KernelVision YOLOv8n ONNX deployment graphs

This repository publishes the two validated ONNX graphs produced during KernelVision, a correctness-first study of GPU optimization in a YOLOv8n object-detection pipeline.

KernelVision did not train a new detector. Both graphs derive from the official COCO-pretrained Ultralytics YOLOv8n model. They are published as reproducible deployment artifacts so the project is not presented as owning a new set of learned weights.

Files

File	Graph precision	Input	Raw output	Size
yolov8n_fp32.onnx	FP32	FP32 [1, 3, 640, 640]	FP32 [1, 84, 8400]	12.2 MiB
yolov8n_mixed_fp16.onnx	ModelOpt mixed FP16/FP32	FP16 [1, 3, 640, 640]	FP16 [1, 84, 8400]	6.2 MiB

Both graphs use ONNX opset 17 and a static batch size and image shape. The raw output contains pre-NMS box and class predictions; confidence filtering, non-maximum suppression (NMS), and coordinate scaling are not embedded in the graphs.

Provenance

Ultralytics YOLOv8n PyTorch checkpoint
        ↓ export, opset 17
yolov8n_fp32.onnx
        ↓ NVIDIA ModelOpt AutoCast
yolov8n_mixed_fp16.onnx
        ↓ TensorRT 11.2.1.2 build on NVIDIA L4
TensorRT FP16 engine (not distributed)

The TensorRT engines are intentionally excluded because serialized engines are tied to their build environment and target GPU stack.

Correctness evidence

The FP32 ONNX graph was compared with the PyTorch reference over all 705,600 raw output elements:

Check	Result
Shape	[1, 84, 8400]
All-close (atol=1e-4, rtol=1e-4)	Pass
Mean absolute difference	2.981e-06
P99 absolute difference	6.866e-05
Maximum absolute difference	0.004059

The mixed graph was used to build the TensorRT FP16 candidate. Against the PyTorch FP32 reference on the final application-level case, it produced five matched detections with no unmatched boxes, mean matched IoU 0.997469, and a maximum coordinate difference of 0.546 px.

NVIDIA L4 benchmark context

These are results from the KernelVision benchmark harness, not hosted Hugging Face inference measurements:

Backend	Model-only median	Complete-pipeline median
PyTorch FP32	9.529 ms	13.600 ms
TensorRT FP16	1.428 ms	9.664 ms
Median speedup	6.671×	1.407×

The complete pipeline includes image decode, preprocessing, host-to-device transfer, model execution, NMS, and in-memory visualization. It excludes model loading and output-file writing.

Minimal ONNX Runtime example

The caller must supply an already letterboxed, RGB, normalized BCHW tensor:

import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("yolov8n_fp32.onnx")
input_tensor = np.zeros((1, 3, 640, 640), dtype=np.float32)
raw_output = session.run(["output0"], {"images": input_tensor})[0]
print(raw_output.shape)  # (1, 84, 8400)

For actual detections, apply the same letterbox metadata, confidence filtering, NMS, and coordinate scaling described in the KernelVision source repository.

Artifact integrity

File	SHA-256
yolov8n_fp32.onnx	3db80127ae56dae98402da2b3bd11ef2214a61c1f30e69166148087e43f3bc5e
yolov8n_mixed_fp16.onnx	5d319e1d30f30d6e64bcfab3ca86a4d4844b1d8f740f0495d4d211fc0b069759

Limitations

• The graphs accept only batch 1 at 640 × 640.
• The model is the general COCO-pretrained YOLOv8n detector, not a KernelVision-trained model.
• Dataset-wide COCO mAP was not re-evaluated for these exports.
• The FP16 graph is a mixed-precision deployment input for TensorRT; reduced precision does not imply identical raw floating-point values.
• The reported performance covers one Modal-hosted NVIDIA L4 and should not be generalized to other hardware without measurement.

Reproduction and license

• Source, scripts, and raw benchmark records
• Complete final report
• Ultralytics YOLOv8 base model

The artifacts are distributed under AGPL-3.0 in accordance with the upstream Ultralytics licensing terms. See the KernelVision third-party notices for complete attribution.