Metric	Value
Quantization method	nvidia-modelopt FP8 PTQ
Cosine similarity vs FP32 (mean)	0.999
Cosine similarity vs FP32 (min)	0.992
Quantizer nodes	111 (25 Linear + 12 Conv2d)
Save/load roundtrip	1.000 (perfect)
Model size	5.3 MB
GPU tested	RTX 5090 (Blackwell)

Usage

python

import torch, copy
import modelopt.torch.opt as mto
from huggingface_hub import hf_hub_download
from olmoearth_pretrain.model_loader import ModelID, load_model_from_id

model = load_model_from_id(ModelID.OLMOEARTH_V1_NANO, load_weights=True).cuda().eval()
encoder = copy.deepcopy(model.encoder)

ckpt = hf_hub_download("2imi9/olmoearth-nano-fp8", "model.pth")
mto.restore(encoder, ckpt)
encoder.cuda().eval()

output = encoder(sample, patch_size=2)
embedding = output["project_aggregated"]  # [B, 128]

Hardware

FP8: Hopper (H100) or Blackwell (RTX 5090, B100)
CUDA 13.0+
nvidia-modelopt >= 0.31