[CI][BugFix] Fix and Validate FP8 Z-Image quality gate

[david6666666]

Summary

Unskips test_quantization_quality[fp8_z_image] and updates the Z-Image FP8 quality prompt to the requested long floating-archipelago prompt.

Following PR #3279's text-encoder FP8 path, this uses regular online FP8 quantization. The text encoder is FP8 for the early/mid blocks, with only the final 8 text-encoder blocks listed in ignored_layers for BF16 fallback. The Z-Image transformer also keeps the final 15 main attention/FFN blocks in BF16 through ignored_layers.

The final PR diff is intentionally small and only changes tests/diffusion/quantization/test_quantization_quality.py. The visual artifacts are linked below and are not part of the final file diff.

E2E Result

Command:

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 ""

Result: passed with max_lpips=0.15.

Metric	Value
LPIPS	0.128773
PSNR	20.125577 dB
MAE	0.047712

I also tested the simpler "text encoder fully FP8 + transformer final 15 fallback" variant. It failed the requested 0.15 gate with LPIPS 0.199243, so the final config keeps text encoder layers 28-35 as BF16 fallback.

Quantized Config

{
    "method": "fp8",
    "ignored_layers": [
        "img_mlp",
        "layers.15..29.{attention.to_qkv,attention.to_out.0,feed_forward.w13,feed_forward.w2}",
        "model.layers.28..35.{self_attn.q_proj,self_attn.k_proj,self_attn.v_proj,self_attn.o_proj,mlp.gate_proj,mlp.up_proj,mlp.down_proj}",
    ],
}

Note: img_mlp appears in Qwen-Image but not in Z-Image, so it is included for parity with PR #1034 but does not match a Z-Image layer by itself.

Visual Comparison

BF16 baseline:

FP8 quantized:

Validation

.venv/bin/python -m pytest tests/diffusion/quantization/test_fp8_config.py -q

test_fp8_config.py: 28 passed.

The local E2E environment used vllm==0.21.0, torch==2.11.0+cu130, lpips==0.1.4, and a single NVIDIA B300 GPU via CUDA_VISIBLE_DEVICES=0.

[congw729]

Maybe upload these images to AWS S3? or move to tests/assets with other images/wavs

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[david6666666]

@yuanheng-zhao @RuixiangMa ptal thx

[Gaohan123]

LGTM. Thanks

[yuanheng-zhao]

LGTM. Btw, are the layers for the final 15 attention/FFN blocks ignored for their weighted influences on generated output diff

[david6666666]

LGTM. Btw, are the layers for the final 15 attention/FFN blocks ignored for their weighted influences on generated output diff

Yes