Files
vllm-omni/tests/diffusion/quantization/artifacts/fp8_z_image_quality

FP8 Z-Image Quality Artifact

Generated by:

CUDA_VISIBLE_DEVICES=0 \
VLLM_OMNI_QUALITY_OUTPUT_DIR=$PWD/tests/diffusion/quantization/artifacts/fp8_z_image_quality \
.venv/bin/python -m pytest \
  tests/diffusion/quantization/test_quantization_quality.py::test_quantization_quality[fp8_z_image] \
  -s -v -m ""

Quantized config:

{"method": "fp8", "ignored_layers": ["img_mlp"]}

Result: failed with the current max_lpips=0.15 threshold.

Metric Value
LPIPS 0.250578
PSNR 17.042153 dB
MAE 0.076899

Comparison

BF16 baseline vs FP8 quantized

Individual Outputs

BF16 baseline

FP8 quantized