Text encoder validation
Validation date: 2026-07-27
Environment:
- NVIDIA GeForce RTX 4080 Laptop GPU
- ONNX 1.22.0, full model check enabled
- ONNX Runtime 1.20.1
DmlExecutionProviderwith CPU fallback available
The INT4 ONNX output was compared with the stored FP16 pre-quantization reference embedding for each prompt.
| Prompt | Cosine similarity | Max absolute difference | Mean absolute difference |
|---|---|---|---|
walk forward calmly |
0.981569 | 2.437500 | 0.303064 |
walk forward confidently with long energetic strides |
0.982333 | 2.742188 | 0.288607 |
sneak forward cautiously |
0.981617 | 2.507813 | 0.271095 |
wave with the right hand while standing |
0.977209 | 1.906250 | 0.318750 |
jog forward quickly |
0.980100 | 2.945313 | 0.294070 |
Summary:
- minimum cosine similarity: 0.9772086
- mean cosine similarity: 0.9805653
- maximum cosine similarity: 0.9823330
- output shape:
[1, 4096] - all outputs finite