tools/nemotron/tune_vllm_ssu.py	41.77%	92 Missing ⚠️
src/nemo_safe_synthesizer/llm/model_policy.py	68.75%	25 Missing ⚠️
src/nemo_safe_synthesizer/llm/metadata.py	96.05%	3 Missing ⚠️
...o_safe_synthesizer/training/huggingface_backend.py	40.00%	3 Missing ⚠️
tests/smoke/test_training_cpu.py	70.00%	3 Missing ⚠️
..._safe_synthesizer/generation/timeseries_backend.py	92.59%	2 Missing ⚠️
src/nemo_safe_synthesizer/config/autoconfig.py	96.29%	1 Missing ⚠️
tests/tools/test_tune_vllm_ssu.py	98.03%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!