x86 Model Prep

Jetson Thor can't do FP8 quantization or ONNX export on its own — those steps live on an x86 GPU host (A100/H100). thor-cosmos ships one recipe that chains the four steps.

One-liner

just prep-edge-model reason2-2b ./models/R2-fp8

Expands to:

just download reason2-2b           ./models/R2-fp8/hf
just quantize ./models/R2-fp8/hf   ./models/R2-fp8/quantized fp16 fp8
just export-llm ./models/R2-fp8/quantized ./models/R2-fp8/onnx
just export-visual ./models/R2-fp8/hf      ./models/R2-fp8/onnx/visual_enc_onnx fp16 fp8
# ✅ ONNX ready → ./models/R2-fp8/onnx  (scp to Thor next)

Shortcuts: model names

Name	HF repo
reason2-2b	nvidia/Cosmos-Reason2-2B
reason2-7b	nvidia/Cosmos-Reason2-7B
predict2.5-2b	nvidia/Cosmos-Predict2.5-2B
predict2.5-14b	nvidia/Cosmos-Predict2.5-14B
transfer2.5-2b	nvidia/Cosmos-Transfer2.5-2B
transfer2.5-edge	nvidia/Cosmos-Transfer2.5-Edge
transfer2.5-depth	nvidia/Cosmos-Transfer2.5-Depth
transfer2.5-seg	nvidia/Cosmos-Transfer2.5-Seg

Any other name is passed through as-is to hf download.

Shortcuts: datasets

Name	HF repo
gr1	nvidia/PhysicalAI-Robotics-GR00T-GR1
gr1-100	nvidia/GR1-100
gr00t-eval	nvidia/PhysicalAI-Robotics-GR00T-Eval
safe-unsafe	pjramg/Safe_Unsafe_Test

just download-dataset gr1 ./datasets/gr1

Quantization precision

just quantize <model_dir> <output_dir> <dtype> <quantization>

dtype	quantization	use case
fp16	fp8	default for Thor — 2x smaller, negligible quality loss
bf16	fp8	alternative base dtype
fp16	int8	older GPUs (no FP8 HW)
fp16	int4	aggressive compression, notable quality drop

ONNX export

just export-llm    <model_dir> <output_dir>
just export-visual <model_dir> <output_dir> <dtype> <quantization>

LLM export takes the quantized weights. Visual export takes the original HF weights — the visual encoder uses its own dtype/quant settings.

Ship to Thor

scp -r ./models/R2-fp8/onnx cagatay@thor.local:~/R2-fp8-onnx

Then on Thor:

just build-engines ~/R2-fp8-onnx ~/R2-fp8-engines

→ Thor deployment