nitrogen-quant

nitrogen-quant — pre-calibrated FP8 / NVFP4 NitroGen DiT exports

Pre-built artifacts for the inference-pipeline-benchmark harness's FP8 / NVFP4 NitroGen rounds. Customers do not need to recalibrate — bench sweep downloads these automatically on first FP8/NVFP4 use.

What's in here

Each folder ships:

• ng_dit.onnx — ~466 MB. The DiT submodule of NitroGen 500M after modelopt PTQ calibration + QDQ-instrumented ONNX export. Quantized layers: DiT + SigLIP-2 vision tower Linears. Excluded layers (kept at compute dtype): action_head, action_decoder, proprio_emb, time_emb, timestep*, norm*, ada_ln*, pos_embed*.
• meta.json — calibration metadata (precision, steps, opset, input shapes).

How to use

In your fork of the benchmark harness:

bench setup --backend nitrogen-quant
bench sweep --gpu rtx_pro6000 --sweep nitrogen-backends
# fp8 / nvfp4 rounds will hf-download from here on first run

Force local recalibration (e.g., your production frame distribution differs meaningfully from these mixed-genre calibration samples):

NITROGEN_FORCE_RECALIBRATE=1 bench sweep ...

Provenance

• Source model: nvidia/NitroGen (ng.pt, June 2026)
• Calibration tool: nvidia-modelopt 0.44.0
• Export tool: PyTorch 2.12.1 torch.onnx.export(dynamo=False), opset 20
• Builder script: scripts/build_nitrogen_artifacts.py
• Manifest (sha256s): benchmarks/nitrogen_artifacts.yaml

License & attribution

These artifacts derive directly from nvidia/NitroGen weights — same NVIDIA Open Model License applies. We've only added QDQ instrumentation; the underlying weights are unchanged.

The NitroGen model itself is the work of NVIDIA / MineDojo. This repo only republishes a quantized export for benchmark convenience.