examples/benchmark_overview.md

# Formal Benchmark Overview

## Variant Summary

| Variant | Avg Latency (ms) | Avg Peak VRAM (MiB) | vs Base Latency | vs Base Peak VRAM |
| --- | ---: | ---: | ---: | ---: |
| transformer-int8 + pe-bf16 + use_pe=true | 78053 | 28516 | +0 | +0 |
| transformer-int8 + pe-int8 + use_pe=true | 81412 | 28721 | +3359 | +205 |
| transformer-int8 + use_pe=false | 60287 | 28339 | -17766 | -177 |
| ERNIE-Image-Turbo Reference | 32535 | 35255 | -45518 | +6739 |

## Notes

- `pe-int8` is a runtime-quantized benchmark-only variant and does not change the packaged release precision matrix.
- Peak VRAM is the peak reserved CUDA memory of the current PyTorch process during each generation call.
- The prompt suite contains 7 prompts, including the added portrait case `zh_portrait_studio_east_asian`.